Цю сторінку перекладено машинно. Читати англійський оригінал.
Перегляд
Сутності каталогу
Шукайте, фільтруйте та сортуйте всі опубліковані сутності. Кожна картка веде до типізованого контракту сутності в її найновішій опублікованій версії.
Сутності: 803
- v0.9.1
Plugin
terraform
ОфіційнийterraformВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/provider/terraform:0.9.1 - v0.9.1
Plugin
helm
ОфіційнийhelmВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/provider/helm:0.9.1 - v0.9.1
Plugin
helm-chart
Офіційнийhelm-chartВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/provider/helm-chart:0.9.1 - v0.9.1
Plugin
compose
ОфіційнийcomposeВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/provider/compose:0.9.1 - v0.9.1
Plugin
cloudformation
ОфіційнийcloudformationВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/provider/cloudformation:0.9.1 - v0.9.1
Plugin
yaml
Parse YAML source into a canonical JSON tree.
Офіційнийparse:yamlВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/parser/yaml:0.9.1 - v0.9.1
Plugin
toml
Parse TOML source into a canonical JSON tree.
Офіційнийparse:tomlВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/parser/toml:0.9.1 - v0.9.1
Plugin
jsonc
Parse JSON with comments — line and block comments and trailing commas — into the canonical JSON tree the JSON parser yields for the same document without them.
Офіційнийparse:jsoncВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/parser/jsonc:0.9.1 - v0.9.1
Plugin
json
Parse JSON source into a canonical JSON tree with sorted keys and exact integers.
Офіційнийparse:jsonВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/parser/json:0.9.1 - v0.9.1
Plugin
hcl
Parse HCL (HashiCorp Configuration Language) source into a canonical JSON tree.
Офіційнийparse:hclВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/parser/hcl:0.9.1 - v0.9.1
Plugin
publishing
ОфіційнийВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/linter/publishing:0.9.1 - v0.9.1
Plugin
n8n
Import an exported n8n workflow into a draft Aggregate: by default the n8n platform carrying the workflow, with the workflow's models, tools and vector stores that match the workspace's catalog wired to it as Components; with the layout setting nodes, every node drafted as a Component of its own.
Офіційнийimport:n8nВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/n8n:0.9.1 - v0.9.1
Plugin
mcp
Import an MCP client configuration (.mcp.json, claude_desktop_config.json, .vscode/mcp.json, .cursor/mcp.json) or an MCP registry server description (server.json) into a draft Aggregate holding one core/ai-tool/mcp instance per server, and a draft Release beside it carrying each server's command line, address or image and its environment.
Офіційнийimport:mcpВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/mcp:0.9.1 - v0.9.1
Plugin
langgraph
Import the graphs a LangGraph project declares in langgraph.json into a draft Aggregate holding one core/ai-agent/langgraph instance per graph, wired to the datastore its checkpointer keeps state in, and a draft Release carrying the environment the configuration states; the project's code is not read, nor any value of an environment file.
Офіційнийimport:langgraphВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/langgraph:0.9.1 - v0.9.1
Plugin
langflow
Import Langflow agent-flow JSON into a draft Aggregate: by default the Langflow platform carrying the flow, with the flow's models, tools and stores that match the workspace's catalog wired to it as Components; with the layout setting nodes, every node drafted as a Component of its own.
Офіційнийimport:langflowВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/langflow:0.9.1 - v0.9.1
Plugin
google-adk
Import the agents declared in the Agent Config files of an application built on Google ADK (root_agent.yaml and the configurations its sub-agents name) into a draft Aggregate holding one core/ai-agent/google-adk instance per agent tree, wired to the Gemini model the tree names, and a draft Release beside it; the application's code is not read, nor its environment file.
Офіційнийimport:google-adkВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/google-adk:0.9.1 - v0.9.1
Plugin
flowise
Import Flowise agent-flow JSON into a draft Aggregate: by default the Flowise platform carrying the flow, with the flow's models, tools and stores that match the workspace's catalog wired to it as Components; with the layout setting nodes, every node drafted as a Component of its own.
Офіційнийimport:flowiseВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/flowise:0.9.1 - v0.9.1
Plugin
dify
Import a Dify application, in the YAML DSL Dify exports, into a draft Aggregate composing the Dify installation core/dify with the application carried in its app_definition parameter; the models, tools, knowledge bases and plugins the application names are reported as configured inside Dify.
Офіційнийimport:difyВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/dify:0.9.1 - v0.9.1
Plugin
crewai
Import the agents and tasks a CrewAI project declares (config/agents.yaml and config/tasks.yaml, or crew.jsonc and agents/*.jsonc, or their .json spellings) into a draft Aggregate holding one core/ai-agent/crewai instance per agent, wired to the model its llm names, and a draft Release beside it; the crew's program is not read.
Офіційнийimport:crewaiВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/crewai:0.9.1 - v0.9.1
Plugin
compose
Import a Compose application (compose.yaml, docker-compose.yml and their override files) into a draft Aggregate holding one instance per service — of the catalog Component its image matches, or of core/compute/server running its image — connected as depends_on says, an instance per model of core/ai-model/docker-model-runner where the workspace's catalog holds it, or of a model Component drafted for it otherwise, and a draft Release carrying each service's image, ports, volumes and environment. Images are matched through a table generated from the Compose Recipes and Components of aggrega-hub at b90d8a27764785bb314595ad8c1493b1ae6423aa.
Офіційнийimport:composeВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/compose:0.9.1 - v0.9.1
Plugin
agent-framework
Import a Microsoft Agent Framework declarative agent (a YAML document of kind Prompt) into a draft Aggregate holding one core/ai-agent/agent-framework instance per declarative agent, wired to the OpenAI model it names, with each MCP server it names written in a comment, since the model's provider reaches it, and a draft Release beside it; the application's code is not read.
Офіційнийimport:agent-frameworkВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/importer/agent-framework:0.9.1 - v0.9.1
Plugin
selector-tagger
Set a tag label on each topology node from a component-to-tag map.
ОфіційнийВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/governance/selector-tagger:0.9.1 - v0.9.1
Plugin
rego
Evaluate operator-authored Rego against a build with an embedded Rego engine.
ОфіційнийВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/governance/rego:0.9.1 - v0.9.1
Plugin
label-injector
Add labels to topology nodes whose existing labels match a rule set.
ОфіційнийВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/governance/label-injector:0.9.1 - v0.9.1
Plugin
cel
Evaluate operator-authored Common Expression Language against a build with an embedded expression evaluator.
ОфіційнийВерсії: 6Опубліковано 2026-11-15oci://hub.aggrega.ai/plugin/governance/cel:0.9.1 - v1.1.0
Recipe
server-ecs
The generic long-running server primitive as one Amazon ECS service on Fargate, running infra.replicas tasks of the Release's image, rendered by vendoring a Terraform module kept beside this Recipe. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/server on a Kubernetes cluster. The cluster, subnets, security groups and execution role come from the operator, under infra.overrides, and each is required; a task role, a public IP, a load balancer target group and a log group are optional, and nothing is emitted for one left unset. Only a task definition and a service are emitted — no cluster, role, security group, log group or load balancer, each of which the operator already runs. No container health check is declared, because ECS runs one inside an image this Recipe does not name; a target group's own check is the operator's. Each env entry an edge's inject block, or the component's own, dictates becomes an entry of the container's environment under that name, its value written whole, so a Terraform reference a neighbour publishes is evaluated by the engine. A file entry is refused by name, as a Fargate task definition writes no file, and so are an arg entry, as no command line is rendered, and any other mechanism. The task runs at the smallest Fargate size meeting both infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 512 MiB. The terraform-aws-modules ECS module is not sourced: its root runs services only in a cluster it creates itself, so it is not complete for a service in the operator's cluster. AWS App Runner is not used: it is no longer open to new customers.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/terraform/server-ecs:1.1.0 - v1.0.1
Recipe
task-ecs
The generic ephemeral compute primitive as an Amazon ECS task on Fargate that an EventBridge Scheduler schedule starts, rendered by vendoring a Terraform module kept beside this Recipe. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/task on Kubernetes, which also runs a task with no schedule once. A task with no schedule carrying core/manual-trigger is deployed without starting, its task definition alone; the operator starts each run with aws ecs run-task, giving the cluster, the task definition and the network configuration from infra.overrides. A scheduled task runs on its schedule, attached or not, and can be started that way too. Any other unscheduled task is refused by name, as nothing on AWS starts it. The cron fields become the scheduler's own expression where the two mean the same times, evaluated in UTC unless the Release names a time zone, and are refused by name where they do not, as for a schedule restricting both the day of the month and the day of the week. The cluster, subnets, security groups, execution role and, with a schedule, the scheduler's role come from infra.overrides and are required. Only a task definition and a schedule are emitted: no role, log group or network resource, each of which the operator already runs, and no service, as a run ends. Each env entry an edge or the component's own inject block dictates becomes a container variable; timeout arrives by default as TASK_TIMEOUT, which no edge may dictate while it is the default, for the code to honour, as ECS sets no deadline on a run. File, arg and other entries are refused by name. The task runs at the smallest Fargate size meeting infra.compute.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/terraform/task-ecs:1.0.1 - v1.1.0
Recipe
server-container-apps
The generic long-running server primitive as one Azure Container Apps app running the Release's image on infra.port, through a Terraform module kept beside this Recipe. Reach for it in an Azure subscription with a Container Apps environment; for terraform/server-ecs or cloudformation/server-ecs in an AWS account, terraform/server-cloud-run on Google Cloud, and helm/server on Kubernetes. Its ingress is internal: only the other apps in its environment reach it, and it is not published at the environment's inbound address. The environment and resource group are the operator's, each required under infra.overrides; user-assigned identities and a private registry, its password read from Key Vault by reference, are optional. Only the app is emitted — no environment, network, identity, registry, Key Vault or log workspace, each the operator's. Each env entry an edge's inject block, or the component's own, dictates becomes a container variable of that name, written whole, so a neighbour's Terraform reference is evaluated by the engine, and each file entry a secret mounted alone at the path named. Neither may dictate a name Container Apps sets itself; arg and other entries are refused by name. The app runs exactly infra.replicas replicas at the smallest Consumption size meeting infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 0.5 GiB. Azure/avm-res-app-containerapp/azurerm is not sourced: at 0.9.0, its newest, it calls Azure/avm-utl-interfaces/azure from the registry, which init fetches outside the pin. It takes hashicorp/azurerm 4 or 5, so it shares a terraform boundary with terraform/postgres-flexible-server or with the azurerm 5 Recipes.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/terraform/server-container-apps:1.1.0 - v1.0.1
Recipe
retrieval-knative
The similarity-retrieval tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The store the tool searches arrives on the tool's datastore socket, and the recipe delivers that store's host, port and protocol into the container alongside the developer's match count, so the running tool addresses the store the topology resolved rather than one named in its own configuration. A connection writing an `inject:` block receives instead the names it dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/retrieval-knative:1.0.1 - v1.1.0
Recipe
server
The generic long-running server primitive on Kubernetes via the helm provider's Workload synthesis: a Deployment plus a ClusterIP Service carrying inbound traffic to the container's port. The Service is what makes the operational type — compute that expects inbound application traffic — readable in the emitted bytes, and it is what separates this realization from the background-worker primitive's, which emits no Service at all. The component declares no developer fields of its own, so everything the workload needs comes from the operator: the image, the port, the replica count and the resource quantities. It needs no cluster control plane, so it carries no prerequisite affirmation and emits no custom resource. No provides contract is emitted, because the component declares none: this primitive publishes nothing across an edge. What reaches the container from the component's own edges is what each edge's inject block dictates, because the image is the Release's own code: env entries become variables and file entries read-only files, and an arg entry or any other delivery mechanism is refused by name. An edge that dictates nothing delivers nothing. The component's own inject block delivers env entries the same way. A credential attached to the component reaches the container the same way, through the delivery pipe: a core/secret-value attachment, its value written into the rendered Deployment, or a core/secret-provided one. A credential published for the pod to reference, such as core/referenced-secret or core/random-password, cannot reach it, because the component declares no credential slot a reference could name, so the build halts naming the attachment.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/server:1.1.0 - v1.0.1
Recipe
task
The generic ephemeral compute primitive on Kubernetes via the helm provider's Workload synthesis, in a run-to-completion shape decided by two things the Aggregate's author writes and by nothing else: the task's `schedule` value and a core/manual-trigger attachment on it. No operator value and no Manifest entry reaches that choice. With a schedule, the task is a CronJob running on it, attached or not. With none, it is a Job that runs once at deploy, unless core/manual-trigger is attached: then it is a suspended CronJob on the fixed schedule 0 0 30 2 *, a date no month holds, so no run starts at deploy, nor if someone resumes it. Start a run of either CronJob with kubectl create job <run-name> --from=cronjob/<cronjob>, the CronJob named as the component is rendered. No Service and no port are emitted: nothing connects to a run that ends. The developer's `timeout` value reaches the container as TASK_TIMEOUT by default for the application's own code to honour, as no deadline is set on the run. An edge's inject block dictates what reaches the container, in every shape: each env entry a variable under that name, beside TASK_TIMEOUT, which no edge may dictate while it is the default, and each file entry a read-only file from a ConfigMap. The component's own inject block delivers env entries the same way, and one reading timeout alone replaces TASK_TIMEOUT, its default name. An arg entry, or any other mechanism, is refused by name: the recipe renders no command line. A credential published for the pod to reference, such as core/referenced-secret, halts the build by name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/task:1.0.1 - v1.1.0
Recipe
server-cloud-run
The generic long-running server primitive as one Cloud Run service running the Release's image on infra.port, through a Terraform module kept beside this Recipe. Reach for it in a Google Cloud project with a VPC network; for terraform/server-ecs or cloudformation/server-ecs in an AWS account, and helm/server on Kubernetes. Its ingress is internal: VPC networks in the project, a Shared VPC network it sends traffic to and an internal load balancer reach it, the internet does not. No invoker binding is emitted, so only a caller the operator grants the Cloud Run Invoker role gets through — reached from inside, as a Kubernetes server is. The project, region, VPC network and service account are the operator's, each required under infra.overrides; a subnet and network tags are optional. Only the service is emitted — no service account, IAM binding, network, firewall rule, load balancer or domain mapping, each the operator's. Each env entry an edge's inject block, or the component's own, dictates becomes a container variable of that name, written whole, so a neighbour's Terraform reference is evaluated by the engine. Neither may dictate a name Cloud Run reserves or refuses; file, arg and other entries are refused by name. Scaling is manual, at infra.replicas instances. The container runs at the smallest whole-vCPU size meeting infra.compute.cpu and infra.compute.memory, in the second-generation environment at Cloud Run's default concurrency, so the Component's defaults run at 1 vCPU and 512 MiB. GoogleCloudPlatform/cloud-run/google is not sourced: at v0.34.2, its newest, it caps hashicorp/google below 8, excluding the range every Google body here declares.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/terraform/server-cloud-run:1.1.0 - v1.1.0
Recipe
mcp
The Model Context Protocol server tool realized on Kubernetes by the helm provider's Workload synthesis, rendered offline as plain manifests. The Workload becomes a Deployment that stays up between calls and a ClusterIP Service that fronts the port it listens on, because an agent reaches this tool by connecting to it rather than by triggering it. The developer's declared effect, timeout and description override reach the running server as environment variables. The image is the Release's, because this tool packages no upstream product: a build with none set halts and names the key. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Deployment. provides is the synthesized Service, so a consumer wires to a host, port, endpoint and protocol and never to the server behind them; it also publishes the request header the server reads a caller's credential from, Authorization unless infra.credential_header names another. Reach for this realization where the server ships as a container image serving the protocol over HTTP; reach for helm/mcp-stdio-bridge where it is a package speaking stdio, and for helm/mcp-remote where someone else already runs it. The two stdio keys, a server_env attachment and infra.endpoint are refused here, each naming the realization that reads it, and so is infra.published_port, a Docker Compose setting. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/mcp:1.1.0 - v1.1.0
Recipe
n8n
The n8n automation platform on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis: a Deployment serving the n8n editor and webhook endpoint on a ClusterIP Service. A workflow definition, when supplied, is carried in a ConfigMap and imported by an init container with `n8n import:workflow` on every start; with none, neither is rendered. The data directory is a scratch volume: the supplied definition returns after a restart because it is imported again, and everything else authored in the editor is lost. The model arrives as OPENAI_BASE_URL and its key as OPENAI_API_KEY; each `tools` and `vector_store` connection under variables named for its alias, as the comments above the workload set out; every credential through a secretKeyRef. A connection writing an `inject:` block receives the names it dictates in their place, as variables or as files mounted at the paths it names; its credentials stay. N8N_BLOCK_ENV_ACCESS_IN_NODE is true unless the component's allow_env_access_in_nodes value is true, so a workflow reads these values through $env only then. `infra.external_base_url`, where set, becomes the editor address and the webhook base. With `values.metrics_enabled` true, N8N_METRICS turns n8n's metrics on, at /metrics on the Service's port, which answers without a sign-in, and provides.metrics_endpoint publishes the address. `infra.published_port` is a Docker Compose setting and halts the build. Reach for this realization where the build must complete with nothing fetched; reach for `n8n-chart`, rendering n8n's own chart, where a chart repository is reachable at build time, or editor state must survive a restart.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/n8n:1.1.0 - v1.1.0
Recipe
n8n-chart
The n8n automation platform on Kubernetes via n8n's own Helm chart, pulled from its OCI registry: one Deployment serving the editor and the webhook endpoint on a ClusterIP Service, on n8n's own SQLite database, with no queue workers, external database or cache. The image is the one the Release names in `infra.image`, never the chart's own default, so the build halts until it is set. The data directory is a persistent volume sized by `infra.storage.size`, so the workflows and credentials authored in the editor, and the encryption key n8n generates on its first start, survive a restart. A workflow definition, when supplied, is carried in a ConfigMap and imported by an init container with `n8n import:workflow` on every start. The `model`, `tools` and `vector_store` connections arrive under the variables the workload-synthesis realization delivers, every credential through a secretKeyRef, a connection writing an `inject:` block receiving the names it dictates as there, and N8N_BLOCK_ENV_ACCESS_IN_NODE is true unless the component's allow_env_access_in_nodes value is true. `infra.external_base_url`, where set, becomes the editor address and the webhook base. With `values.metrics_enabled` true, N8N_METRICS turns n8n's metrics on, at /metrics on the Service's port, which answers that path without a sign-in, and provides.metrics_endpoint publishes the address. `infra.published_port` is a Docker Compose setting and halts the build. The chart's own connection-check Pod is not rendered. Reach for this realization where a chart repository is reachable at build time; reach for `n8n`, the workload-synthesis realization, where the build must complete with nothing fetched.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/n8n-chart:1.1.0 - v1.1.0
Recipe
mcp-stdio-bridge
A Model Context Protocol server published as an npm or PyPI package that speaks only stdio, run on Kubernetes behind supergateway 4.1.0, which serves it over Streamable HTTP. The helm provider synthesizes a Deployment running the bridge image, pinned by digest, for the runtime infra.stdio_runtime names, and a ClusterIP Service on infra.port. The bridge starts the command line infra.stdio_command names, with each item quoted so it reaches the server as written; the launcher fetches the package when the server starts, so the cluster reaches the package registry then. One server process runs per client session, at the path /mcp, and the bridge's /healthz endpoint is the readiness probe. provides.endpoint is the Service's address with that path. Variables the server reads its credentials from arrive through the server_env slot: every data key of the Secret its attachment names becomes one variable in the bridge's environment, by reference, and the server inherits that environment. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates into that environment too, or as files from a ConfigMap emitted beside the Deployment. Reach for this realization where the server is a package speaking stdio; reach for helm/mcp where it ships as an image serving HTTP, for helm/mcp-remote where someone else already runs it, and for compose/mcp-stdio-bridge under Docker Compose. infra.image, infra.endpoint, infra.published_port and the component's own `inject:` block are refused here. What it renders is checked by building it; the bridge's behaviour is read from its documentation and source at the pinned release, and no run of this realization has been observed.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/mcp-stdio-bridge:1.1.0 - v1.1.0
Recipe
mcp-remote
A Model Context Protocol server someone else already runs, reached at the address infra.endpoint names, as a reference realization: nothing is provisioned and no compute runs. The helm provider emits an ExternalName Service mapping an in-cluster name onto the server's host. provides.endpoint is the address as written, host, port and protocol are read from it, and credential_header is the request header the server reads a caller's credential from, Authorization unless infra.credential_header names another. No credential is held here: a consumer that needs one holds it on its own connection to the tool and presents it under that header, because the credential belongs to the consumer rather than to the address. infra.endpoint is the address without any part that carries a key, at least a scheme and a host; a consumer whose server reads its key from the address supplies the whole address through the endpoint slot on its own connection, from a Secret. An address carrying a user name or password is refused. Reach for this realization where the server already runs outside the Release; reach for helm/mcp to run a server shipped as an image serving HTTP, and for helm/mcp-stdio-bridge to run one published as a package speaking stdio. infra.image, the two stdio keys, infra.published_port and a server_env attachment are refused here, and so is an `inject:` block, on a connection from the tool or on the component itself, since no container runs to receive it.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/mcp-remote:1.1.0 - v1.2.0
Recipe
librechat
A self-hosted LibreChat installation on Kubernetes via the helm provider's Workload synthesis: a single-replica Deployment, a ClusterIP Service on the listen port, and one persistent volume for the files people upload; the accounts, conversations, agents and stored per-user credentials live in the document store wired on the `backend` edge. The model edge's endpoint becomes the platform's OpenAI-compatible base URL and its model name the one model offered; the backend edge's host and port compose the document-store connection string, with the sign-in attached on that connection, by reference, as its user part. infra.external_base_url, when set, reaches the platform as DOMAIN_CLIENT and DOMAIN_SERVER; unset, both carry its in-cluster address. Seven credential slots are wired: the model API key on the `model` connection, the store sign-in on the `backend` connection, and five on the component. The two token secrets, the credential-encryption key and its initialization vector have no default: the Component requires their slots. With no model API key attached, each person enters their own key in the interface. With values.metrics_enabled true, the secret attached to metrics_secret reaches the platform as METRICS_SECRET, turning on its Prometheus metrics at /metrics on the listen port, and provides.metrics_endpoint publishes the address with metrics_credential_required true; the build halts when that slot is unfilled. False delivers nothing. Reach for this realization where the build must complete with nothing fetched; reach for `librechat-chart`, rendering the project's own chart, where a chart repository is reachable at build time.
ОфіційнийВерсії: 3Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/librechat:1.2.0 - v1.2.0
Recipe
librechat-chart
A self-hosted LibreChat installation on Kubernetes via the project's own Helm chart, pulled from its OCI registry: one platform process with a ClusterIP Service on the listen port and one persistent volume for the files people upload; the accounts, conversations, agents and stored per-user credentials live in the document store wired on the `backend` edge. The chart was checked for being official, complete and maintained against the chart itself, and its bundled document store, search server, cache and retrieval service all stay off. The model edge's endpoint becomes the platform's OpenAI-compatible base URL and its model name the one model offered, and the backend edge's host and port compose the connection string, with the sign-in attached on that connection, by reference, as its user part. Every credential reaches the platform by a secretKeyRef to the Secret its slot names; the two token secrets, the credential-encryption key and its initialization vector are required. The chart's own connection check, a Pod carrying the `helm.sh/hook: test` annotation whose image no setting pins, is left out of the render. Reach for this realization where a chart repository is reachable at build time; reach for `librechat`, the workload-synthesis realization, where the build must complete with nothing fetched.
ОфіційнийВерсії: 3Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/librechat-chart:1.2.0 - v1.1.0
Recipe
langgraph
A LangGraph agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The endpoint of the model wired into the `model` socket is injected as OPENAI_BASE_URL; the checkpointer datastore wired into the `checkpointer_db` socket contributes its host and port as LANGGRAPH_CHECKPOINTER_HOST and LANGGRAPH_CHECKPOINTER_PORT, and its user and database where it publishes them as LANGGRAPH_CHECKPOINTER_USER and LANGGRAPH_CHECKPOINTER_DATABASE; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY, and a password attached on the `checkpointer_db` connection as LANGGRAPH_CHECKPOINTER_PASSWORD, each via a secretKeyRef. `values.recursion_limit` is delivered as LANGGRAPH_DEFAULT_RECURSION_LIMIT, which LangGraph reads as the limit of a run whose configuration sets none. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL, the API key and the checkpointer password stay. A checkpointer datastore served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, naming the connection, since such a connection hands the application the store's host and port and not its protocol. Its own value graph_entrypoint reaches the container as LANGGRAPH_GRAPH_ENTRYPOINT by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name. infra.published_port, a Docker Compose setting, is refused by name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/langgraph:1.1.0 - v1.1.0
Recipe
langflow
The Langflow visual flow-builder on Kubernetes, synthesized by the helm provider: a Deployment serving the editor on a ClusterIP Service, with a persistent volume holding the flow database and its encryption key; nothing else survives a restart. A supplied flow definition is carried in a ConfigMap Langflow loads at each start, even with signing in required, which an image at Langflow 1.5.1 or older does not do. Signing in is required, as an administrator seeded from values.admin_user and the password attached to admin_password; values.require_authentication false leaves the editor open to anyone reaching it, and doing neither halts the build. The `model` socket's endpoint arrives as OPENAI_BASE_URL and a key attached on that connection as OPENAI_API_KEY. Each `vector_store` and `tools` connection arrives as the Component describes, every credential by a secretKeyRef, and LANGFLOW_SSRF_ALLOWED_HOSTS admits the hosts of both. A `vector_store` connection writing an `inject:` block receives the names it dictates in place of its address and connection-string variables, as variables a flow reads as global variables or as files mounted at the paths it names; its credential stays. With `values.metrics_enabled` true, LANGFLOW_PROMETHEUS_PORT turns on Langflow's metrics at port 9090, which checks no caller and joins the Service, and provides.metrics_endpoint publishes their address; false, it is empty. Reach for this realization where the build must complete with nothing fetched or the metrics are wanted; reach for `langflow-chart`, rendering Langflow's own chart, where a chart repository is reachable at build time. A set `infra.published_port`, a Docker Compose setting, halts the build.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/langflow:1.1.0 - v1.1.0
Recipe
google-adk
An application built on Google's Agent Development Kit (ADK), on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the image the Release supplies, and a ClusterIP Service fronting infra.port, which the application receives as PORT. A readiness probe opens a TCP connection to that port, so the Service sends no traffic until the server accepts one; provides is that Service's host, port, endpoint and protocol. The credential on the `model` connection reaches the application as GOOGLE_API_KEY through a secretKeyRef, so the key never renders into the manifest. Each tool wired into the `tools` socket reaches the application under variables named for the connection's alias, upper-cased with hyphens as underscores: its address as AGGREGA_TOOL_<ALIAS>_URL, and, where the connection carries a credential, that credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL through a secretKeyRef, with the header the tool publishes for it as AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection holding the tool's whole address as a secret delivers it as AGGREGA_TOOL_<ALIAS>_URL through a secretKeyRef instead. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names: a tools connection in place of its address and header variables, while what its attachments deliver stays, and the model connection beside its key. The component's own `inject:` block delivers its entries as a connection's are. infra.published_port, a Docker Compose setting, is refused by name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/google-adk:1.1.0 - v1.0.1
Recipe
http-request-knative
The HTTP-request tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between calls. The replica floor and ceiling carry through as the revision's own scaling bounds. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, so a consumer wires to the tool's declared contract rather than to anything Knative-specific. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/http-request-knative:1.0.1 - v1.1.0
Recipe
langflow-chart
The Langflow visual flow-builder on Kubernetes via the project's own langflow-ide Helm chart, served from its HTTPS chart repository: one StatefulSet pod serving the editor and the flows' endpoint, with a persistent volume holding the flow database and its encryption key. The chart was checked for being official, complete and maintained against the chart itself; its separate front-end pod and its bundled PostgreSQL stay off, and the image is the one the Release names in `infra.image`, never the chart's floating default, so the build halts until it is set. A supplied flow definition is mounted as the directory Langflow loads at each start. Signing in is required, as an administrator seeded from values.admin_user and the password attached to admin_password; values.require_authentication false leaves the editor open, and doing neither halts the build. The `model`, `tools` and `vector_store` connections arrive as the workload-synthesis realization delivers them, every credential through a secretKeyRef, a `vector_store` connection writing an `inject:` block receiving the names it dictates as there, and LANGFLOW_SSRF_ALLOWED_HOSTS admits the wired hosts. The chart's Service carries no port for Langflow's own metrics, so values.metrics_enabled true halts the build on this Recipe. Reach for this realization where a chart repository is reachable at build time; reach for `langflow`, the workload-synthesis realization, where the build must complete with nothing fetched or the metrics are wanted. A set `infra.published_port`, a Docker Compose setting, halts the build.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/langflow-chart:1.1.0 - v1.1.0
Recipe
flowise
The Flowise visual agent-builder on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis: a Deployment serving the editor on a ClusterIP Service, backed by a persistent volume at the platform's home directory holding the chatflow database and the key encrypting the editor's credentials, so both survive a restart; nothing outside that directory does. infra.published_port is refused; reach the Service with kubectl port-forward. The model endpoint wired into the `model` socket is injected as OPENAI_BASE_URL and a key attached on that connection as OPENAI_API_KEY, which a chatflow's OpenAI node reads only while its Connect Credential and Base Path fields are empty, through the client library's environment fallback; a credential on the node wins. Each `tools` and `vector_store` connection reaches the platform under variables named for its alias, as the comments above the workload set out: AGGREGA_TOOL_<ALIAS>_URL, AGGREGA_VECTOR_STORE_<ALIAS>_URL and their parts, each credential by a secretKeyRef. A chatflow reads each through a Runtime variable of the same name. A connection writing an `inject:` block receives the names it dictates instead of its address and part variables, as variables or as files mounted at the paths it names; its credentials stay. A flow_definition, when set, is carried in a ConfigMap and seeded by a one-shot Job in the platform's image: it registers infra.owner_email as the owner on a fresh instance, the owner_password credential read by secretKeyRef, signs in, and updates the chatflow at chatflow_id or imports it there. With none, neither is rendered. The contract is authored against Flowise 3.1.3.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/flowise:1.1.0 - v2.0.0
Recipe
dify
The Dify API and console server on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment and a ClusterIP Service on the API port. The identities its uses-edges resolve compose its connection settings and the in-cluster URLs of its siblings, and its outbound requests to a user-supplied address leave through the proxy. A cache publishing valkey-tls or redis-tls is reached over TLS; another TLS protocol halts the build, and so does a vector-store connection whose target is not a vector engine served over http. By default nothing reaches Dify's vendor from it: infra.telemetry_enabled allows the install event, update check, template lists and Creators Platform, and infra.marketplace_enabled the plugin marketplace. infra.console_url sets the console origin. Its application secret comes from the credential attached to secret_key, which has no default: the Component requires the slot, and the build halts when nothing fills it. Each of its other four credentials comes from a credential when one is attached and otherwise from a pinned development literal; the three it presents to the backend, the sandbox and the plugin runtime are held on those connections. Those defaults are for evaluation and local development. The two keys the server shares with the task worker and the plugin runtime are attached on all three or on none. An app_definition, when set, is carried in a ConfigMap and imported by a one-shot Job running the server's image: it performs Dify's first-time setup if needed as infra.owner_email, with the owner_password credential read by secretKeyRef, signs in, and imports the application unless one of that name is present. With no definition, neither is rendered.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/dify:2.0.0 - v2.0.0
Recipe
dify-worker
The Dify task worker on Kubernetes via the helm provider's Workload synthesis. The worker runs the same image as the API server under a different MODE, takes its work from the Celery broker rather than from a listening socket, and therefore renders as a stateless Deployment with no ports and no Service — nothing connects to it, so it publishes nothing. By default it sends Dify's vendor nothing: infra.telemetry_enabled allows its usage events, and infra.marketplace_enabled the plugin marketplace. The identities its uses-edges resolve compose its connection settings and the single URL each sibling subsystem is addressed by. A cache publishing valkey-tls or redis-tls is reached over TLS; another TLS protocol halts the build, and so does a vector-store connection whose target is not a vector engine served over http. The application secret comes from the credential attached to secret_key, which has no default: the Component requires the slot, and the build halts when nothing fills it. The backend password, the sandbox execution key and the two keys the worker and the plugin runtime present to each other each come from a credential when one is attached and from an inline development default otherwise, matching the defaults the other Recipes in this stack emit; the three the worker presents to the backend, the sandbox and the plugin runtime are held on those connections. These defaults are for evaluation and local development. The two handshake keys are read by the worker, the API server and the plugin runtime alike, so a credential for either is attached on all three or on none of them.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/dify-worker:2.0.0 - v2.0.0
Recipe
dify-worker-beat
The Dify task scheduler on Kubernetes via the helm provider's Workload synthesis. The scheduler runs the same image as the API server under a different MODE and does one thing: read the periodic-task schedule from the backend and enqueue each task onto the Celery broker at its due time. It opens no listening socket, so it renders as a stateless Deployment with no ports and no Service, and it publishes nothing. By default it schedules nothing that reaches Dify's vendor: infra.telemetry_enabled allows the daily usage event, and infra.marketplace_enabled the plugin upgrade check. Exactly one replica is rendered and the count is not a knob. Two schedulers reading one schedule and enqueuing onto one broker would put every periodic task on the queue twice, so scaling this component out multiplies the work rather than sharing it. Throughput is scaled on the task worker, which is what consumes what the scheduler enqueues. The backend and cache identities resolved from the component's uses-edges arrive on .context.inject.env under the names the scheduler itself reads. A cache publishing valkey-tls or redis-tls is reached over TLS, its certificate checked against the cache connection's server_ca, mounted, or the image's default trust store; another TLS protocol on the cache halts the build. The backend password comes from a credential when the backend connection carries one and from an inline development default otherwise, matching the standard-library PostgreSQL Recipe's own default.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/dify-worker-beat:2.0.0 - v1.1.0
Recipe
dify-web
The Dify front end on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the HTTP port). The front end holds no state and owns no datastore, so it renders stateless. The API server's identity resolved from the `api` uses-edge arrives on .context.inject.env and becomes SERVER_CONSOLE_API_URL, the in-cluster base URL the server-side rendering path calls. That is a different address from the one the browser uses: a page rendered in the cluster reaches the API server by its Service name, while the browser reaches it through the ingress tier. So where the operator supplied an external base URL, it is emitted separately as CONSOLE_API_URL and APP_API_URL, which is what the browser reads; left empty, the browser derives the address from the page it was served on. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/dify-web:1.1.0 - v1.1.0
Recipe
crewai
A CrewAI agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. The model name published by the same component is delivered as MODEL, the variable CrewAI resolves first when no model is set in code. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names, beside these. The agent's own values reach the container as CREWAI_ROLE, CREWAI_GOAL, CREWAI_BACKSTORY, CREWAI_ALLOW_DELEGATION and CREWAI_LOG_LEVEL by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name. infra.published_port, a Docker Compose setting, is refused by name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/crewai:1.1.0 - v1.0.1
Recipe
task
The generic ephemeral compute primitive on Docker Compose: one service running infra.image to completion, with restart set to no, so Compose never starts it again when it exits. Reach for helm/task on Kubernetes, which also runs a task on a schedule. With nothing attached, the service runs once each time the stack is started with docker compose up. With core/manual-trigger attached, it carries the profile manual, so docker compose up leaves it stopped and each run is one someone starts with docker compose run --rm <service>, the service named after the component. A task with a schedule is refused by name, as Docker Compose has no scheduler. The service is sized from infra.compute and publishes no port. The developer's timeout value reaches the container as TASK_TIMEOUT by default for the application's own code to honour, as Compose sets no deadline on a run. An edge's inject block dictates what reaches the container: each env entry a variable under that name, beside TASK_TIMEOUT, which no edge may dictate while it is the default, and each file entry a file from a Compose config. The component's own inject block delivers env entries the same way, and one reading timeout alone replaces TASK_TIMEOUT, its default name. An arg entry, or any other mechanism, is refused by name: the recipe renders no command line.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/task:1.0.1 - v1.1.0
Recipe
agent-framework
An application built on Microsoft Agent Framework, on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the image the Release supplies, with no Service and no port. The model wired into the `model` socket reaches the application as the variables Agent Framework's Python OpenAI clients read: its endpoint as OPENAI_BASE_URL, its model name as OPENAI_MODEL, and the connection's credential as OPENAI_API_KEY through a secretKeyRef, so the key never renders into the manifest. Each tool wired into the `tools` socket reaches the application under variables named for the connection's alias, upper-cased with hyphens as underscores: its address as AGGREGA_TOOL_<ALIAS>_URL, and, where the connection carries a credential, that credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL through a secretKeyRef, with the header the tool publishes for it as AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection holding the tool's whole address as a secret, for a server that reads its key from its address, delivers that secret as AGGREGA_TOOL_<ALIAS>_URL through a secretKeyRef instead. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names: a tools connection in place of its address and header variables, while what its attachments deliver stays, and the model connection beside the variables Agent Framework reads. The component's own `inject:` block delivers its entries as a connection's are. infra.published_port, a Docker Compose setting, is refused by name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/helm/agent-framework:1.1.0 - v2.0.0
Recipe
server
The generic long-running server primitive on Docker Compose: one service running infra.image, listening on infra.port. Where the Release sets infra.published_port, that port is published on the host's loopback address, 127.0.0.1, forwarding to infra.port, so the application answers there and on no other host interface; unset, nothing is published, and the service is reached at its name on the Compose network. Reach for helm/server on Kubernetes. The service is sized from infra.compute. It runs one container, so an infra.replicas above one is refused by name, as Docker Compose keeps no replica count for a service it is handed this way. An edge's inject block dictates what reaches the container: each env entry a variable under that name, and each file entry a file from a Compose config. An arg entry, or any other mechanism, is refused by name: the recipe renders no command line. An edge that dictates nothing delivers nothing. The component's own inject block delivers env entries the same way.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/server:2.0.0 - v1.0.1
Recipe
nginx
NGINX reverse-proxy gateway for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload into a service entry. The upstream host and port arrive on .context.inject.env under the two variables the locals name, which the `upstream` edge's inject block names, and are written into a minimal NGINX server configuration whose proxy_pass forwards matching requests to that upstream. With an edge wired on any socket but guard and either variable missing from the pipe, the build halts naming the edge and its socket; with neither on the pipe and no such edge, the location answers 503. A core/prompt-injection-shield attachment's realization (prompt-injection-shield-nginx) writes the two variables with the address of the guard wired to the guard socket. provides.host is the bare service name, resolved on the compose network. Each core/traffic-shadowing attachment on the component adds a mirror to that location, sending a copy of every forwarded request, with its original URI, to the service the attachment's infra names, whose response NGINX discards. NGINX cannot copy a share of the requests, so the FeatureRecipe realizing the attachment (traffic-shadowing-nginx) halts on a share below 100, and this Recipe halts on an attachment to a gateway that forwards nothing. With no such attachment the configuration carries no mirror. Each core/concurrency-limit attachment adds a cap, one total across NGINX's worker processes, and a request past it is answered 503; concurrency-limit-nginx refuses scope model.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/nginx:1.0.1 - v1.0.0
Recipe
retrieval
The similarity-retrieval tool on Docker Compose: one service running the image the Release names in infra.image, listening on infra.port on the compose network, where an agent reaches it at the service's own name. Nothing is published to the host. The store the tool searches arrives on the tool's datastore socket, and its host, port and protocol reach the container as RETRIEVAL_TOOL_STORE_HOST, RETRIEVAL_TOOL_STORE_PORT and RETRIEVAL_TOOL_STORE_PROTOCOL beside the developer's match count, the names helm/retrieval-knative delivers them under; a connection writing an `inject:` block receives instead the names it dictates, as variables or as files from compose configs. It differs from the Knative realization in one way a reader must know: the service is one container that runs while the stack runs. Docker Compose has no autoscaler, so the tool does not scale to zero between calls and runs no second replica; an infra.min_replicas above one is refused by name, and infra.max_replicas, a ceiling one container never passes, is honoured as it stands. No control plane is involved, so infra.control_plane_present is not read. Reach for this realization under Docker Compose, and for helm/retrieval-knative on a Kubernetes cluster running Knative Serving. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name. An arg entry, or any other mechanism, is refused by name.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/retrieval:1.0.0 - v1.0.0
Recipe
mcp
The Model Context Protocol server tool on Docker Compose, for local development: one service running the image the Release names in infra.image, serving the protocol over HTTP on infra.port, where an agent on the compose network reaches it at the service's name. Where the Release sets infra.published_port, that port on the host's loopback address, 127.0.0.1, forwards to infra.port; unset, nothing is published, and two servers that each publish take distinct published_port values. The tool packages no upstream product, so a build with no image set halts and names the key. provides publishes host, port, endpoint, protocol and credential_header. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from compose configs; the component's own `inject:` block delivers its entries the same way. The server's credentials arrive through the server_env slot: attach core/referenced-secret-set mapped to compose/dotenv-secret-set, each entry's value a Compose variable reference such as ${GITHUB_TOKEN}, which Docker Compose substitutes when the stack starts. An attachment publishing only a secret resource's name, or delivering no variable, halts the build. Reach for this realization under Docker Compose where the server ships as a container image serving the protocol over HTTP; reach for helm/mcp on Kubernetes, for compose/mcp-stdio-bridge where the server is a package speaking stdio, and for compose/mcp-remote where someone else already runs it. The two stdio keys and infra.endpoint are refused here.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/mcp:1.0.0 - v1.1.0
Recipe
n8n
The n8n automation platform for local development via Docker Compose, synthesized from the Workload kind by the compose provider. The data directory is a named volume, the only home of the workflows and the credentials authored in the editor, which survive a restart; state written anywhere else in the container does not. A workflow definition, when supplied on `values.flow_definition`, is carried in a compose config and imported by an init service with `n8n import:workflow` before the platform starts, on every start, so it is applied again rather than seeded once; with none, neither is rendered. The model wired into the `model` socket arrives as OPENAI_BASE_URL and its key as OPENAI_API_KEY. Each tool and vector store arrives under the variables helm/n8n names, AGGREGA_TOOL_<ALIAS>_URL, AGGREGA_VECTOR_STORE_<ALIAS>_URL, their parts and credentials, each credential written into the file as it was delivered. A connection writing an `inject:` block receives the names it dictates as on helm/n8n, a file as a compose config mounted at the path it names. N8N_BLOCK_ENV_ACCESS_IN_NODE is true unless the component's allow_env_access_in_nodes value is true, so a workflow reads these values through $env only then. `infra.external_base_url`, where set, becomes the editor address and the webhook base. With `values.metrics_enabled` true, N8N_METRICS turns n8n's metrics on, at /metrics on `infra.port`, which answers that path without a sign-in and which any peer on the compose network reaches, and provides.metrics_endpoint publishes the address; false, that key is empty. With `infra.published_port` set, the editor is published on that port of the host's loopback address, 127.0.0.1, alone.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/n8n:1.1.0 - v1.1.0
Recipe
mcp-stdio-bridge
A Model Context Protocol server published as an npm or PyPI package that speaks only stdio, run for local development via Docker Compose behind supergateway 4.1.0, which serves it over Streamable HTTP. One service runs the bridge image, pinned by digest, for the runtime infra.stdio_runtime names, and starts the command line infra.stdio_command names; the launcher fetches the package when the server starts. One server process runs per client session, at the path /mcp on infra.port, where a consumer on the compose network reaches it at the service's name. Where the Release sets infra.published_port, that port on the host's loopback address, 127.0.0.1, forwards to infra.port; unset, nothing is published, and two servers that each publish take distinct published_port values. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates into the bridge's environment, which the server inherits, or as files from compose configs. The server's credentials arrive through the server_env slot: attach core/referenced-secret-set mapped to compose/dotenv-secret-set, each entry's value a Compose variable reference such as ${GITHUB_TOKEN}, which Docker Compose substitutes when the stack starts. An attachment publishing only a secret resource's name, or delivering no variable, halts the build. Reach for this realization under Docker Compose; reach for helm/mcp-stdio-bridge on Kubernetes, for compose/mcp or helm/mcp where the server ships as an image serving HTTP, and for compose/mcp-remote where someone else already runs it. infra.image, infra.endpoint and the component's own `inject:` block are refused here.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/mcp-stdio-bridge:1.1.0 - v1.1.0
Recipe
mcp-remote
A Model Context Protocol server someone else already runs, reached at the address infra.endpoint names, as a reference realization for local development via Docker Compose: no service is defined for it and nothing runs. The address reaches a consuming service through the injection context rather than through a service definition, because compose has no counterpart to an in-cluster alias. provides.endpoint is the address as written, host, port and protocol are read from it, and credential_header is the request header the server reads a caller's credential from, Authorization unless infra.credential_header names another. No credential is held here: a consumer that needs one holds it on its own connection to the tool and presents it under that header, because the credential belongs to the consumer rather than to the address. infra.endpoint is the address without any part that carries a key, at least a scheme and a host; a consumer whose server reads its key from the address supplies the whole address through the endpoint slot on its own connection, as a Compose variable reference. An address carrying a user name or password is refused. Reach for this realization under Docker Compose where the server already runs outside the Release; reach for helm/mcp-remote on Kubernetes, for compose/mcp to run a server shipped as an image serving HTTP, and for compose/mcp-stdio-bridge to run a server published as a package speaking stdio. infra.image, the two stdio keys, infra.published_port and a server_env attachment are refused here, and so is an `inject:` block, on a connection from the tool or on the component itself, since no container runs to receive it.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/mcp-remote:1.1.0 - v2.0.0
Recipe
librechat
A self-hosted LibreChat installation for local development via Docker Compose, synthesized from the Workload kind by the compose provider: one service and a named volume for the files people upload. The accounts, conversations, agents and stored per-user credentials live in the document store wired on the `backend` edge. The model edge's endpoint becomes the platform's OpenAI-compatible base URL and its model name the one model offered; the backend edge's host and port compose the document-store connection string, with the sign-in attached on that connection as its user part. With infra.published_port set, the interface is published on the host's loopback address and opens at http://127.0.0.1:<published_port>; unset, nothing is published and only peers on the compose network reach it. The four required secrets, an optional model API key and an optional store sign-in arrive as literal environment values in the Compose file, as every credential does on this target, unless the attached value is a Compose variable reference. With no model API key attached, each person enters their own key in the interface. With values.metrics_enabled true, the secret attached to metrics_secret turns the platform's metrics on at /metrics on the listen port, and provides.metrics_endpoint publishes the address; the build halts when that slot is unfilled.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/librechat:2.0.0 - v1.1.0
Recipe
langgraph
A LangGraph agent application for local development via Docker Compose. The `model` socket's endpoint is rendered as OPENAI_BASE_URL; the checkpointer datastore wired into the `checkpointer_db` socket contributes its host and port as LANGGRAPH_CHECKPOINTER_HOST and LANGGRAPH_CHECKPOINTER_PORT, and its user and database where it publishes them as LANGGRAPH_CHECKPOINTER_USER and LANGGRAPH_CHECKPOINTER_DATABASE; a model API key attached on the `model` connection is rendered as OPENAI_API_KEY, and a password attached on the `checkpointer_db` connection as LANGGRAPH_CHECKPOINTER_PASSWORD. `values.recursion_limit` is rendered as LANGGRAPH_DEFAULT_RECURSION_LIMIT, which LangGraph reads as the limit of a run whose configuration sets none. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL, the API key and the checkpointer password stay. A checkpointer datastore served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, since such a connection hands the application the store's host and port and not its protocol. With infra.image unset it runs the catalog's quickstart agent, a small LangGraph application whose own code reads the model's default variables and the store's host and port, and no other. infra.published_port, where set, publishes the application on 127.0.0.1, forwarding to infra.port. Its own value graph_entrypoint reaches the container as LANGGRAPH_GRAPH_ENTRYPOINT by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийВерсії: 3Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/langgraph:1.1.0 - v1.1.0
Recipe
langflow
The Langflow visual flow-builder for local development via Docker Compose, synthesized from the Workload kind by the compose provider. The named volume holds the flow database and its encryption key; nothing else survives a restart. A run-once step gives that volume to the user the platform's published image runs as, uid 1000, before the platform starts. With `infra.published_port` set, the editor is published at that port on 127.0.0.1. Signing in is required, as an administrator seeded from values.admin_user and the password attached to admin_password; values.require_authentication false leaves the editor open to anyone who can reach it, and doing neither halts the build. The `model` socket's endpoint is rendered as OPENAI_BASE_URL and a key attached on that connection as OPENAI_API_KEY. A definition on `values.flow_definition` is carried in a compose config Langflow loads at each start, even with signing in required, which an image at Langflow 1.5.1 or older does not do. Each `vector_store` connection arrives under the variables helm/langflow names, a credential written as it was delivered, and a store publishing a database and a user also as AGGREGA_VECTOR_STORE_<ALIAS>_CONNECTION_STRING for Langflow's PGVector component. Each `tools` connection's address is written into the flow's MCP Tools node named for its alias, whose header names the variable holding its credential, as on helm/langflow. LANGFLOW_SSRF_ALLOWED_HOSTS admits the hosts of both. With `values.metrics_enabled` true, LANGFLOW_PROMETHEUS_PORT turns Langflow's metrics on, on port 9090, which checks no caller and which any peer on the compose network reaches, and provides.metrics_endpoint publishes the address.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/langflow:1.1.0 - v1.1.0
Recipe
google-adk
An application built on Google's Agent Development Kit (ADK), for local development via Docker Compose. Synthesized-mode Workload: the image the Release supplies and its environment are translated from the Workload kind by the compose provider. The application receives infra.port as PORT, and provides is the service's name on the compose network with that port. Where the Release sets infra.published_port, the application is published on the host's loopback address, 127.0.0.1, on that port, forwarding to infra.port; unset, nothing is published on the host. The credential on the `model` connection reaches the application as GOOGLE_API_KEY, as on helm/google-adk. Each tool wired into the `tools` socket reaches it under the variables helm/google-adk names: its address as AGGREGA_TOOL_<ALIAS>_URL, or the whole address the connection holds as a secret in its place, and the connection's credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header the tool publishes, AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block also receives the names it dictates, as on helm/google-adk, a file as a compose config mounted at the path it names. Compose has no secret reference, so a secret held on a connection is written into the file as it was delivered; a Compose variable reference such as ${MODEL_API_KEY} is written as that reference, and Docker Compose substitutes it when the stack starts. Two secrets the Release alone attaches, a model credential beside a tool credential, both arrive under the key value and halt the build naming both; an Aggregate naming values.key on one of them avoids it. The component's own `inject:` block delivers its entries as a connection's are.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/google-adk:1.1.0 - v1.0.0
Recipe
http-request
The HTTP-request tool on Docker Compose: one service running the image the Release names in infra.image, listening on infra.port on the compose network, where an agent reaches it at the service's own name. Nothing is published to the host. The developer's method, effect and timeout reach the container as HTTP_TOOL_METHOD, HTTP_TOOL_SIDE_EFFECTS and HTTP_TOOL_TIMEOUT_SECONDS, the names helm/http-request-knative delivers them under. It differs from the Knative realization in one way a reader must know: the service is one container that runs while the stack runs. Docker Compose has no autoscaler, so the tool does not scale to zero between calls and runs no second replica; an infra.min_replicas above one is refused by name, and infra.max_replicas, a ceiling one container never passes, is honoured as it stands. No control plane is involved, so infra.control_plane_present is not read. Reach for this realization under Docker Compose, and for helm/http-request-knative on a Kubernetes cluster running Knative Serving. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from compose configs. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name. An arg entry, or any other mechanism, is refused by name.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/http-request:1.0.0 - v1.1.0
Recipe
flowise
The Flowise visual agent-builder for local development via Docker Compose, synthesized from the Workload kind by the compose provider. The chatflow database and the key encrypting the editor's credentials live on the named data volume, mounted at the platform's home directory, and survive a restart; nothing outside that directory does. infra.published_port publishes the editor on 127.0.0.1. The model endpoint wired into the `model` socket is rendered as OPENAI_BASE_URL and a key attached on that connection as OPENAI_API_KEY, which a chatflow's OpenAI node reads only while its Connect Credential and Base Path fields are empty, through the client library's environment fallback, as on helm/flowise. Each `tools` and `vector_store` connection reaches the platform under the variables helm/flowise names, AGGREGA_TOOL_<ALIAS>_URL, AGGREGA_VECTOR_STORE_<ALIAS>_URL, their parts and credentials, read by a chatflow through a Runtime variable of the same name. Compose has no secret reference, so a secret is written into the file as it was delivered, a Compose variable reference such as ${STORE_API_KEY} as that reference. A connection writing an `inject:` block receives the names it dictates as on helm/flowise, a file as a compose config mounted at the path it names. A flow_definition, when set, is carried in a compose config and seeded by a service that runs once, in the platform's image, after the platform reports healthy: it registers infra.owner_email as the owner on a fresh instance, with the owner_password credential as delivered, signs in, and updates the chatflow at chatflow_id or imports it there. With none, neither is rendered. The contract is authored against Flowise 3.1.3.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/flowise:1.1.0 - v1.0.0
Recipe
dify
The Dify API and console server on Docker Compose via the compose provider's Workload synthesis: one service on the API port, infra.port, reached on the compose network under its service name. A Release setting infra.published_port publishes it on the host's loopback address, 127.0.0.1, at that port, where a client on the machine running the stack calls it and any process on that machine can reach it; nothing on the host's other interfaces does. Unset, nothing is published. It reads the same values, connections and credentials as helm/dify and writes the same settings, and it applies outstanding schema migrations at start. By default nothing reaches Dify's vendor from it: infra.telemetry_enabled allows the install event, update check, template lists and Creators Platform, and infra.marketplace_enabled the plugin marketplace. infra.console_url is written as the console's address and the origin the server admits browser requests from. Unlike on helm, nothing restarts a service that exits, so a step first waits until the backend accepts TCP connections. Each credential is the value its attachment delivers, written into the file as delivered, so an origin written as a Compose variable reference keeps it out of the file; unattached, the pinned development literals are written. An app_definition, when set, is imported by a one-shot service running the server's image with the owner password as delivered. It halts the build on a credential publishing a secret resource's name, on one that delivers no value, and on a cache connection holding a server_ca.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify:1.0.0 - v1.0.0
Recipe
dify-worker
The Dify task worker on Docker Compose via the compose provider's Workload synthesis: one service running the API server's image under a different MODE, taking its work from the Celery broker, so it publishes no port and nothing. It reads the same values, connections and credentials as helm/dify-worker and writes the same settings. By default it sends Dify's vendor nothing: infra.telemetry_enabled allows its usage events, and infra.marketplace_enabled the plugin marketplace. Unlike on helm, nothing restarts a service that exits, so a step first waits until the backend accepts TCP connections. Each credential is the value its attachment delivers, written into the file as delivered, so an origin written as a Compose variable reference keeps it out of the file; unattached, the pinned development literals are written. It halts the build on a credential publishing a secret resource's name, on one that delivers no value, and on a cache connection holding a server_ca.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-worker:1.0.0 - v1.0.0
Recipe
dify-worker-beat
The Dify task scheduler on Docker Compose via the compose provider's Workload synthesis: one service running the API server's image under a different MODE, enqueuing each periodic task onto the Celery broker at its due time. It opens no listening socket and publishes nothing, and it writes the same settings as helm/dify-worker-beat. One service runs, and a service on this target has one container, so the count is not a knob. By default it schedules nothing that reaches Dify's vendor: infra.telemetry_enabled allows the daily usage event, and infra.marketplace_enabled the plugin upgrade check. Unlike on helm, nothing restarts a service that exits, so a step first waits until the backend accepts TCP connections. The backend password is the value the backend connection's credential delivers, written into the file as delivered, or the pinned development literal when none is attached. It halts the build on a credential publishing a secret resource's name, on one that delivers no value, and on a cache connection holding a server_ca.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-worker-beat:1.0.0 - v1.0.0
Recipe
dify-ssrf-proxy
The Dify egress proxy on Docker Compose via the compose provider's Workload synthesis: one service running Squid, reached on the compose network under its service name and published on no host port. It keeps nothing across a restart, so it declares no volume. It renders the same Squid configuration as helm/dify-ssrf-proxy, resolved when the stack is compiled, with the same access rules in the same order: any destination in private, loopback, link-local, shared-CGN or multicast address space is refused before the named domains and the calling services are permitted, so a supplied address reaches the public internet and cannot reach the stack's own services. A private range becomes reachable only by being named in allowed_private_networks. The configuration is a compose config mounted as the one file Squid reads, where the helm Recipe mounts a ConfigMap directory. Its health check opens a TCP connection to the proxy port with the image's own bash, since the image carries no nc. It reads every value, infra key and sizing the helm Recipe reads, and refuses nothing beyond the Component's schema.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-ssrf-proxy:1.0.0 - v1.0.0
Recipe
dify-web
The Dify front end on Docker Compose via the compose provider's Workload synthesis: one service on the HTTP port, infra.port, reached on the compose network under its service name, so a reverse proxy or an ingress tier in front of it routes the browser path to it. A Release setting infra.published_port publishes it on the host's loopback address, 127.0.0.1, at that port, where a browser on the machine running the stack opens it and any process on that machine can reach it; nothing on the host's other interfaces does. Unset, nothing is published. It holds no state and declares no volume. It writes the same settings as helm/dify-web. The API server's address, resolved from the api connection, becomes SERVER_CONSOLE_API_URL, the base URL the server-side rendering path calls on the compose network; an external base URL the operator supplies is written separately as CONSOLE_API_URL and APP_API_URL, the addresses the browser calls, and left empty the browser derives them from the page it was served on. Its health check opens a TCP connection to the HTTP port on the container's IPv4 loopback with the image's own nc. It halts the build where the api connection leaves API_HOST or API_PORT undelivered.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-web:1.0.0 - v1.0.0
Recipe
dify-sandbox
The Dify code sandbox on Docker Compose via the compose provider's Workload synthesis: one service on the execution port, reached on the compose network under its service name and published on no host port. Each execution is independent, so it declares no volume. Its outbound client is pointed at the proxy wired on the ssrf_proxy connection, as with helm/dify-sandbox, and it shares the stack's one network rather than sitting on an isolated one as in Dify's own Compose file. The execution key is the value attached to code_execution_key, written into the emitted file as delivered, so an origin written as a Compose variable reference keeps the key out of the file; unattached, the pinned development literal is written. Declared Python requirements are one compose config mounted as the requirements file. It halts the build where the ssrf_proxy connection leaves SSRF_PROXY_HOST or SSRF_PROXY_PORT undelivered, and where the attached key's realization publishes the name of a secret resource or delivers no value.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-sandbox:1.0.0 - v1.0.0
Recipe
dify-plugin-daemon
The Dify plugin runtime on Docker Compose via the compose provider's Workload synthesis: one service on the dispatch port, reached on the compose network under its service name and published on no host port. Installed plugin packages and their prepared environments live on a named volume, whose declared size reaches no emitted byte on this target. It writes the same settings as helm/dify-plugin-daemon, keeping its catalogue in a database of its own on the backend's server, and publishes its Prometheus metrics address as provides.metrics_endpoint. As on helm, a step first waits until the backend accepts TCP connections. Each credential is the value its attachment delivers, written into the file as delivered, so an origin written as a Compose variable reference keeps it out of the file; unattached, the pinned development literals are written. It halts the build on a credential publishing a secret resource's name, on one that delivers no value, and on a cache connection holding a server_ca.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/dify-plugin-daemon:1.0.0 - v1.1.0
Recipe
crewai
A CrewAI agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The container declares infra.port, the port the application listens on, which is delivered to it under no variable. Where the Release sets infra.published_port, the application is published on the host's loopback address, 127.0.0.1, on that port, forwarding to infra.port; unset, nothing is published on the host. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; the model name is delivered as MODEL, the variable CrewAI resolves first when no model is set in code. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names, beside these. The agent's own values reach the container as CREWAI_ROLE, CREWAI_GOAL, CREWAI_BACKSTORY, CREWAI_ALLOW_DELEGATION and CREWAI_LOG_LEVEL by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/crewai:1.1.0 - v1.0.1
Recipe
task-ecs
The generic ephemeral compute primitive as an Amazon ECS task on Fargate that an EventBridge Scheduler schedule starts, rendered as a CloudFormation fragment the provider assembles into the boundary's stack. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/task on Kubernetes, which also runs a task with no schedule once. A task with no schedule carrying core/manual-trigger is deployed without starting, its task definition alone; the operator starts each run with aws ecs run-task, giving the cluster, task definition and network configuration from infra.overrides. A scheduled task runs on its schedule, attached or not, and can be started that way too. Any other unscheduled task is refused by name, as nothing on AWS starts it. The cron fields become the scheduler's own expression where the two mean the same times, evaluated in UTC unless the Release names a time zone, and are refused by name where they do not, as for a schedule restricting both the day of the month and the day of the week. The cluster, subnets, security groups, execution role and, with a schedule, the scheduler's role come from infra.overrides and are required. Only a task definition and a schedule are emitted: no role, log group or network resource, each of which the operator already runs, and no service, as a run ends. Each env entry an edge or the component's own inject block dictates becomes a container variable; timeout arrives by default as TASK_TIMEOUT, which no edge may dictate while it is the default, for the code to honour, as ECS sets no deadline on a run. File, arg and other entries are refused by name. The task runs at the smallest Fargate size meeting infra.compute.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/cloudformation/task-ecs:1.0.1 - v1.1.0
Recipe
agent-framework
An application built on Microsoft Agent Framework, for local development via Docker Compose. Synthesized-mode Workload: the image the Release supplies and its environment are translated from the Workload kind by the compose provider. Where the Release sets infra.published_port, the application is published on 127.0.0.1 on that port, forwarding to infra.port; unset, nothing is published. The model wired into the `model` socket reaches the application as the variables Agent Framework's Python OpenAI clients read: its endpoint as OPENAI_BASE_URL, its model name as OPENAI_MODEL, and the connection's credential as OPENAI_API_KEY. Each tool wired into the `tools` socket reaches it under the variables helm/agent-framework names: its address as AGGREGA_TOOL_<ALIAS>_URL, or the whole address the connection holds as a secret in its place, and the connection's credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header the tool publishes, AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block also receives the names it dictates, as on helm/agent-framework, a file as a compose config mounted at the path it names. Compose has no secret reference, so a secret held on a connection is written into the file as it was delivered; a Compose variable reference such as ${MODEL_API_KEY} is written as that reference, and Docker Compose substitutes it when the stack starts. Two secrets the Release alone attaches, a model credential beside a tool credential, both arrive under the key value and halt the build naming both; an Aggregate naming values.key on one of them avoids it. The component's own `inject:` block delivers its entries as a connection's are.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/compose/agent-framework:1.1.0 - v1.1.0
Recipe
server-ecs
The generic long-running server primitive as one Amazon ECS service on Fargate, running infra.replicas tasks of the Release's image, rendered as a CloudFormation fragment the provider assembles into the boundary's stack. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/server on a Kubernetes cluster. The cluster, subnets, security groups and execution role come from the operator, under infra.overrides, and each is required; a task role, a public IP, a load balancer target group and a log group are optional, and nothing is emitted for one left unset. Only a task definition and a service are emitted — no cluster, role, security group, log group or load balancer, each of which the operator already runs. No container health check is declared, because ECS runs one inside an image this Recipe does not name; a target group's own check is the operator's. Each env entry an edge's inject block, or the component's own, dictates becomes an entry of the container's environment under that name, its value written whole, so a neighbour's address arrives as an Fn::GetAtt in the same stack and as an Fn::ImportValue from another. A file entry is refused by name, as a Fargate task definition writes no file, and so are an arg entry, as no command line is rendered, and any other mechanism. The task runs at the smallest Fargate size meeting both infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 512 MiB. It is written as a fragment, as the catalog writes its CloudFormation Recipes, rather than sourced from a published template. AWS App Runner is not used: it is no longer open to new customers.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/recipe/cloudformation/server-ecs:1.1.0 - v1.0.1
FeatureRecipe
k8s-existing-secret
Points a workload at a Kubernetes Secret that already exists in the namespace it runs in. origin — from values or infra, whichever carries it — is read as that Secret's name and values.key as the data key inside it, and both are published exactly as written, as the locator a consuming workload references with a secretKeyRef or a chart reads through its existing-secret setting, so an operator can name the Secret in the Release. It renders no resource of its own, so the credential material never passes through an authored file or a rendered artifact: whoever deploys creates the Secret, and a workload referencing a Secret that does not exist fails to start rather than starting with a default. Because the name is published as written, it must already be a valid Kubernetes object name, and an empty origin names no Secret, so the build halts on it, naming the component. Reach for this realization where the Secret already exists in the target namespace — created by hand, by a pipeline, or by a tool outside the Release. Reach for k8s-external-secret where a secret-store operator runs in the cluster and should fetch the credential from its store, and for k8s-secret-from-parameter where neither holds and the material may sit in the rendered output, as it may for evaluation.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/featurerecipe/helm/k8s-existing-secret:1.0.1 - v1.0.1
Component
core/compute/task
A generic ephemeral compute primitive. Its operational type is `task` — ephemeral compute that runs to completion and dies, such as a database migration — with two optional fields of its own, `schedule` and `timeout`. An Aggregate declares this for work that ends, rather than `server` or `worker`, which are both long-running. helm/task runs it on Kubernetes, on its `schedule` as a CronJob and once at deploy without one. compose/task runs it on Docker Compose once at each `docker compose up`, and refuses a `schedule`. terraform/task-ecs and cloudformation/task-ecs run it as an Amazon ECS task on Fargate, on its schedule through EventBridge Scheduler, and refuse it with none, as nothing on AWS would start it. With no `schedule`, `core/manual-trigger` holds it until someone starts a run, with the command each realization's description names. Nothing renders until the Release sets `infra.image`. The image is the Release's own code, so an edge delivers what its `inject:` block dictates, and no declared socket narrows the mechanism: `env` entries become variables beside `TASK_TIMEOUT`, `timeout`'s default name, which the component's own `env` entries may replace and, while it is the default, no edge may dictate; `file` entries files, read-only on Kubernetes, and none on Amazon ECS; and `arg` or any other mechanism is refused by name, as no realization renders a command line. A `core/secret-value` attachment's value is written into the rendered workload in the clear, and on Docker Compose so is a `core/referenced-secret` attachment's origin; no other realization delivers `core/referenced-secret`, none delivers `core/random-password`, and the helm realization refuses either by name.
ОфіційнийcomputeЗалежні: 6Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/compute/task:1.0.1 - v1.0.0
FeatureRecipe
dotenv-referenced-secret
Compose realization of core/referenced-secret, for local development. It delivers origin — from values or infra, whichever carries it — under values.key on the consuming component's inject pipe, and a compose Recipe reading the slot writes that value into the emitted docker-compose file, exactly as authored. A literal origin therefore lands in the file as the credential itself. An origin written as a Compose variable reference, such as ${DIFY_DB_PASSWORD}, lands as that reference instead, and Docker Compose substitutes it when the stack starts — from the shell environment, a file named with --env-file, or a .env file in the project directory — and substitutes an empty string, with a warning, for a variable it cannot resolve and that names no default. The compiler neither reads nor writes a .env file; whoever starts the stack maintains it. Docker Compose has no secret resource for a consumer to reference, so this realization emits none and the name it publishes is empty; the key it publishes is values.key, the name the value travels under. origin is therefore the value, never the name of a Secret. Reach for it under Docker Compose; on Kubernetes reach for helm/k8s-existing-secret, helm/k8s-external-secret or helm/k8s-secret-from-parameter.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/featurerecipe/compose/dotenv-referenced-secret:1.0.0 - v1.0.0
FeatureRecipe
dotenv-secret-set
Compose realization of core/referenced-secret-set, for local development. It delivers every entry — from values or infra, whichever carries them — on the consuming component's inject pipe, one environment variable per entry, named as the entry's key and carrying its value, and a compose Recipe that delivers its attachments' variables writes each into the emitted docker-compose file as the service's environment value, exactly as authored. A literal value therefore lands in the file as the credential itself. A value written as a Compose variable reference, such as ${GITHUB_TOKEN}, lands as that reference instead, and Docker Compose substitutes it when the stack starts — from the shell environment, a file named with --env-file, or a .env file in the project directory — and substitutes an empty string, with a warning, for a variable it cannot resolve and that names no default. The compiler neither reads nor writes a .env file and emits no env_file entry; whoever starts the stack maintains the .env file. Docker Compose has no secret resource for a consumer to reference, so this realization emits none, and the name it publishes is empty: a consumer reads the entries from its inject pipe rather than from a resource. Reach for it under Docker Compose; reach for helm/k8s-secret-set on Kubernetes, where the entries are held in a Secret the consumer references by name.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/featurerecipe/compose/dotenv-secret-set:1.0.0 - v1.1.0
Component
core/compute/server
A generic long-running server primitive. It adds no fields of its own; what it contributes is its operational type, `server` — long-running, stateless compute that expects inbound application traffic. An Aggregate declares this for a component that stays up and serves requests, rather than `worker`, with zero inbound application ports, or the run-once `task`. On a Kubernetes cluster it is a Deployment behind a ClusterIP Service. In an AWS account with an ECS cluster and a VPC, terraform/server-ecs and cloudformation/server-ecs run it as an Amazon ECS service on Fargate. In a Google Cloud project with a VPC network, terraform/server-cloud-run runs it as a Cloud Run service; in an Azure subscription with a Container Apps environment, terraform/server-container-apps as a container app. On Docker Compose, compose/server runs it, publishing no host port unless asked. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The image is the Release's own code, so an edge from this component delivers what its `inject:` block dictates, and no declared socket narrows the mechanism: `env` entries become variables on every realization, `file` entries files everywhere but on Amazon ECS and Cloud Run, and `arg` or any other mechanism is refused by name, as no realization renders a command line. The component's own `inject:` block takes `env` entries too. A `core/secret-value` attachment's value is written into the rendered workload in the clear on every realization, and on Docker Compose so is a `core/referenced-secret` attachment's origin; no other realization delivers one, none delivers `core/random-password`, and the helm realization refuses either by name.
ОфіційнийcomputeЗалежні: 17Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/compute/server:1.1.0 - v1.0.1
Component
core/ai-tool/http-request
A tool that lets an agent make an HTTP request, for reaching a service that has no more specific tool of its own. The developer fixes the method the tool may use, defaulting to POST, and may reference a schema the request body is expected to satisfy. Fixing the method at authoring time keeps one tool from being repurposed as another: a tool declared for GET cannot be directed to write. Marked mutating, since a request may change state in the service it reaches. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between calls, and on Docker Compose as one service that runs while the stack runs. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 3Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-tool/http-request:1.0.1 - v1.1.0
Component
core/ai-tool/mcp
A tool that attaches a Model Context Protocol server to an agent, so the tools that server publishes become available to the agent without each one being declared separately in the catalog. The server is reached over HTTP, and the software behind it is the Release's, so a server can be wired to an agent without a catalog entity of its own. Marked mutating, because what an attached server exposes is decided by that server rather than by this declaration. Every realization publishes a host, port, endpoint, protocol and credential header, so an agent wires to the tool the same way whichever one answers it. Which realization to reach for follows how the server is published. Where it ships as a container image serving the protocol over HTTP, reach for helm/mcp on Kubernetes or compose/mcp for local development, which run the image infra.image names. Where it is an npm or PyPI package speaking only stdio, reach for helm/mcp-stdio-bridge on Kubernetes or compose/mcp-stdio-bridge for local development, which run the command line infra.stdio_command names inside a pinned bridge serving it over HTTP. Where someone else already runs it, reach for helm/mcp-remote or compose/mcp-remote, which run nothing and publish the address infra.endpoint names. A server's own credentials arrive through the server_env slot, on the realizations that run a server, helm/mcp aside. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, wherever a server runs; the mcp-remote realizations run nothing and refuse one. helm/mcp and compose/mcp deliver the component's own `inject:` block; the others refuse it.
Офіційнийai-toolЗалежні: 30Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-tool/mcp:1.1.0 - v1.0.1
Component
core/ai-tool/retrieval
A read-only tool that fetches reference material by similarity, so an agent can ground an answer in the documents its store holds rather than only in what the model already holds. It declares a required socket onto a datastore whose engine serves similarity search — vector, semantic-memory or search — so the store is resolved in the topology rather than configured in a prompt. The developer sets how many matches a lookup returns, defaulting to five. Its effect is pinned read-only in its own schema. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it, and on Docker Compose as one service that runs while the stack runs. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The store's address reaches the tool's code under default variable names; an `inject:` block on the connection delivers instead, as variables or files, the names that code reads. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 3Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-tool/retrieval:1.0.1 - v1.1.0
Component
core/ai-agent/langflow
The Langflow visual flow-builder, run in-cluster as the image a Release names. Deploying it gives the browser editor and the endpoint serving its flows, on a volume holding them and their encryption key. This component exposes the flow a caller invokes, the per-request tweaks that call may carry, and how the editor is reached: signing in is required by default, so the seeded administrator's password is attached or the build halts, while an Aggregate wanting the editor open to anyone says so. A flow authored in the Aggregate is loaded at each start on both targets, which an image at Langflow 1.5.1 or older does only with signing in off. The catalog pins no image: the Release names it in `infra.image`, and the build halts until it does. Each vector store wired into the optional `vector_store` socket arrives under variables named for the connection's alias: its address as AGGREGA_VECTOR_STORE_<ALIAS>_URL, and the connection's credential as AGGREGA_VECTOR_STORE_<ALIAS>_CREDENTIAL. A connection writing an `inject:` block receives the names it dictates instead, read by a flow as global variables; its credential stays. Each tool wired into the optional `tools` socket, a Model Context Protocol server reached over HTTP, is written into the flow's MCP Tools node named for the connection's alias. With `values.metrics_enabled` true Langflow's metrics are served at /metrics on port 9090, which checks no caller, and published as `metrics_endpoint`. On Kubernetes, reach for `langflow`, the workload synthesis, where the build must fetch nothing or the metrics are wanted, and for `langflow-chart`, rendering Langflow's own chart, which refuses them, where a chart repository is reachable at build time.
Офіційнийai-agentЗалежні: 29Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/langflow:1.1.0 - v1.1.0
Component
core/ai-agent/langgraph
A LangGraph agent, run in-cluster as an application image the Release supplies. LangGraph models an agent as a graph whose nodes resume from persisted state rather than as a single call, which is why a datastore is a required topology edge here rather than an option: the store wired into `checkpointer_db` contributes its host and port, and its user and database where it publishes them; a credential attached on that connection contributes its password; and the store is where a run's state survives between steps. What this component exposes is the graph entrypoint inside the application image and the recursion ceiling that stops a graph which is not converging. LangGraph is published as a library an application is built around, so there is no upstream image to pin: the Release supplies its application in `infra.image`. On Kubernetes, until it does, the build halts naming the container image; on Docker Compose the image defaults to the catalog's own quickstart agent, whose own code reads the model's default variables and the store's host and port, and no other, and it is published on 127.0.0.1 only where `infra.published_port` asks. The variables each connection is delivered under are defaults: an `inject:` block on the connection replaces them with the names the application reads, as variables or as files, and a checkpointer connection writing none refuses a store served over TLS alone or over HTTPS, since the application would be handed its host and port and not its protocol. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/langgraph:1.1.0 - v1.2.0
Component
core/ai-agent/librechat
A self-hosted LibreChat installation — the browser platform a person holds conversations in, and the tier that turns a bare model endpoint into a working assistant surface. It reaches its model through the `model` uses-socket and adds agents an author assembles without writing code, tool invocation over the tool servers an administrator registers, a code interpreter, and retrieval over files a user uploads. Every agent run, a delegated child run included, is entered from a conversation a person started: the platform runs no background loop of its own. Its accounts, conversations, agents and stored per-user credentials live in a document store wired through the `backend` socket, the system of record; only uploaded files sit on a volume of its own. A store that requires a sign-in, as FerretDB does by default, takes the user and password from the credential attached on that connection. The two token-signing secrets and the credential-encryption pair decide what its sign-ins and stored credentials are worth, and each has a slot of its own. None of the four has a default, and the build halts until each slot is filled. On Kubernetes, reach for `librechat-chart`, rendering the project's own chart, where a chart repository is reachable at build time, and for `librechat`, the workload synthesis, where the build must fetch nothing. With `values.metrics_enabled` true each Recipe turns on the platform's Prometheus metrics, at /metrics on its port and answered only to a bearer token equal to the secret attached to metrics_secret, and publishes the address as `metrics_endpoint`; a scrape presents that secret on its connection's `credential` slot.
Офіційнийai-agentЗалежні: 29Версії: 3Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/librechat:1.2.0 - v1.1.0
Component
core/ai-agent/n8n
The n8n automation platform, run in-cluster as the image a Release names. n8n runs workflows built from connected nodes and started by a webhook, a schedule or an event, so deploying it gives the editor and the webhook endpoint together, with a model reachable from whichever node needs one. What this component exposes is the workflow a caller invokes and, optionally, the workflow definition itself: supplied in the Aggregate, it is imported before the platform starts and on every start, so the platform runs the workflow as authored, whatever its data directory holds. The catalog pins no image: the Release names it in `infra.image`, and the build halts naming the container image until it does. The model, and each tool and store wired into the optional `tools` and `vector_store` sockets, reach the platform as environment variables, each credential by reference; each socket's comments and slots name them. A connection writing an `inject:` block receives the names it dictates instead of those its target's values fill; its credentials stay. A workflow reads these through n8n's $env only where the Aggregate sets `values.allow_env_access_in_nodes` to true. By default the Recipes keep n8n's block on environment access in expressions and the Code node, and lifting it lets every workflow read every variable of the platform's environment, the delivered credentials included. On Kubernetes, reach for `n8n-chart`, rendering n8n's own chart, where a chart repository is reachable at build time, and for `n8n`, the workload synthesis, where the build must fetch nothing. Of the two, only `n8n-chart` keeps the data directory on a volume that survives a restart, sized by `infra.storage.size`.
Офіційнийai-agentЗалежні: 29Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/n8n:1.1.0 - v1.1.0
Component
core/ai-agent/dify
The Dify API and console server — the control plane of a self-hosted Dify installation. It serves the console API the browser front end calls, the published application API, and the file endpoints, and it owns the platform's relational model: workspaces, applications, workflow definitions, datasets and the encrypted model-provider credentials. It stores that model in a PostgreSQL backend wired through the `backend` uses-socket, keeps sessions and the Celery task queue in a Redis-compatible cache wired through `cache`, and reads and writes dataset embeddings in a vector datastore wired through `vector_store`. Three sibling services complete the platform and reach it through their own sockets: code nodes execute in the sandbox wired through `sandbox`, model and tool plugins run in the plugin runtime wired through `plugin_daemon`, and outbound requests to addresses a user supplied are routed through the forward proxy wired through `ssrf_proxy`, so a workflow cannot reach a private address directly. Holds no durable state of its own — the backend, the cache and the vector datastore are the systems of record — so it renders as a stateless server. Exports its host, port and protocol to consumers via the uses topology. An application written in Dify's own export format, set as app_definition, is imported into the installation at deploy under the owner account the operator names. By default the server sends nothing to Dify's vendor: infra.telemetry_enabled turns on its usage events, update check, remote template lists and Creators Platform, and infra.marketplace_enabled its plugin marketplace, which model-provider plugins such as OpenAI's install from.
Офіційнийai-agentЗалежні: 27Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/dify:1.1.0 - v1.1.0
Component
core/ai-agent/dify-worker
The Dify task worker — the Celery consumer that runs everything a Dify installation does off the request path: indexing uploaded documents into the vector datastore, executing scheduled and queued workflow runs, generating dataset segments, and delivering mail and webhooks. It takes work from the Redis-compatible broker wired through the `cache` uses-socket, reads and writes the platform's relational model in the PostgreSQL backend wired through `backend`, and writes embeddings into the vector datastore wired through `vector_store`. Workflow code nodes it executes run in the sandbox wired through `sandbox`, and outbound requests to addresses a user supplied are routed through the forward proxy wired through `ssrf_proxy`. It serves no port and nothing connects to it — work reaches it through the broker — so it publishes no provides contract and renders as a stateless worker tier. By default nothing reaches Dify's vendor: infra.telemetry_enabled turns on its daily usage event, and infra.marketplace_enabled its daily plugin upgrade check against the plugin marketplace.
Офіційнийai-agentЗалежні: 20Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/dify-worker:1.1.0 - v1.1.0
Component
core/ai-agent/flowise
The Flowise visual agent-builder, run in-cluster as the image a Release names. Flowise is a browser editor in which a chatflow is assembled from nodes and served over HTTP, so deploying it gives the editor and the endpoint together, backed by a volume holding the chatflow database and the key encrypting the editor's credentials. What this component exposes is the chatflow a caller invokes and the per-request overrides that call may carry; the model wired into the `model` socket arrives as the endpoint an OpenAI-compatible client reads, while the chatflow's own node chooses the model it calls. The catalog pins no image for it: which Flowise build to run is named by the Release, and until that Release sets `infra.image` the build halts naming the container image. A chatflow set as flow_definition is seeded at deploy, on both targets, under the owner account the operator names. Each tool wired into the optional `tools` socket, such as a Model Context Protocol server reached over HTTP, and each vector store wired into the optional `vector_store` socket reach the platform under variables named for the connection's alias, as each socket's comments set out: AGGREGA_TOOL_<ALIAS>_URL or AGGREGA_VECTOR_STORE_<ALIAS>_URL, their parts and the connection's credential, by reference. A chatflow reads each through a Runtime variable of the same name created on Flowise's Variables page, as $vars.<name>; nothing creates one at startup. A connection writing an `inject:` block receives the names it dictates instead of its address and part variables, read the same way; its credentials stay.
Офіційнийai-agentЗалежні: 28Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/flowise:1.1.0 - v1.1.0
Component
core/ai-agent/google-adk
An agent built on Google's Agent Development Kit (ADK), run in-cluster as an application image the Release supplies. ADK is a library an application is built around, so the catalog pins no image; the build halts naming `infra.image` until a Release sets it. It runs as a server on `infra.port`, 8000 by default like ADK's own API server: the REST API of `adk api_server` or of an application built on ADK's `get_fast_api_app()`, with `/list-apps`, session routes under `/apps/`, `/run` and `/run_sse`. The image listens on every interface, not `adk api_server`'s default 127.0.0.1, and receives the port as PORT, which ADK's container guide reads. ADK documents those endpoints as unauthenticated, for a trusted network: the component publishes them to its peers on the cluster or compose network, as host, port, endpoint and protocol, and on the host's 127.0.0.1 only where `infra.published_port` asks; attaching core/public-ingress or core/http-route halts the build until `infra.routes_authenticated` affirms the application authenticates them. The `model` socket takes a model, never a gateway, carrying `model.aggrega.ai/api: gemini`, such as `core/ai-model/gemini`, and delivers the connection's credential, by reference, as GOOGLE_API_KEY, which ADK's Gemini client sends to the Gemini API; the agent's code names the model. Each tool wired to the optional `tools` socket, such as an HTTP Model Context Protocol server, reaches the application as AGGREGA_TOOL_<ALIAS>_URL, beside any credential its connection carries. An `inject:` block on any connection, or on the component itself, delivers names the application reads, as variables or files, a tools connection's in place of the tool's address and header.
Офіційнийai-agentЗалежні: 28Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/google-adk:1.1.0 - v1.1.0
Component
core/ai-agent/agent-framework
An agent built on Microsoft Agent Framework, run in-cluster as an application image the Release supplies. Agent Framework is a library an application is built around: it serves nothing itself, and an application that serves HTTP does so through a web server of its own. So there is no upstream image to pin: the catalog supplies none, and until a Release sets `infra.image` the build halts naming the container image and renders nothing. Deploying it gives that application as a worker wired to a model serving the OpenAI API. It publishes no address to its peers; on Docker Compose the application is published on 127.0.0.1 only where `infra.published_port` asks. The Recipe delivers the model's endpoint as OPENAI_BASE_URL, its model name as OPENAI_MODEL and the connection's credential as OPENAI_API_KEY. These are the variables Agent Framework's Python OpenAI clients read, each taking OPENAI_MODEL where its own model variable is unset; a .NET application reads its configuration in its own code, where the same variables are available to it. Agent Framework installs no telemetry exporter, so it sends its OpenTelemetry data nowhere unless the application's code configures one. Each tool wired into the optional `tools` socket, such as a Model Context Protocol server reached over HTTP, reaches the application under variables named for the connection's alias, its address as AGGREGA_TOOL_<ALIAS>_URL, which the application hands to its MCP tool in its own code. An `inject:` block on a connection, or on the component itself, delivers the names the application reads, as variables or files: on a tools connection in place of the tool's address and header, elsewhere beside the variables Agent Framework reads.
Офіційнийai-agentЗалежні: 21Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/agent-framework:1.1.0 - v1.1.0
Component
core/ai-agent/crewai
A CrewAI agent, run in-cluster as an application image the Release supplies. CrewAI casts each agent as a member of a crew with a stated part to play, so what this component exposes is that part — the role the agent occupies, the goal it is working toward and the backstory it reasons from — together with whether it may hand work to another member and how much of its reasoning it narrates. Deploying it gives that agent as a worker wired to a model: the component on the `model` socket publishes its endpoint and model name, and the Recipe delivers them as the environment an OpenAI-compatible client reads, so one image runs against any model in the catalog without an endpoint in its code. The catalog supplies no image for it — CrewAI is published as a library an application is built around, so there is no upstream image to pin — and until a Release sets `infra.image` the build halts naming the container image and renders nothing. On Docker Compose the application is published on the host's 127.0.0.1 only where `infra.published_port` asks, forwarding to `infra.port`, the port it listens on. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, beside the variables the framework and its client read. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/crewai:1.1.0 - v1.1.0
Component
core/ai-agent/dify-web
The Dify front end — the server-rendered web application behind the console and the published application pages. It renders the workflow editor, the dataset and application management screens, and the chat and completion surfaces an end user reaches, and it calls the API server for every piece of data it shows: the API server's in-cluster address arrives on the `api` uses-socket and becomes the base URL the server-side rendering path uses. Holds no state of its own and owns no datastore, so it renders as a stateless server. Exports its host, port and protocol to consumers via the uses topology, so an ingress tier can route the browser path to it. The console page, in the user's browser, loads an icon from assets.dify.ai and a star count from ungh.cc, and no setting turns either off.
Офіційнийai-agentЗалежні: 27Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/dify-web:1.1.0 - v1.1.0
Component
core/ai-agent/dify-worker-beat
The Dify task scheduler — the Celery beat process that puts the platform's periodic work on the queue at its due time: dataset re-indexing sweeps, expired-file and orphaned-document cleanup, workspace usage rollups, and scheduled application runs. It reads the schedule from the platform's relational model in the PostgreSQL backend wired through the `backend` uses-socket and enqueues onto the Redis-compatible broker wired through `cache`, where the task worker picks the work up; it never runs a task itself. Exactly one instance runs — two schedulers on one queue would enqueue every periodic task twice — so it renders as a single-replica stateless worker. It serves no port and nothing connects to it, so it publishes no provides contract. By default it schedules nothing that reaches Dify's vendor: infra.telemetry_enabled turns on the daily usage event, and infra.marketplace_enabled the daily plugin upgrade check.
Офіційнийai-agentЗалежні: 21Версії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/component/core/ai-agent/dify-worker-beat:1.1.0 - v1.0.1
Aggregate
core/dify
A self-contained Dify installation packaged as a reusable module: the platform on which an application author builds retrieval-augmented assistants and agent workflows from a browser. It composes ten single-workload components: an API and console server, a task worker, a task scheduler, a web front end, a code sandbox, a plugin runtime, an egress proxy, a PostgreSQL backend, a Redis-compatible cache, and a Weaviate vector datastore. Every address a workflow fetches leaves through the proxy, which refuses private destinations. Five credentials are module parameters an importing Aggregate passes down, each attached through core/referenced-secret. secret_key, the application secret, has no default, so an importer passes it down and its Release fulfills it, or the build halts naming the parameter; the other four have development defaults every reader holds, for evaluation only. The cache and the vector datastore run without authentication. Its app_definition parameter carries an application in Dify's export format, imported at deploy. A Release filling it sets infra.owner_email and attaches owner_password on the api component, or the build halts. Realized on helm and on compose. The Manifest of every Release importing it maps core/referenced-secret, or the build halts, and maps the backend to a PostgreSQL Recipe other than postgres-cloudnative-pg. On compose, infra.published_port on its api and console components publishes the API and the front end on 127.0.0.1. It ships no reverse proxy, no agent shell-workspace tier and no real-time workflow collaboration, and it wires no shared file storage, so document upload and dataset indexing need a storage backend this module does not compose.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/aggregate/core/dify:1.0.1 - v1.0.1
Aggregate
core/ferretdb
A self-contained document-database stack: a MongoDB-wire-compatible proxy in front of a PostgreSQL backend with the DocumentDB extension. Packaged as a reusable module composing two single-workload components — the proxy and its backend — wired by an internal edge that carries the backend's host, port, user, and database into the proxy's connection. The proxy delegates authentication to the backend, so both share one password; the module uses a shared inline development default. The module exports the proxy as `mongo`, the document-database endpoint consumers connect to. The module attaches no credential Feature, so its two components run on the inline development password their Recipes carry, which every reader of this catalog holds, and the module is for evaluation and local development. An Aggregate whose Release needs a real password composes the proxy and its backend directly, rather than importing this module, and attaches a credential Feature to the credential slot of each.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/aggregate/core/ferretdb:1.0.1 - v1.0.1
Aggregate
core/kong
A self-contained API-gateway stack: a database-backed gateway in front of a PostgreSQL backend. Packaged as a reusable module composing two single-workload components — the gateway and its backend — wired by an internal edge that carries the backend's host, port, user, and database into the gateway's connection. The gateway provisions its database schema through a migrations step that runs to completion before it serves. The gateway and the backend share one password; the module uses a shared inline development default. The module exports the gateway as `proxy`, the API endpoint consumers connect to. The module attaches no credential Feature, so its two components run on the inline development password their Recipes carry, which every reader of this catalog holds, and the module is for evaluation and local development. An Aggregate whose Release needs a real password composes the gateway and its backend directly, rather than importing this module, and attaches a credential Feature to the credential slot of each.
ОфіційнийВерсії: 2Опубліковано 2026-11-15oci://hub.aggrega.ai/aggregate/core/kong:1.0.1 - v1.0.0
ArtifactLibrary
ComposeInjectFiles
The compose configs that carry the files a component's inject blocks dictate, as one bundle compose Recipes import, so every file entry reaches a Docker Compose service the same way whichever of them renders it. Configs writes one compose config per file entry a connection's inject block, or the component's own, dictates: numbered in path order and named after the component instance, each holding the entry's content with every $ doubled, so the file holds the value itself rather than a variable Docker Compose would substitute. It writes nothing while no file is dictated, so a build with none renders what it would without the bundle. The importing Recipe mounts each config by name from its own workload, which a bundle cannot write, so it declares the name's prefix as the local inject_file_config_prefix, and the bundle stops the build when that local is missing or differs from its own.
ОфіційнийОпубліковано 2026-11-15oci://hub.aggrega.ai/artifactlibrary/composeinjectfiles:1.0.0 - v0.2.4
Plugin
moo
This Aggrega has Super Cow Powers.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/plugin/cli/moo:0.2.4 - v1.0.0
Plugin
litellm-gateway-config
Write runtime controls into an AI gateway's rendered configuration. The gateway's realization emits its configuration as a text block inside a ConfigMap, and the proxy reads that file through its --config flag; this step selects that entry by the routing table it carries, writes what its feature-recipe step asked for into the section it named, and replaces the entry. A section of named settings is written with settings; a section that is a list is written with entries, and identifiedBy names the key inside an entry that identifies it, so an entry the section already carries is replaced rather than duplicated whatever position it sits at. Asking for the shape a section does not hold is refused rather than written. Every line the request does not touch is carried through unchanged, so the routing table keeps its authored order and formatting. Compatible with the helm provider, whose Workload synthesis emits the documents it patches. It ships in neither the compiler nor a catalog, so it is reached through a registry. For a Release whose build includes a Feature realized through this step, `aggrega pull` fetches this entity and its compiled `.wasm` from the registry named by the workspace's `default_registry` setting, which must be reachable when it does, and records the content digest of each in the Release's lockfile. A later `aggrega build` that does not pull again first (no `--auto-pull`, no `auto_pull` workspace setting) reads both from the cache the pull filled, without contacting the registry, and halts if either no longer matches the digest recorded for it.
ОфіційнийhelmОпубліковано 2026-11-10oci://hub.aggrega.ai/plugin/pipeline-step/litellm-gateway-config:1.0.0 - v1.0.0
WorkloadLibrary
BaseWorkloads
Reusable Workload template fragments shared across recipes that synthesize provider-native manifests from a Workload body. A recipe consumes an entry with `workload: { $import: BaseWorkloads#/workloads/<Entry> }`; the version is supplied by the per-Release lockfile and no registry host enters the reference. Each entry is a Go template evaluated in the consuming recipe's render context (.values, .infra, .features), so a single authored shape serves every recipe that imports it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/workloadlibrary/baseworkloads:1.0.0 - v1.0.0
SchemaLibrary
LiteLLMGuardSchemas
Schema fragments for the Components that run the LiteLLM proxy behind the request guard of the LiteLLMGuard#/artifacts/RequestGuard bundle. RequestGuardInfra declares admitted_request_fields, the infra key a Release names further request fields in for the guard to forward. It is declared here once for every such Component, so what the key admits, and what it cannot re-open, is stated in one place wherever the guard runs.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/schemalibrary/litellmguardschemas:1.0.0 - v1.0.0
SchemaLibrary
TextMaskingSchemas
Schema fragments for masking sensitive text by listed pattern and by entity type. MaskedTextParameter declares patterns and entities once, for core/data-masking, which masks what travels through a gateway, and for values.log_masking on core/gateway/litellm, which masks what that gateway writes to its logs, so a pattern or an entity type means the same thing on both.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/schemalibrary/textmaskingschemas:1.0.0 - v1.0.0
SchemaLibrary
BaseComponents
Universal base schemas shared by every Component in the core catalog. BaseComponentParameter carries the display_name + description fields every Component supports as developer-tunable documentation knobs.
ОфіційнийЗалежні: 122Опубліковано 2026-11-10oci://hub.aggrega.ai/schemalibrary/basecomponents:1.0.0 - v1.0.0
SchemaLibrary
AIDatastores
Shared schema fragments for ai-datastore Components. Carries the universal datastore base shape plus two specialised variants for the dominant operational shapes used by AI agents: vector stores (Qdrant, Milvus, Pinecone) and graph stores (Neo4j, Memgraph). Each variant composes the universal datastore base via cross-document `$ref:` and adds shape-specific fields.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/schemalibrary/aidatastores:1.0.0 - v1.0.0
Recipe
worker-ecs
The generic background-worker primitive as one Amazon ECS service on Fargate, running infra.replicas tasks of the Release's image, rendered by vendoring a Terraform module kept beside this Recipe. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/worker on a Kubernetes cluster. The cluster, subnets, security groups and execution role come from the operator, under infra.overrides, and each is required; a task role, a public IP and a log group are optional, and nothing is emitted for one left unset. Only a task definition and a service are emitted — no cluster, role, security group, log group or load balancer, each of which the operator already runs — and the container maps no port, so nothing connects to the worker. No container health check is declared, because ECS runs one inside an image this Recipe does not name. Each env entry an edge's inject block, or the component's own, dictates becomes an entry of the container's environment under that name, its value written whole, so a Terraform reference a neighbour publishes is evaluated by the engine; concurrency and queue_depth_hint arrive by default as WORKER_CONCURRENCY and WORKER_QUEUE_DEPTH_HINT, which no edge may dictate while they are the defaults. File, arg and other entries are refused by name: Fargate writes no file, and no command line is rendered. The task runs at the smallest Fargate size meeting both infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 512 MiB. The terraform-aws-modules ECS module is not sourced: its root runs services only in a cluster it creates itself, so it is not complete for a service in the operator's cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/worker-ecs:1.0.0 - v1.0.0
Recipe
worker-container-apps
The generic background-worker primitive as one Azure Container Apps app running the Release's image, through a Terraform module kept beside this Recipe. Reach for it in an Azure subscription with a Container Apps environment; for terraform/worker-ecs or cloudformation/worker-ecs in an AWS account, terraform/worker-cloud-run on Google Cloud, and helm/worker on Kubernetes. The app declares no ingress, so nothing reaches it. The environment and resource group are the operator's, each required under infra.overrides; user-assigned identities and a private registry, its password read from Key Vault by reference, are optional. Only the app is emitted — no environment, network, identity, registry, Key Vault or log workspace, each the operator's. Each env entry an edge's inject block, or the component's own, dictates becomes a container variable under that name, written whole, so a neighbour's Terraform reference is evaluated by the engine, and each file entry a secret mounted alone at the path named; concurrency and queue_depth_hint arrive by default as WORKER_CONCURRENCY and WORKER_QUEUE_DEPTH_HINT. No edge may dictate those while they are the defaults, or a name Container Apps sets itself; arg and other entries are refused by name. The app runs exactly infra.replicas replicas at the smallest Consumption size meeting infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 0.5 GiB. Azure/avm-res-app-containerapp/azurerm is not sourced: at 0.9.0, its newest, it calls Azure/avm-utl-interfaces/azure from the registry, which init fetches outside the pin.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/worker-container-apps:1.0.0 - v1.0.0
Recipe
worker-cloud-run
The generic background-worker primitive as one Cloud Run worker pool running infra.replicas instances of the Release's image, rendered by vendoring a Terraform module kept beside this Recipe. Reach for it in a Google Cloud project with a VPC network; for terraform/worker-ecs or cloudformation/worker-ecs in an AWS account, and helm/worker on Kubernetes. A worker pool is Cloud Run's resource for continuous background work: it has no endpoint or URL, the container declares no port, and it runs a fixed instance count with no autoscaling. The project, region, VPC network and service account are the operator's, each required under infra.overrides; a subnet and network tags are optional. Only the worker pool is emitted — no service account, IAM binding, network, subnet or firewall rule. Private addresses are reached by Direct VPC egress. Each env entry an edge's inject block, or the component's own, dictates becomes a container variable under that name, its value written whole, so a Terraform reference a neighbour publishes is evaluated by the engine; concurrency and queue_depth_hint arrive by default as WORKER_CONCURRENCY and WORKER_QUEUE_DEPTH_HINT. No edge may dictate those while they are the defaults, or a name Cloud Run reserves or refuses. File, arg and other entries are refused by name: the worker pool writes no file, and no command line is rendered. The container runs at the smallest worker pool size meeting infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 1 vCPU and 512 MiB. GoogleCloudPlatform/cloud-run/google is not sourced: at v0.34.2, its newest version, it declares no worker pool; only the hashicorp/google provider's resource does.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/worker-cloud-run:1.0.0 - v1.0.0
Recipe
valkey-memorystore
Valkey on Google Cloud Memorystore for Valkey, as one instance in Cluster Mode Disabled reached from the operator's VPC network, rendered by vendoring a Terraform module kept beside this Recipe. The published terraform-google-modules/memorystore/google module is not sourced, because its valkey module calls a module from the Terraform registry, which initialising it would fetch outside any pin. Reach for it where the cache is a Memorystore instance in the operator's Google Cloud project; reach for terraform/valkey-elasticache on AWS, and for valkey or valkey-chart on helm, or valkey on compose, where it runs as a container. The project, region and network are the operator's, named in infra.overrides.project_id, region and private_network, and the Recipe emits none of them. The service connects only through Private Service Connect, which the network's service connection policy for the gcp-memorystore service class deploys; that policy must exist first. The cache runs without authentication, so the private network alone gates it: the service accepts no password a Release supplies, and an attached password halts naming the slot. host publishes a reference to the instance's primary endpoint, or the DNS name in infra.overrides.host; port is 6379. With infra.client_tls true the service serves TLS alone with the instance's own certificate authority, protocol is valkey-tls, and each consumer's server_ca slot takes a Secret holding that authority under ca.crt. infra.engine_version must name a version the service offers, so the Component's default halts here.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/valkey-memorystore:1.0.0 - v1.0.0
Recipe
valkey-elasticache
Valkey on Amazon ElastiCache, as one replication group with cluster mode disabled, realized by the published ElastiCache module of the terraform-aws-modules project, terraform-aws-modules/elasticache/aws at v1.11.2, Apache-2.0. The module is fetched from its git repository at that tag and vendored into the component's own directory; the pin is the commit the tag names and the digest of its archive, recorded in the entity.lock beside this Recipe. Reach for this realization where the cache is an ElastiCache service in the operator's AWS account; reach for valkey or valkey-chart on helm, or valkey on compose, where it runs as a container. host publishes a Terraform reference to the module's primary endpoint output, which resolves for a consumer authored as a source: Recipe in the same root configuration, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the cache instead. infra.overrides.subnet_group is required. maxmemory_policy and databases are set through a parameter group. A value ElastiCache cannot apply — maxmemory, appendonly, persistence — halts the build naming the key; compute and storage are not read. With infra.client_tls true the service encrypts in transit and protocol is valkey-tls; an attached password becomes the AUTH token, and needs client_tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/valkey-elasticache:1.0.0 - v1.0.0
Recipe
s3-object-store-gcs
An S3-compatible object store on Google Cloud Storage: one bucket in the operator's project, reached over Cloud Storage's XML API, which S3 clients interoperate with, and realized by the terraform-google-modules project's published cloud-storage module, terraform-google-modules/cloud-storage/google at v12.4.1, Apache-2.0, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where a consumer speaks the S3 API with an access key to a bucket on Google Cloud; for core/storage/gcs, realized by terraform/gcs, where it uses Google's client libraries with a Google identity; for terraform/s3-object-store-amazon-s3 or cloudformation/s3-object-store-amazon-s3 in an AWS account; and for helm/s3-object-store or compose/s3-object-store where the Release runs the store itself, on SeaweedFS, in a cluster or on a Compose host. The bucket is never public: public access prevention is enforced and uniform bucket-level access is on, and no key a Release sets changes either. So an S3 client setting a bucket or object ACL is answered 400 Bad Request, and access is granted by IAM on the bucket. Cloud Storage takes no key from a bucket: each consumer brings a Cloud Storage HMAC key, created for a service account outside the build, on its own connection, and an attachment on access_key_id or secret_access_key halts naming the slot. It publishes the bucket as written, the location as region, the endpoint https://storage.googleapis.com, its host, port 443 and protocol s3. The project and location are the operator's, in infra.overrides.project_id and infra.region.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/s3-object-store-gcs:1.0.0 - v1.0.0
Recipe
s3-object-store-amazon-s3
An S3-compatible object store on Amazon S3: one general purpose bucket in the operator's AWS account and the Region infra.region names, realized by the terraform-aws-modules project's published s3-bucket module, terraform-aws-modules/s3-bucket/aws at v5.17.0, Apache-2.0, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where a consumer speaks the S3 API to a bucket in an AWS account through Terraform; for cloudformation/s3-object-store-amazon-s3 in a CloudFormation stack; for terraform/s3-object-store-gcs on Google Cloud; and for helm/s3-object-store or compose/s3-object-store where the Release runs the store itself, on SeaweedFS, in a cluster or on a Compose host. Every public access is blocked: all four S3 Block Public Access settings are on and ACLs are disabled, and no key a Release sets changes either, so S3 rejects a public ACL or a public bucket policy. S3 takes no key from a bucket: each consumer brings the access key of an IAM principal, created outside the build, on its own connection, and an attachment on access_key_id or secret_access_key halts naming the slot. No IAM user, role, key or policy is created; a bucket policy is written only where infra.overrides.policy supplies one. infra.region is required. It publishes the bucket as written, infra.region as region, the Regional endpoint https://s3.<region> under the partition's domain suffix, its host, port 443 and protocol s3. A Region outside the partitions S3 serves, or a bucket name S3 does not take, halts naming the key.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/s3-object-store-amazon-s3:1.0.0 - v1.0.0
Recipe
redis-elasticache
Redis OSS on Amazon ElastiCache, as one replication group with cluster mode disabled, realized by the published ElastiCache module of the terraform-aws-modules project, terraform-aws-modules/elasticache/aws at v1.11.2, Apache-2.0, fetched from its git repository at that tag, vendored, and pinned in the entity.lock beside this Recipe. ElastiCache runs Redis OSS up to 7.1, so infra.engine_version takes 7.1, 7.0, 6.2, 6.0, 5.0.6 or 4.0.10, read off the start of the value; from 8 on it halts naming the key, the Component's default included, so the Release sets the version on this target. ElastiCache serves vector search on Valkey alone, so vector_search is empty and a gateway's semantic cache halts on this store. Reach for this realization where the cache is an ElastiCache service in the operator's AWS account; reach for valkey-elasticache for a newer engine, or redis on helm or compose where it runs as a container. host is a Terraform reference to the module's primary endpoint, or infra.overrides.host while set; port is infra.port. infra.overrides.subnet_group is required. maxmemory_policy and databases are set through a parameter group. maxmemory, appendonly, persistence, infra.overrides.image, and databases below 6.0 halt naming the key; compute and storage are not read. With infra.client_tls the service encrypts in transit and protocol is redis-tls; an attached password becomes the AUTH token, and needs client_tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/redis-elasticache:1.0.0 - v1.0.0
Recipe
redis-azure-managed
Redis on Azure Managed Redis, Microsoft's service running Redis Enterprise, offered under Microsoft's terms of service: one instance with its one database, rendered as Terraform for the azurerm provider. The published Azure/avm-res-cache-redisenterprise/azurerm module is not sourced, because it is not complete: it takes no persistence setting and no access-key setting, and leaves out two eviction policies the service offers. So this Recipe writes the body. Reach for it where the cache is an Azure Managed Redis instance in the operator's Azure subscription; reach for the ElastiCache Recipes on AWS, and redis on helm or compose for a container. The resource group exists beforehand; the Release names it and the region in infra.overrides, and the SKU sizes the cache. The service serves port 10000 and authenticates every client, with Microsoft Entra ID or with an access key it generates once access keys are turned on; it takes no password, so an attached one halts. It runs the Redis version it chooses, so infra.engine_version is not read, and the cache is created with no module, so vector_search is empty. host is a Terraform reference to the instance's host name, or infra.overrides.host while set; port is 10000. With infra.client_tls the service serves TLS alone and protocol is redis-tls; off, it serves plaintext, which needs access keys. maxmemory, databases, another port, infra.overrides.image and a server certificate halt naming the key; compute and storage are not read. It needs hashicorp/azurerm 5, so it goes in a different terraform boundary from terraform/postgres-flexible-server.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/redis-azure-managed:1.0.0 - v1.0.0
Recipe
rabbitmq-amazon-mq
RabbitMQ on Amazon MQ, as one broker in the operator's VPC, rendered by vendoring a Terraform module kept beside this Recipe. The published cloudposse/mq-broker/aws module is not sourced, because its root calls two more modules from the Terraform registry, which initialising it would fetch outside any pin. Reach for it where the broker is an Amazon MQ broker in the operator's AWS account; reach for rabbitmq or rabbitmq-cluster on helm, or rabbitmq on compose, where it runs as a container. Amazon MQ serves AMQP over TLS alone, so port is 5671 and protocol amqps, which is what tells a consumer to connect over TLS. host publishes a reference to the broker's endpoint; infra.overrides.host publishes the operator's DNS name instead, which a client checking the certificate cannot use, since the certificate names the broker's own endpoint. The password is required: with nothing attached the build halts naming the slot. It is the core/random-password attachment drawn at apply time and handed over as a Terraform reference, so the rendered files carry none, and the attachment must leave values.symbols off and keep values.length at 12 or more, as the service's password rules need. Any other credential halts. infra.overrides.subnet_ids is required. infra.engine_version must name a version the service offers, so the Component's default halts here; the container-sizing, port and controller keys, and values.metrics_enabled, halt away from the Component's defaults.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/rabbitmq-amazon-mq:1.0.0 - v1.0.0
Recipe
postgres-rds
PostgreSQL via AWS RDS, rendered as an authored Terraform body, main.tf. The managed-service realization: the operator's ha sizing maps onto the instance's native multi-AZ and read-replica settings, an attached core/backup onto its automated backups, and RDS owns the master password, in AWS Secrets Manager: a credential on the password slot other than core/generated-secret, or any on app_user_password, halts the build. host publishes a Terraform reference to the instance's hostname, an output of this component's module, and port publishes the port infra.port configures. The reference resolves only for a consumer authored as a source: Recipe in the same Terraform root configuration; a consumer on another target receives it as text. It names this component's module, so it resolves only where the Aggregate the Release deploys declares this component, not where it is composed in from another Aggregate. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead, which a consumer on any target can use. Pointing that name at the instance is the operator's. Reach for this realization where nothing may be fetched or where the instance needs read replicas; reach for terraform/postgres-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module, where the pull may fetch it. A value RDS cannot apply — an encoding, a locale, init scripts, a raw configuration file — halts the build naming the key; compute.cpu and compute.memory are not read. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/postgres-rds:1.0.0 - v1.0.0
Recipe
postgres-rds-terraform-aws-modules
PostgreSQL via AWS RDS, realized by the published RDS module of the terraform-aws-modules project, terraform-aws-modules/rds/aws at v7.2.2, Apache-2.0. The module is fetched from its git repository at that tag and vendored into the component's own directory; the pin is the git commit the tag names and the digest of its archive, recorded in the entity.lock beside this Recipe. RDS owns the master password, in AWS Secrets Manager: a credential on the password slot other than core/generated-secret, or any on app_user_password, halts the build. Reach for this realization where the pull may fetch the module from its git repository; reach for terraform/postgres-rds, which authors its own body and fetches nothing, where nothing may be fetched or where the instance needs read replicas, which one call of this module does not provision. host publishes a Terraform reference to the module's db_instance_address output, which resolves for a consumer authored as a source: Recipe in the same root configuration, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the instance instead. infra.overrides.subnet_group is required. A value RDS cannot apply — an encoding, a locale, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu and compute.memory are not read, because the instance class sizes the instance. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/postgres-rds-terraform-aws-modules:1.0.0 - v1.0.0
Recipe
postgres-flexible-server
PostgreSQL on Azure Database for PostgreSQL flexible server, one server with private access in a delegated subnet of the operator's virtual network, realized by Microsoft's Azure Verified Modules module Azure/avm-res-dbforpostgresql-flexibleserver/azurerm at v0.2.3, MIT, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where the database is a flexible server in the operator's Azure subscription; reach for the RDS Recipes on AWS, terraform/postgres-cloud-sql on Google Cloud, and the helm or compose Recipes for a container. It needs hashicorp/azurerm 4 and terraform/azure-api-management azurerm 5, so the two go in different terraform boundaries. The resource group, region, delegated subnet and private DNS zone are the operator's, named in infra.overrides, and none is emitted: other servers share the subnet and the zone. The module deploys a telemetry resource by default, reporting the subscription, region and module to Microsoft; infra.overrides.enable_telemetry: false turns it off. The administrator values.user names takes the core/random-password attachment on the password slot as a Terraform reference, so the rendered files carry no password; with nothing attached, or another credential, the build halts naming the slot. host publishes a reference to the server's FQDN, or infra.overrides.host; port is 5432. Init scripts, a configuration file, replicas, another port, a disk type and a storage size the service does not offer halt naming the key; compute is not read, because infra.overrides.sku_name sizes the server. Point-in-time recovery is always on, so infra.point_in_time_recovery: false halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/postgres-flexible-server:1.0.0 - v1.0.0
Recipe
postgres-cloud-sql
PostgreSQL on Google Cloud SQL, as one instance with a private IP address in the operator's VPC network, realized by the postgresql module of the terraform-google-modules project's published Cloud SQL module, terraform-google-modules/sql-db/google at v28.3.0, Apache-2.0, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where the database is a Cloud SQL instance in the operator's Google Cloud project, reached over a VPC network that already has private services access to Cloud SQL; on AWS, reach for the terraform/postgres-rds Recipes, and for the helm or compose Recipes where PostgreSQL runs as a container. The project, region and network are the operator's, named in infra.overrides.project_id, region and private_network, and the Recipe emits none of them. The instance gets no public address. Point-in-time recovery is on, as Cloud SQL's high availability needs, unless the core/backup attachment sets infra.point_in_time_recovery to false, which halts beside infra.ha.enabled. The user values.user names logs in with the core/random-password attachment on the password slot, handed over as a Terraform reference, or, with none attached, a password the module draws; either way the rendered files carry no password, and another credential halts naming the slot. host publishes a reference to the instance's private address, or the DNS name in infra.overrides.host; port is 5432, the one port the service serves. An encoding or locale the service cannot set, init scripts, a raw configuration file and another infra.port halt naming the key; compute.cpu and compute.memory are not read, because infra.overrides.tier sizes the instance.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/postgres-cloud-sql:1.0.0 - v1.0.0
Recipe
opensearch-service
OpenSearch on Amazon OpenSearch Service, as one domain of one data node in the operator's VPC, through the terraform-aws-modules project's published module terraform-aws-modules/opensearch/aws at v2.11.1, Apache-2.0, vendored into the component's directory; entity.lock pins the tag's commit and archive digest. Reach for it where the search engine is an OpenSearch Service domain in the operator's AWS account; reach for helm/opensearch-chart or compose/opensearch where it runs as a container. The domain serves HTTPS alone on 443, with fine-grained access control and one account, admin, whose password is the core/random-password attachment drawn at apply time and handed over as a Terraform reference, so the rendered files carry none; with no password attached, or another credential, the build halts naming the slot. The attachment must set values.symbols true, so the password holds every character class. host publishes a reference to the domain's endpoint, port 443, protocol https, endpoint https://<host> and user admin. infra.overrides.host publishes the operator's DNS name instead and names the domain's custom endpoint, presenting the certificate infra.overrides.custom_endpoint_certificate_arn names, so a client checking the certificate reaches the domain by it; either key alone halts, as does a name the endpoint cannot take, such as an IP literal. infra.overrides.subnet_ids is required. infra.engine_version must name a version the service offers and infra.storage.size 10Gi or more, so both Component defaults halt here; an infra.port other than 9200 halts, and so do an io1 volume and any instance or volume type the service cannot serve this domain on.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/opensearch-service:1.0.0 - v1.0.0
Recipe
mysql-rds
MySQL via AWS RDS, rendered as a Terraform HCL body. Authored as a full-body artifact named main.tf that emits verbatim as a standalone file alongside any other artifacts contributed at the same boundary. host publishes a Terraform reference to the instance's hostname, an output of this component's module, and port publishes the port infra.port configures. The reference resolves only for a consumer authored as a source: Recipe in the same Terraform root configuration; a consumer on another target receives it as text. It names this component's module, so it resolves only where the Aggregate the Release deploys declares this component, not where it is composed in from another Aggregate. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead, which a consumer on any target can use. Pointing that name at the instance is the operator's. Reach for this realization where nothing may be fetched or where the instance needs read replicas; reach for terraform/mysql-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file — halts the build naming the key; compute.cpu, compute.memory, charset and collation are not read. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mysql-rds:1.0.0 - v1.0.0
Recipe
mysql-rds-terraform-aws-modules
MySQL via AWS RDS, realized by the published RDS module of the terraform-aws-modules project, terraform-aws-modules/rds/aws at v7.2.2, Apache-2.0. The module is fetched from its git repository at that tag and vendored into the component's own directory; the pin is the git commit the tag names and the digest of its archive, recorded in the entity.lock beside this Recipe. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. Reach for this realization where the pull may fetch the module from its git repository; reach for terraform/mysql-rds, which authors its own body and fetches nothing, where nothing may be fetched or where the instance needs read replicas, which one call of this module does not provision. Unlike the authored Recipes, it sets charset and collation, in the parameter group the module creates for the instance. host publishes a Terraform reference to the module's db_instance_address output, which resolves for a consumer authored as a source: Recipe in the same root configuration, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the instance instead. infra.overrides.subnet_group is required. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu and compute.memory are not read, because the instance class sizes the instance. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mysql-rds-terraform-aws-modules:1.0.0 - v1.0.0
Recipe
mysql-rds-module
MySQL via AWS RDS, rendered by vendoring a local Terraform module and supplying its input variables. host publishes a Terraform reference to the instance's hostname, an output of the vendored module, and port publishes the port infra.port configures. The reference resolves only for a consumer authored as a source: Recipe in the same Terraform root configuration; a consumer on another target receives it as text. It names this component's module, so it resolves only where the Aggregate the Release deploys declares this component, not where it is composed in from another Aggregate. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead, which a consumer on any target can use. The Recipe emits nothing for it — no DNS record, no alias — so pointing the name at the instance is the operator's. Reach for this realization where nothing may be fetched; reach for terraform/mysql-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository, and for terraform/mysql-rds where the instance needs read replicas. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute, charset and collation are not read. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mysql-rds-module:1.0.0 - v1.0.0
Recipe
mysql-cloud-sql
MySQL on Google Cloud SQL, as one instance with a private IP address in the operator's VPC network, realized by the mysql module of the terraform-google-modules project's published Cloud SQL module, terraform-google-modules/sql-db/google at v28.3.0, Apache-2.0, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where the database is a Cloud SQL instance in the operator's Google Cloud project, reached over a VPC network that already has private services access to Cloud SQL; reach for the terraform/mysql-rds Recipes on AWS, and for the helm or compose Recipes where MySQL runs as a container. The project, region and network are the operator's, named in infra.overrides.project_id, region and private_network, and the Recipe emits none of them. The instance gets no public address. Binary logging, its point-in-time recovery, is on, as high availability and read replicas need, unless the core/backup attachment sets infra.point_in_time_recovery to false, which halts beside either. Cloud SQL creates root with no password, so root_password must be the core/random-password attachment, handed over as a Terraform reference; the user values.user names, where set, logs in with the attachment on user_password, or a password the module draws. The rendered files carry no password, and another credential halts naming the slot. host publishes a reference to the instance's private address, or the DNS name in infra.overrides.host; port is 3306, the one port the service serves. Init scripts, a raw configuration file and another infra.port halt naming the key; compute.cpu and compute.memory are not read, because infra.overrides.tier sizes the instance.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mysql-cloud-sql:1.0.0 - v1.0.0
Recipe
memcached-memorystore
Memcached on Google Cloud Memorystore for Memcached, as one instance of one node in the operator's VPC network, rendered by vendoring a Terraform module kept beside this Recipe. The published terraform-google-modules/memorystore/google module is not sourced, because its memcache module calls a module from the Terraform registry, which initialising it would fetch outside any pin. Google is deprecating the service: it creates no instance in a new project after February 1, 2027, and shuts down on January 31, 2029. Reach for it where the cache is a Memorystore instance in the operator's Google Cloud project; reach for terraform/memcached-elasticache on AWS, and for memcached on helm or on compose where it runs as a container. The project, region and network are the operator's, named in infra.overrides.project_id, region and private_network, and the Recipe emits none of them: the network needs a private services access connection before the instance is created, which every service using it shares. The service offers no encryption in transit and no authentication, so the private network alone gates the cache, and infra.client_tls, a server_certificate and a password each halt naming the key or slot. host publishes a reference to the node's address, or the DNS name in infra.overrides.host; port is 11211. values.memory_limit and values.threads size the node, and max_item_size is a parameter; max_connections, an infra.replicas above 1 and another infra.port halt naming the key.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/memcached-memorystore:1.0.0 - v1.0.0
Recipe
memcached-elasticache
Memcached on Amazon ElastiCache, as one cluster of infra.replicas nodes, realized by the published ElastiCache module of the terraform-aws-modules project, terraform-aws-modules/elasticache/aws at v1.11.2, Apache-2.0. The module is fetched from its git repository at that tag and vendored into the component's own directory; the pin is the commit the tag names and the digest of its archive, recorded in the entity.lock beside this Recipe. Reach for this realization where the cache is an ElastiCache service in the operator's AWS account; reach for memcached on helm or on compose where it runs as a container. host publishes a Terraform reference to the cluster's configuration endpoint, which resolves for a consumer authored as a source: Recipe in the same root configuration, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the cache instead. infra.overrides.subnet_group is required, and infra.engine_version must name a full version ElastiCache offers, such as 1.6.22. max_item_size is set through a parameter group. A value ElastiCache cannot apply — memory_limit, max_connections, threads, a password — halts the build naming the key; compute is not read. With infra.client_tls true the service encrypts in transit and protocol is memcached-tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/memcached-elasticache:1.0.0 - v1.0.0
Recipe
mariadb-rds
MariaDB via AWS RDS, rendered as a Terraform HCL body. Authored as a full-body artifact named main.tf that emits verbatim as a standalone file alongside any other artifacts contributed at the same boundary. host publishes a Terraform reference to the instance's hostname, an output of this component's module, and port publishes the port infra.port configures. The reference resolves only for a consumer authored as a source: Recipe in the same Terraform root configuration; a consumer on another target receives it as text. It names this component's module, so it resolves only where the Aggregate the Release deploys declares this component, not where it is composed in from another Aggregate. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead, which a consumer on any target can use. Pointing that name at the instance is the operator's. Reach for this realization where nothing may be fetched or where the instance needs read replicas; reach for terraform/mariadb-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file — halts the build naming the key; compute.cpu, compute.memory, charset and collation are not read. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mariadb-rds:1.0.0 - v1.0.0
Recipe
mariadb-rds-terraform-aws-modules
MariaDB via AWS RDS, realized by the published RDS module of the terraform-aws-modules project, terraform-aws-modules/rds/aws at v7.2.2, Apache-2.0. The module is fetched from its git repository at that tag and vendored into the component's own directory; the pin is the git commit the tag names and the digest of its archive, recorded in the entity.lock beside this Recipe. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. Reach for this realization where the pull may fetch the module from its git repository; reach for terraform/mariadb-rds, which authors its own body and fetches nothing, where nothing may be fetched or where the instance needs read replicas, which one call of this module does not provision. Unlike the authored Recipes, it sets charset and collation, in the parameter group the module creates for the instance. host publishes a Terraform reference to the module's db_instance_address output, which resolves for a consumer authored as a source: Recipe in the same root configuration, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the instance instead. infra.overrides.subnet_group is required. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu and compute.memory are not read, because the instance class sizes the instance. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mariadb-rds-terraform-aws-modules:1.0.0 - v1.0.0
Recipe
mariadb-rds-module
MariaDB via AWS RDS, rendered by vendoring a local Terraform module and supplying its input variables. host publishes a Terraform reference to the instance's hostname, an output of the vendored module, and port publishes the port infra.port configures. The reference resolves only for a consumer authored as a source: Recipe in the same Terraform root configuration; a consumer on another target receives it as text. It names this component's module, so it resolves only where the Aggregate the Release deploys declares this component, not where it is composed in from another Aggregate. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead, which a consumer on any target can use. The Recipe emits nothing for it — no DNS record, no alias — so pointing the name at the instance is the operator's. Reach for this realization where nothing may be fetched; reach for terraform/mariadb-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository, and for terraform/mariadb-rds where the instance needs read replicas. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute, charset and collation are not read. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/mariadb-rds-module:1.0.0 - v1.0.0
Recipe
kafka-msk
Apache Kafka on Amazon MSK (Managed Streaming for Apache Kafka), as one provisioned cluster in KRaft mode in the operator's VPC, through the terraform-aws-modules project's published module terraform-aws-modules/msk-kafka-cluster/aws at v3.3.0, Apache-2.0, fetched from its git repository at that tag and vendored; entity.lock pins the tag's commit and archive digest. The cluster runs one broker in each client subnet, the smallest cluster MSK runs, so infra.overrides.subnet_ids names two or three subnets in distinct Availability Zones and the cluster has two or three brokers. Clients reach the brokers inside the VPC with no authentication, on the plaintext listener, port 9092, under protocol kafka, beside MSK's TLS listener on 9094; with infra.client_tls true, on that TLS listener alone, under protocol kafka-ssl, whose certificate AWS issues and default trust stores trust. A certificate attached to server_certificate halts, and so does infra.overrides.host with client_tls: that certificate does not name it. bootstrap_servers publishes a reference to the brokers' whole list, host the first broker in it, port 9092, or 9094 with client_tls; infra.overrides.host publishes the operator's DNS name instead. Reach for it where Kafka is an MSK cluster in the operator's AWS account; reach for kafka or kafka-strimzi on helm where it runs in Kubernetes. infra.engine_version takes 3.8, 3.9, 4.0 or 4.1, the Component's default included. An infra.broker_port other than 9092 halts, and so do values.metrics_enabled, values.cluster_id, infra.controller_port, infra.compute and infra.control_plane_present away from the Component's defaults: MSK decides each of them itself or has no use for it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/kafka-msk:1.0.0 - v1.0.0
Recipe
function-lambda
The generic event-driven compute primitive as one AWS Lambda function run from a container image, by the terraform-aws-modules project's Lambda module, terraform-aws-modules/lambda/aws at v8.12.0, Apache-2.0, fetched from its git repository at that tag and pinned in the entity.lock beside this Recipe. Its packaging, which runs Python on the applying machine, is off, as an image function has no package. Reach for it where the target is an AWS account; reach for helm/function-knative or helm/function-keda on Kubernetes. Only the function is emitted, with provisioned concurrency for a floor — no role, log group, trigger, URL or event source mapping: what invokes it, its role and its log group are the operator's, the last two required under infra.overrides. The image is in Amazon ECR, in the function's Region, and implements Lambda's runtime API — an AWS base image for Lambda, or a linux/amd64 image with a runtime interface client; one serving HTTP carries an adapter forwarding Lambda's invocations. An image outside ECR is refused by name. compute.memory is the function's memory, max_replicas its reserved concurrency, and a min_replicas above zero provisioned concurrency on the published version, serving invocations of that version. port, compute.cpu, control_plane_present and scaler are not read; the comment above the assertions gives each reason. Each env entry an edge or the component's own inject block dictates becomes a variable of the function's environment, its value written whole; runtime arrives by default as FUNCTION_RUNTIME. A name Lambda reserves or cannot take, or FUNCTION_RUNTIME while it is that default, is refused by name, and so are file, arg and other entries.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/function-lambda:1.0.0 - v1.0.0
Recipe
gcs
A Google Cloud Storage bucket in the operator's project, reached through Google's own API, realized by the terraform-google-modules project's published cloud-storage module, terraform-google-modules/cloud-storage/google at v12.4.1, Apache-2.0, fetched from its git repository at that tag; entity.lock pins the tag's commit and archive digest. Reach for it where a consumer uses Google's client libraries with a Google identity, or needs Cloud Storage's own storage class or versioning; for core/storage/s3-object-store, realized by terraform/s3-object-store-gcs, where it speaks the S3 API with an access key. The bucket is never public: public access prevention is enforced and uniform bucket-level access is on, and no key a Release sets changes either. So access is granted by IAM on the bucket alone, to each consumer's own Google identity outside the build, and a request setting a bucket or object ACL is answered 400 Bad Request. It publishes the bucket as written, its gs:// address as url, and the location. The project and location are the operator's, in infra.overrides.project_id and infra.location; infra.storage_class and infra.versioning set the bucket's storage class and object versioning.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/gcs:1.0.0 - v1.0.0
Recipe
external-object-store
An externally-operated object-storage bucket, realized by vendoring a local Terraform module and binding its input variables from the developer's values. The module files copy into the node's output verbatim, in that component's own directory; the values pillar becomes the arguments of the component's module block in the boundary's root configuration. This Recipe declares a source and NO valuesMapping, and that combination is what makes the sensitivity marking reach the output. Without a mapping, the argument names ARE the value leaf paths the Component declared, so the Component's `sensitive: true` on a leaf names one of the module's inputs exactly. A mapping's output keys are invented inside its template, which no declaration names, so a Recipe that declared one would render the same values with no marking at all.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/external-object-store:1.0.0 - v1.0.0
Recipe
cloudflare-ai-gateway
Creates a Cloudflare AI Gateway in the account infra.account_id names, rendered as Terraform for the Cloudflare provider, and publishes its address. Reach for it where the Release is to create the gateway and set its controls; reach for the helm or compose reference where the gateway already exists and is configured outside the Release. It writes the gateway's authentication and request-log settings from infra, its cache off and its request-rate limit as zero, and what an attached core/period-budget or core/retry-policy sets into the same resource. Terraform authenticates to Cloudflare with the API token in the environment it runs in, so the rendered configuration names no credential. Applying it is outside any build: whether the gateway then behaves as Cloudflare documents is not observed by the compiler.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/cloudflare-ai-gateway:1.0.0 - v1.0.0
Recipe
azure-openai
Creates the Azure OpenAI resource core/ai-model/azure-openai is served from, and its chat model deployment, rendered as Terraform for the azurerm provider: an account of kind OpenAI under the custom subdomain the Release names, and one model deployment named values.deployment_name serving values.model. The published Azure/avm-res-cognitiveservices-account/azurerm module is not sourced, because it calls a module from the Terraform registry, so this Recipe writes the body. Reach for it where the Release is to create the resource; reach for the helm or compose reference where the resource already exists outside the Release. The resource group exists beforehand; the Release names it, the region and the subdomain in infra.overrides. Which models, versions and deployment types a region offers, and the quota the subscription holds for them, are the subscription's and no build sees them: a model the region does not offer, or one beyond the quota, fails when Terraform applies. endpoint is https://<subdomain>.openai.azure.com, the address Azure serves a resource under its custom subdomain, so it is known before the deploy and a consumer on any target can use it; model is the deployment name. The account holds its API keys and the rendered files carry none: a consumer's credential is a key read from the resource once it exists, attached on the consumer's connection. Public network access is the service's default, every network, unless the Release turns it off. It needs hashicorp/azurerm 5, as terraform/azure-api-management does, so it goes in a different terraform boundary from terraform/postgres-flexible-server, which needs azurerm 4.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/azure-openai:1.0.0 - v1.0.0
Recipe
azure-openai-embedding
Creates the Azure OpenAI resource core/ai-model/azure-openai-embedding is served from, and its embedding model deployment, rendered as Terraform for the azurerm provider: an account of kind OpenAI under the custom subdomain the Release names, and one model deployment named values.deployment_name serving values.model. The published Azure/avm-res-cognitiveservices-account/azurerm module is not sourced, because it calls a module from the Terraform registry, so this Recipe writes the body. Reach for it where the Release is to create the resource; reach for the helm or compose reference where the resource already exists outside the Release. The resource group exists beforehand; the Release names it, the region and the subdomain in infra.overrides. Which models, versions and deployment types a region offers, and the quota the subscription holds for them, are the subscription's and no build sees them: a model the region does not offer, or one beyond the quota, fails when Terraform applies. endpoint is https://<subdomain>.openai.azure.com, the address Azure serves a resource under its custom subdomain, so it is known before the deploy and a consumer on any target can use it; model is the deployment name. The account holds its API keys and the rendered files carry none: a consumer's credential is a key read from the resource once it exists, attached on the consumer's connection. Public network access is the service's default, every network, unless the Release turns it off. It needs hashicorp/azurerm 5, as terraform/azure-api-management does, so it goes in a different terraform boundary from terraform/postgres-flexible-server, which needs azurerm 4.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/azure-openai-embedding:1.0.0 - v1.0.0
Recipe
azure-api-management
Adds an API serving a model to an Azure API Management instance that already exists, rendered as Terraform for the azurerm provider: the API, its chat completions operation, the backend it forwards to and its policy. Reach for it where the instance exists and the Release is to configure the API; reach for the helm or compose reference where the API is configured outside the Release. It never creates the instance, so a Release without one creates it outside first. The policy sends each request to the backend, removes the caller's subscription key from what it forwards, and authenticates as infra.backend_authentication says; what an attached core/semantic-caching, core/period-budget, core/prompt-injection-shield or core/retry-policy sets is written into the same policy. Terraform authenticates to Azure with the credentials in the environment it runs in, and a provider key reaches the configuration only as the identifier of the Key Vault secret holding it, so the rendered files carry no credential. The azurerm provider, which does not plan without its features block, is configured in one file beside the boundary's root configuration, shared by every component there that declares it. So a Release that removes the last of them destroys what their modules created first: once the file is gone, Terraform has no provider configuration left to destroy those resources with. It needs hashicorp/azurerm 5 and terraform/postgres-flexible-server azurerm 4, so the two go in different terraform boundaries. Applying it is outside any build: whether the API then behaves as Microsoft documents is not observed by the compiler.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/terraform/azure-api-management:1.0.0 - v1.0.0
Recipe
worker
The generic background-worker primitive on Kubernetes via the helm provider's Workload synthesis: a Deployment and nothing else. No Service and no container port are emitted, and that absence carries the operational type's meaning rather than merely accompanying it — the type describes compute with zero inbound application ports, and the emitted file list is where a reader confirms that. It is what separates this realization from the generic server primitive's, which fronts its workload with a ClusterIP Service. The developer's `concurrency` and `queue_depth_hint` values are delivered to the container under stable names by default for the application's own loop to read, because how many items a worker takes at once and how deep a backlog it expects belong to that loop rather than to any field the workload contract expresses. The component's own inject block delivers env entries too, and one reading either value alone replaces its name. No provides contract is emitted, because the component declares none: this primitive publishes nothing across an edge. What reaches the container from the component's edges is what each edge's inject block dictates, since the image is the Release's own code: each env entry a variable under that name, beside the two developer values, whose names no edge may dictate while they are the defaults, and each file entry a read-only file at the path named, from a ConfigMap. An arg entry is refused by name, as the recipe renders no command line, and so is any other mechanism. A credential published for the pod to reference, such as core/referenced-secret, halts the build by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/worker:1.0.0 - v1.0.0
Recipe
weaviate
Weaviate on Kubernetes via the helm provider's Workload synthesis, running the official upstream image from the project's own registry (no external Helm chart). Vector data is the point of this store rather than a cache of something else, so the workload renders as a StatefulSet with a persistent volume mounted at the server's data path, alongside the governing headless Service the stateful form requires and a ClusterIP Service publishing both the HTTP and the gRPC port. The rendered manifest always states its access posture rather than leaving it to a server default: anonymous access is enabled where no credential is attached, and where the optional api_key slot is attached anonymous access is turned off, API-key authentication is turned on and the key is delivered as an env secretKeyRef, so the plaintext key never appears in the rendered manifest. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). It is synthesized because Weaviate's own Helm chart fails the maintained test: the newest chart its published repository serves packages an older Weaviate than the one this Component pins. With `values.metrics_enabled` true, PROMETHEUS_MONITORING_ENABLED turns on Weaviate's metrics at port 2112, which joins the ClusterIP Service, and provides.metrics_endpoint publishes the address; false, that key is empty. The port checks no caller and also answers /tenant-activity with each collection's tenant names.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/weaviate:1.0.0 - v1.0.0
Recipe
vllm
vLLM as a self-hosted inference server on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). Renders a StatefulSet: a weights-pull initContainer resolves the model pointer into a per-replica model-cache volume, and the serving container mounts the same volume and serves the OpenAI-compatible API behind a ClusterIP Service. The model value is a deterministic pointer the initContainer resolves at pod start; the compiler never fetches weights. When the optional model_pull_token credential slot is attached the token reaches the initContainer as an env secretKeyRef, so it never appears in the rendered manifest; unattached, weights are pulled anonymously, byte-identically to a recipe with no slot. When the optional api_key slot is attached the key reaches the serving container as VLLM_API_KEY the same way, and the server requires it of every caller under /v1; unattached, it requires none. provides.host is the bare, self-contained in-cluster service name; a consumer in another namespace qualifies it on its own side. provides.metrics_endpoint names the server's metrics on the same Service while it runs one replica; above one it is empty and metrics_service names that Service for a scraper to discover every replica behind it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vllm:1.0.0 - v1.0.0
Recipe
vonage
The Vonage voice carrier as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the carrier's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. The two webhook addresses are published to consumers exactly as the operator supplied them. They are opaque pointer strings: this recipe pipes each through unchanged, and nothing here resolves an address, contacts one, or derives one from the state of a cluster. This recipe pins no artifact. It names no image, no chart and no module — the Release calls a service its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to one of the component's slots is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the carrier. All three credential slots are required here: a build leaving any of them empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vonage:1.0.0 - v1.0.0
Recipe
vllm-kubeai
vLLM realized as a KubeAI Model the KubeAI operator reconciles, rather than a directly synthesized StatefulSet — a second controller-driven realization of the same model component, selectable in the Manifest alongside the synthesis recipe and the KServe realization. The recipe emits a Model custom resource whose url is the model pointer; the operator serves it behind its OpenAI-compatible proxy and autoscales it. The model value is a deterministic pointer the operator resolves at reconcile time; the compiler never fetches weights. GPU and CPU sizing carry through resourceProfile. provides is filled from the operator's proxy, which addresses a model by its resource name, so a consumer wires to it exactly as it would to the synthesized realization. An attached api_key reaches the server through the Model's envFrom, which imports the credential's secret by name and names each variable after one of its data keys, so the credential publishes its key as VLLM_API_KEY, and the build halts naming the key when it does not. The KubeAI control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The Model is written into the namespace the Release, or else the Aggregate, names for this component through the aggrega.ai/namespace annotation, where one is set, and the operator creates the model's pods there; the metrics keys publish the pod label, port name and path a scraper discovers those pods by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vllm-kubeai:1.0.0 - v1.0.0
Recipe
vllm-kserve
vLLM realized as a KServe InferenceService the KServe controller reconciles, rather than a directly synthesized StatefulSet — the controller-driven realization of the same model component, selectable in the Manifest alongside the synthesis recipe. The recipe emits an InferenceService custom resource whose predictor resolves the model pointer as a storage URI; the controller turns it into a running serving workload and its predictor Service. The model value is a deterministic pointer the controller resolves at reconcile time; the compiler never fetches weights. GPU is requested directly on the predictor as the extended resource nvidia.com/gpu, with an optional runtime class. provides is filled from the predictor Service the controller creates, so a consumer wires to it exactly as it would to the synthesized realization. An attached api_key reaches the predictor as VLLM_API_KEY, taken from the credential's secret by name, so the key never appears in the custom resource. The KServe control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The InferenceService is written into the namespace the Release, or else the Aggregate, names for this component through the aggrega.ai/namespace annotation, where one is set. Every metrics key is published empty: the port and path of the metrics belong to the serving runtime the cluster registered, which the catalog does not know.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vllm-kserve:1.0.0 - v1.0.0
Recipe
vertex-ai
Google Vertex AI's managed model API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied regional endpoint. No credential is delivered by this recipe: consumers authenticate with ambient GCP workload identity or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vertex-ai:1.0.0 - v1.0.0
Recipe
vertex-ai-embedding
Google Vertex AI's managed embedding API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied regional endpoint. No credential is delivered by this recipe: consumers authenticate with ambient GCP workload identity or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vertex-ai-embedding:1.0.0 - v1.0.0
Recipe
vertex-ai-agent
A Vertex AI agent as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the operator-supplied regional endpoint. No credential is delivered by this recipe — consumers authenticate with ambient GCP workload identity or inject their own credentials. provides.endpoint is the regional host and provides.agent_id is the developer-selected agent.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vertex-ai-agent:1.0.0 - v1.0.0
Recipe
velero-chart
Installs Velero through the project's own Helm chart, at the chart release whose application version is the image the Component pins. The chart is official, published by the Velero maintainers; complete, carrying the custom-resource definitions, the server, its node agent and their permissions; and maintained, its newest release carrying Velero 1.18.2. It renders the server Deployment, which loads the plugin for the S3 API from an init container, the node-agent DaemonSet that copies the data of pod volumes, the custom-resource definitions, and a binding granting the server's service account cluster-admin, which Velero's chart grants by default because a backup reads and restores every kind of resource. It renders no storage location and no schedule: helm/backup-velero writes those for each core/backup attachment. The chart places its objects in the namespace the Release, or else the Aggregate, declares for the component, and Velero reads schedules only from its own namespace, so the build halts unless that namespace is velero. Its chart repository must be reachable from the build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/velero-chart:1.0.0 - v1.0.0
Recipe
valkey
Valkey on Kubernetes via the helm provider's Workload synthesis, running the official upstream valkey/valkey image (no external Helm chart). Used as an in-memory cache with no persistence, so it renders a stateless Deployment plus a ClusterIP Service and declares no storage. Authentication is off by default; when the optional password credential slot is attached the value is delivered as an env secretKeyRef and Valkey is started with --requirepass to enforce it, so the plaintext password never appears in the rendered manifest. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Its readiness and liveness probes send the cache's own PING to the port the Service targets. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `valkey-chart`, which renders the community-maintained Valkey chart, where the cache should run unprivileged on a read-only root filesystem, or express an attached password as a hashed access-control entry instead of a start-up flag. With infra.client_tls true the cache serves TLS alone on its port from the Secret attached to server_certificate, mounted as files, the probes ping over TLS, and protocol is published as valkey-tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/valkey:1.0.0 - v1.0.0
Recipe
vapi
The Vapi voice-agent platform as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the platform's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to the component's slot is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the platform. The credential slot is required here: a build leaving it empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/vapi:1.0.0 - v1.0.0
Recipe
valkey-chart
Valkey in-memory key-value cache on Kubernetes via the community-maintained Valkey Helm chart, served from that project's HTTPS chart repository. The chart is configured for the shape this Component describes: one cache process behind a ClusterIP Service, with replication, network policy, disruption budget and metrics left off, and transport security off unless infra.client_tls turns it on. With a password attached it also renders the chart's own connection check, a Pod carrying the helm test annotation that confirms an unauthenticated connection is refused; the chart offers no setting to decline it, and it is harmless as it stands — it runs the cache's pinned image, reads and writes nothing, and carries no credential material. What it carries that a hand-assembled workload does not is the cache's own start-up and security machinery: a hardened container that runs unprivileged with a read-only root filesystem, probes that run the cache's own client, and an attached password hashed on the pod from a mounted secret, so neither the password nor its hash reaches the rendered artifact. Reach for this realization where a chart repository is reachable at build time; reach for `valkey`, the workload-synthesis realization, where the build must complete with nothing fetched. The cache image and the chart version are pinned. provides.host is the bare in-cluster Service name; a consumer in another namespace qualifies it on its own side. With infra.client_tls true, the chart's transport security serves TLS alone on the operator's port from the Secret attached to server_certificate, the probes ping over TLS, and protocol is published as valkey-tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/valkey-chart:1.0.0 - v1.0.0
Recipe
twilio
The Twilio messaging and voice carrier as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the carrier's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the Release calls a service its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to one of the component's slots is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the carrier. Both credential slots are required here: a build leaving either empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/twilio:1.0.0 - v1.0.0
Recipe
tyk
Tyk Gateway on Kubernetes via the open-source stack chart the gateway's own maintainers publish, served from the project's HTTPS chart repository. The chart was chosen over a hand-authored workload because it meets all three conditions the catalog asks of a chart: it is official, published by the product's maintainers alongside the gateway itself; it is complete, carrying the gateway Deployment, its Service, the storage wiring and the directory setup the gateway needs at start; and it is maintained, its most recent release post-dating the most recent gateway release by two days as read on 2026-09-20. The chart bundles three components — the gateway, an analytics purger, and the vendor's Kubernetes controller — and only the gateway is unconditional. This Recipe deploys the gateway alone: the purger and the controller stay off, and the controller's configuration secret, which the chart emits by default even with the controller off, is turned off too, so the render carries nothing belonging to a component this Recipe does not stand up. The chart names its generated resources after the release, so fullnameOverride is bound to the release name the helm provider gives the chart, which spells the component's name in the alphabet a Kubernetes name allows, and provides.host names the Service the chart emits. It writes each core/upstream-key-pinning attachment into the gateway's environment. A cache publishing valkey-tls or redis-tls is reached through the chart's own Redis TLS setting, checked against the cache connection's server_ca, mounted, or the image's trust store; another TLS protocol halts the build. Default chart version in source.version; override via infra.source_version.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/tyk:1.0.0 - v1.0.0
Recipe
trustyai-huggingface-detector
TrustyAI's Hugging Face detector on Kubernetes via the helm provider's Workload synthesis, running the image in infra.image directly: a stateless Deployment plus a ClusterIP Service on the port the service answers on. Synthesized rather than wrapping a chart because the project publishes none. A weights-pull init container, running the same image and the Hugging Face command-line tool it carries, fetches the model named in values.model at values.model_revision into a pod-local volume before the server starts, and the server loads it from there through the MODEL_DIR variable it reads; the compiler never fetches weights. The volume lasts as long as the pod, so each new pod fetches the weights again. When the optional model_pull_token slot is attached the token reaches the pull step alone as an env secretKeyRef, so it never appears in the rendered manifest; unattached, weights are pulled anonymously, byte-identically to a Recipe with no slot. The server loads the model before it answers anything, and its health route answers only then, so readiness is asked there. The listen port is handed to the server on its command line, from the same infra.port the Service and the published address read. The service's own metrics, in the Prometheus format at /metrics on that port, are published as provides.metrics_endpoint. provides.host is the bare, self-contained in-cluster service name of the synthesized Service; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/trustyai-huggingface-detector:1.0.0 - v1.0.0
Recipe
trustyai-guardrails-gateway
TrustyAI's guardrails gateway in a model's path, on Kubernetes via the helm provider's Workload synthesis, running the image in infra.image directly: a stateless Deployment, a ClusterIP Service on the port the gateway answers on, and a ConfigMap carrying its configuration file. Synthesized rather than wrapping a chart because the project publishes none. The configuration points the gateway at the host and port the orchestrator wired on the orchestrator socket publishes, and declares one route, values.route, applying the detector that orchestrator registered — by the detector_id it publishes — to the latest message, and to the reply when values.check_output is set, with values.threshold, answering a detection with values.fallback_message. The file is mounted on its own and the GATEWAY_CONFIG variable points the gateway at it; the port reaches it through the HTTP_PORT variable it reads. The product serves no health route, so readiness and liveness are a TCP check on its port. provides.host is the bare, self-contained in-cluster service name of the synthesized Service; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/trustyai-guardrails-gateway:1.0.0 - v1.0.0
Recipe
traefik-hub-chart
Traefik Hub on Kubernetes via Traefik's own Helm chart, served from that project's HTTPS chart repository, with the chart's Traefik Hub settings and its AI gateway on. It renders the gateway's Deployment, a ClusterIP Service for its plain HTTP listener, its access control and the chart's custom-resource definitions; the chart's Kubernetes Ingress provider and its IngressClass are off. The image is the one infra.image names and no other, and the build halts naming infra.image when it is unset or names no tag. The licence token is read from the Secret the license slot names, under the data key token, and the build halts on any other key. The export of the gateway's error logs to Traefik's platform and its check for a newer release are both off; the token check against api.traefik.io, which Traefik requires, stays. Each model wired to the models socket gets a route at /<alias>/v1/chat/completions through Traefik Hub's Chat Completion middleware, written beside the chart as Traefik's own resources. Every route checks the key of each caller infra.callers declares against its hash, read by reference, and removes the key before forwarding. The same file writes a core/period-budget, core/prompt-deny-patterns, core/semantic-caching, core/data-masking, core/fallback-chain or core/retry-policy attachment, each FeatureRecipe saying how; the first four halt the build until infra.ai_gateway_entitled affirms the licence's AI gateway. No build observes that the gateway enforces any of them. A budget's store publishing valkey-tls or redis-tls is reached over TLS, checked against the cache connection's server_ca or the system's authorities.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/traefik-hub-chart:1.0.0 - v1.0.0
Recipe
timescaledb
TimescaleDB on Kubernetes via the helm provider's Workload synthesis (StatefulSet + ClusterIP Service + headless Service). provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Reach for this realization where nothing is installed in the target cluster: it needs no control plane and no chart repository, and it builds with nothing fetched. Reach for timescaledb-cloudnative-pg where the CloudNativePG control plane is already in place and the cluster can mount an image as a volume, which is how that realization delivers the extension.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/timescaledb:1.0.0 - v1.0.0
Recipe
thanos-store
Thanos Store Gateway, the read side of a Thanos metrics stack, on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service exposing the StoreAPI gRPC port and the HTTP port). The Store Gateway keeps only an ephemeral local cache of index data under a data-dir — the object store is the system of record — so it renders stateless, with no persistent volume. The endpoint the store wired to the `objstore` uses-edge publishes sets the inline object-store configuration passed on the `--objstore.config` flag: its host and port are the S3 endpoint, and its scheme sets `insecure` — true for http, false for https. A store publishing no endpoint, or one with another scheme or with a path, halts the build naming the connection. The bucket is the component's own value, and a store publishing a different bucket halts the build naming both. That configuration selects the AWS SDK credential chain, so it carries no secret. AWS_ACCESS_KEY_ID is an env secretKeyRef from the `objstore` connection's credential_id, or the s3_access_key value where it carries none; AWS_SECRET_ACCESS_KEY is one from its credential. The secret key has no default: a connection carrying no credential halts the build, naming it, unless its values set unauthenticated: true, and then AWS_SECRET_ACCESS_KEY is not emitted and the gateway takes whatever else the credential chain finds. The gateway's own metrics, in the Prometheus format at /metrics on the HTTP port, are published as provides.metrics_endpoint. provides.host is the bare in-cluster name of the synthesized Service, carrying no namespace.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/thanos-store:1.0.0 - v1.0.0
Recipe
timescaledb-cloudnative-pg
The time-series database realized as a Cluster custom resource the CloudNativePG control plane reconciles, rather than a directly synthesized StatefulSet. The product has no control plane of its own, so this realization runs it under the PostgreSQL controller this catalog already uses for core/datastore/postgres: the server image stays a stock PostgreSQL one and the time-series extension arrives as a second image the controller mounts into it. The recipe emits one Cluster naming both images and the preload list the extension needs; the controller turns it into a running server, its Services and the owner's Secret, and creates the initial database and its owning role from the developer's own values. provides.host is the read-write Service, so a consumer wires to it as it would to a synthesized realization. infra.credentials generated leaves the owner's password with the controller, and supplied points it at the Secret named in infra.credentials_secret. The Cluster is written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation, where one is set, and provides publishes the pod label, port name and path a scraper discovers the instances' metrics by. Two runtime prerequisites are declared, and the build halts on either: the control plane is installed (infra.control_plane_present), and the cluster can mount an image as a volume (infra.image_volumes_supported). Reach for this realization where that control plane is already in place and the cluster can mount an image as a volume; reach for the synthesized timescaledb recipe where it is not — that one needs nothing installed, reaches no chart repository, and builds with nothing fetched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/timescaledb-cloudnative-pg:1.0.0 - v1.0.0
Recipe
thanos-query
Thanos Query, the query front end of a Thanos metrics stack, on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the HTTP port). Query holds no local state — it reads through the Store Gateway — so it renders stateless, with no persistent volume. The Store Gateway identity resolved from the `store` uses-edge (host, port) arrives on .context.inject.env and composes the StoreAPI endpoint Query dials on its `--endpoint` flag; a component wired to `store` that publishes a protocol other than grpc, or none, halts the build, naming the connection. Query reads no object store and needs no credentials of its own. With `infra.external_base_url` set, it reaches Query as its `--alert.query-url`, the base of the links its web interface writes back to itself. Where that address carries a path, Query also takes the path as `--web.external-prefix`, so the links and redirects of its own pages carry it, and `--web.route-prefix=/` keeps it serving at its root; unset, none of these flags is emitted. Query serves its own metrics in the Prometheus format at /metrics on the HTTP port, at the root whatever that address says, published as provides.metrics_endpoint. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/thanos-query:1.0.0 - v1.0.0
Recipe
tgi
Text Generation Inference as a self-hosted inference server on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). Renders a StatefulSet: a weights-pull initContainer resolves the model pointer into a per-replica model-cache volume, and the serving container mounts the same volume and serves the API behind a ClusterIP Service. The model value is a deterministic pointer the initContainer resolves at pod start; the compiler never fetches weights. When the optional model_pull_token credential slot is attached the token reaches the initContainer as an env secretKeyRef, so it never appears in the rendered manifest; unattached, weights are pulled anonymously, byte-identically to a recipe with no slot. When the optional api_key slot is attached the key reaches the serving container as API_KEY the same way, and the server requires it of every caller on its generation routes; unattached, it requires none. provides.host is the bare, self-contained in-cluster service name; a consumer in another namespace qualifies it on its own side. provides.metrics_endpoint names the server's metrics on the same Service while it runs one replica; above one it is empty and metrics_service names that Service for a scraper to discover every replica behind it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/tgi:1.0.0 - v1.0.0
Recipe
tempo
Grafana Tempo distributed-tracing server on Kubernetes via the helm provider's Workload synthesis, running the official upstream grafana/tempo image (no external Helm chart). Tempo is driven in its single-binary shape: every subsystem — distributor, ingester, querier, compactor — runs inside one process backed by a local trace store, so it renders as one StatefulSet with a per-replica data volume, a ConfigMap carrying the Tempo config from the TempoConfig bundle, a ClusterIP Service and a Headless Service exposing the OTLP gRPC and HTTP receivers alongside the HTTP query API. The Tempo server image is pinned to a resolvable tag for reproducible renders. The developer's retention_period flows into the in-process compactor, which deletes stored trace blocks once they age past it. provides.host is the bare in-cluster name of the synthesized ClusterIP Service, with no namespace; a consumer in another namespace qualifies it on its own side. provides.port is the OTLP gRPC receiver port, the endpoint a trace producer wires to, and provides.metrics_endpoint where the server serves its own metrics, in the Prometheus format at /metrics on the HTTP port 3200. Image tag from infra.engine_version. The packaged Tempo server is AGPL-3.0. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `tempo-chart`, which renders the community-maintained Tempo chart, where the server's configuration should follow the chart as it moves upstream rather than the hand-written file this realization imports, or where spans arrive in the Jaeger or Zipkin wire formats this one does not receive.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/tempo:1.0.0 - v1.0.0
Recipe
temporal
Temporal via the project's own Helm chart, served from the temporalio HTTPS chart repository. The chart deploys the engine's four services — the frontend that clients and workers dial, the history service that owns the event histories, the matching service that hands work to workers, and the engine's own internal worker — together with their services, the schema-setup jobs that create the tables before the services start, the browser console and the administrative command-line pod. The relational datastore is not deployed here: it arrives as a topology edge, and this Recipe passes that edge's resolved host, port and user into the engine's datastore configuration. When that connection carries a credential, the Recipe points the engine at the produced Secret by name rather than rendering the value. This Recipe pins the engine, console and administrative container images to explicit release tags for reproducible renders, and pins the chart's resource names to the release name the helm provider gives the chart, so the frontend Service is named predictably, in the alphabet a Kubernetes name allows. Default chart version in source.version; override via infra.source_version. The chart writes no namespace; the helm provider places its resources in the one the Release declares, and with none declared they land where the applying tool puts them. provides also publishes the pod labels, port name and path a scraper discovers every engine service's metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/temporal:1.0.0 - v1.0.0
Recipe
tempo-chart
Grafana Tempo distributed-tracing server on Kubernetes via the community-maintained Tempo Helm chart, served from that project's HTTPS chart repository, in the single-binary shape this Component describes. It renders a StatefulSet with a per-replica trace volume, a ClusterIP Service carrying the OpenTelemetry receivers and the HTTP query API, the Headless Service the StatefulSet needs, the server configuration and a ServiceAccount; the chart's ingress, route, network-policy, monitoring and legacy trace-browsing UI parts are left switched off. The server image and the chart version are pinned. The server's own metrics, in the Prometheus format at /metrics on the HTTP port 3200, which that ClusterIP Service carries, are published as provides.metrics_endpoint. What it carries that a hand-assembled workload does not is the server's configuration assembled by the chart from named settings rather than a hand-written file, so a key moving in a new Tempo release arrives with the chart, and receivers for the Jaeger and Zipkin wire formats beside OpenTelemetry, which the workload-synthesis realization does not carry. Reach for this realization where a chart repository is reachable at build time; reach for `tempo`, the workload-synthesis realization, where the build must complete with nothing fetched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/tempo-chart:1.0.0 - v1.0.0
Recipe
tei
Text Embeddings Inference as a self-hosted embedding server on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). Renders a StatefulSet: a weights-pull initContainer resolves the model pointer into a per-replica model-cache volume, and the serving container mounts the same volume and serves the embeddings API behind a ClusterIP Service. The model value is a deterministic pointer the initContainer resolves at pod start; the compiler never fetches weights. When the optional model_pull_token credential slot is attached the token reaches the initContainer as an env secretKeyRef, so it never appears in the rendered manifest; unattached, weights are pulled anonymously, byte-identically to a recipe with no slot. When the optional api_key slot is attached the key reaches the serving container as API_KEY the same way, the server requires it of every caller, and readiness opens the serving port rather than asking /health, which the key also guards at the default image tag; unattached, the server requires none. provides.host is the bare, self-contained in-cluster service name; a consumer in another namespace qualifies it on its own side. provides.metrics_endpoint names the server's metrics on the same Service while it runs one replica; above one it is empty and metrics_service names that Service for a scraper to discover every replica behind it. While api_key is attached a scrape presents that key.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/tei:1.0.0 - v1.0.0
Recipe
telnyx
The Telnyx messaging and voice carrier as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the carrier's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the Release calls a service its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to the component's slot is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the carrier. The credential slot is required here: a build leaving it empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/telnyx:1.0.0 - v1.0.0
Recipe
supertokens
An authentication core on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port an application's backend calls. The core keeps no durable state of its own — the relational backend is the system of record — so it renders stateless, with no persistent volume and no headless Service. The backend identity resolved from the `backend` uses-edge (host, port, user, database) arrives on .context.inject.env and composes the core's database connection; the core creates and migrates its own tables on startup, so no separate migration step is needed. The connection is rendered password-free where possible: with no credential on the backend connection the Recipe uses an inline development default, and with one attached the password is supplied via the credential subsystem. An API key is emitted only when one is attached, matching the core's own behaviour of issuing none by default. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/supertokens:1.0.0 - v1.0.0
Recipe
strimzi-operator
Installs the Strimzi Cluster Operator as an explicit component: a ServiceAccount, the cluster RBAC the operator needs to reconcile Kafka custom resources, and the operator Deployment on the upstream image. Emitting the control plane as a first-class component lets an operator-driven Kafka ship its runtime prerequisite alongside the Kafka custom resource it emits (the bundled-deploy prerequisite mode), so the custom resource and the operator that reconciles it are deployed together. This is a minimal control-plane bootstrap installing no custom-resource definitions, and at the image the Component pins its role falls short of what the operator needs to run a Kafka cluster: it grants nothing on pods, Services, Secrets, volume claims, roles or the operator's own pod sets, and renders none of the roles the operator binds brokers and the entity operator to. Reach for `strimzi-operator-chart`, which renders the project's own chart. The operator serves its own metrics at /metrics on port 8080, beside its liveness and readiness paths, and checks no caller. With `values.metrics_enabled` true a ClusterIP Service, `<instance>-metrics` in infra.namespace, carries that port and provides.metrics_endpoint publishes its address with the namespace; false, that key is empty and no Service is rendered.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/strimzi-operator:1.0.0 - v1.0.0
Recipe
strimzi-operator-chart
Installs the Strimzi Cluster Operator through the project's own Helm chart, at the chart release whose application version is the operator image's. The chart is official, published by the Strimzi maintainers; complete, carrying the operator, its definitions and every role it binds; and maintained, released with every operator release. Beside what the minimal bootstrap installs — the operator Deployment, its service account and a narrow cluster role — the chart renders the ten Strimzi custom-resource definitions, all served at the `v1` API, the operator's full namespaced, cluster-wide, watched-namespace and leader-election roles, the roles it binds the Kafka brokers, the entity operator and clients to, and the operator's logging configuration. The operator image is infra.image, and every image the operator starts Kafka and its companions on is pinned to the one that release names; nothing is generated when the build runs. Reach for this realization; its chart repository must be reachable from the build. The alternative, `strimzi-operator`, installs no definitions and at the image the Component pins grants less than the operator needs to run a Kafka cluster. The chart places every namespaced object in the namespace the Release, or else the Aggregate, declares for the component, so the build halts unless that namespace is infra.namespace; the operator watches that namespace. With `values.metrics_enabled` true a ClusterIP Service, `<name>-metrics` in infra.namespace, carries the operator's metrics port and provides.metrics_endpoint publishes its address; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/strimzi-operator-chart:1.0.0 - v1.0.0
Recipe
smolagents
A smolagents agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as SMOLAGENTS_AGENT_TYPE and SMOLAGENTS_MAX_STEPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/smolagents:1.0.0 - v1.0.0
Recipe
service
A generic HTTP service on Kubernetes via the helm provider's Workload synthesis: a Deployment plus a ClusterIP Service exposing an `http` port and a named `metrics` port — the port core/prometheus-exporter's ServiceMonitor scrapes by default.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/service:1.0.0 - v1.0.0
Recipe
semantic-kernel
A Semantic Kernel agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; the model name is delivered as OPENAI_CHAT_MODEL_ID, the variable Semantic Kernel's OpenAI chat service reads; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names, beside these. The agent's own values reach the container as SEMANTIC_KERNEL_PLANNER_TYPE and SEMANTIC_KERNEL_EXECUTION_SETTINGS (as JSON text, while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/semantic-kernel:1.0.0 - v1.0.0
Recipe
sagemaker
Amazon SageMaker's managed model serving endpoint as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied regional endpoint. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/sagemaker:1.0.0 - v1.0.0
Recipe
sealed-secrets
sealed-secrets controller via the official sealed-secrets Helm chart, served from an HTTPS chart repository. The chart deploys the controller that decrypts SealedSecret custom resources into native Kubernetes Secrets, together with its RBAC, its Services, and a ServiceAccount. The controller's own metrics, in the Prometheus format at /metrics on port 8081 of the chart's metrics Service, are published as provides.metrics_endpoint, the one address this component publishes. The SealedSecret custom resource definition ships in the chart's Helm-native CRD directory, so it installs with the release unconditionally. The controller generates its sealing keypair in the cluster at runtime, so no key material is produced at render time. This Recipe pins the controller container image to an explicit release tag for reproducible renders. Default chart version in source.version; override via infra.source_version.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/sealed-secrets:1.0.0 - v1.0.0
Recipe
search
The managed web-search API as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the search API's base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a service its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here: a consumer that needs the search API's key declares the `credential` slot on the socket that reaches the search API and delivers it from its own container, because the environment-variable name a credential arrives under belongs to the consumer rather than to the endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/search:1.0.0 - v1.0.0
Recipe
s3-object-store
An S3-compatible object store on Kubernetes: SeaweedFS 4.48, Apache-2.0, from its project's published chrislusf/seaweedfs image, run by the helm provider's Workload synthesis as one weed server process — master, volume server, filer and S3 gateway — in a StatefulSet with a persistent volume and a ClusterIP Service. An init container creates the bucket values.bucket names before the server serves, and leaves an existing one as it is. Reach for it where the store runs in the cluster with nothing else in place; for compose/s3-object-store on a Compose host; for terraform/s3-object-store-amazon-s3 or cloudformation/s3-object-store-amazon-s3 in an AWS account; and for terraform/s3-object-store-gcs in a Google Cloud project. The access_key_id and secret_access_key slots are the server's S3 account, and both are required here: each is read by reference as an env secretKeyRef, so no key reaches the rendered manifest, and an empty slot halts naming it. The server serves plain HTTP inside the cluster; only the S3 gateway listens beyond the pod, on its S3 port and the gRPC port beside it, and each refuses a caller without the key. It publishes the bucket, the endpoint http://<service>:8333, its host, port 8333, protocol s3 and region us-east-1, the one S3 clients read from a server reporting none. infra.region, any infra.overrides key and a bucket name SeaweedFS does not take halt naming the key.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/s3-object-store:1.0.0 - v1.0.0
Recipe
rivet
A Rivet agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied image, which carries both the Rivet runtime and the project file it executes. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as RIVET_PROJECT_FILE, RIVET_GRAPH_NAME and RIVET_INPUTS (as JSON text, while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/rivet:1.0.0 - v1.0.0
Recipe
retell
The Retell AI voice-agent platform as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the platform's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to the component's slot is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the platform. The credential slot is required here: a build leaving it empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/retell:1.0.0 - v1.0.0
Recipe
replicate
Replicate's managed prediction API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/replicate:1.0.0 - v1.0.0
Recipe
redis
Redis on Kubernetes via the helm provider's Workload synthesis, running the official upstream redis image (no external Helm chart). Used as an in-memory cache with no persistence, so it renders a stateless Deployment plus a ClusterIP Service and declares no storage. Authentication is off by default; when the optional password credential slot is attached the value is delivered as an env secretKeyRef and Redis is started with --requirepass to enforce it, so the plaintext password never appears in the rendered manifest. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The packaged Redis server is AGPL-3.0. With infra.client_tls true the cache serves TLS alone on its port from the Secret attached to server_certificate, mounted as files, the probes ping over TLS, and protocol is published as redis-tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/redis:1.0.0 - v1.0.0
Recipe
rabbitmq
RabbitMQ on Kubernetes via the helm provider's Workload synthesis, running the official rabbitmq image directly with no external chart (StatefulSet + ClusterIP Service + Headless Service). Nothing has to be installed in the cluster first and no chart repository has to be reachable, so this is the realization to reach for where no RabbitMQ control plane runs and where the build has to complete without network access; reach for the controller-driven recipe where the control plane is already in place. RabbitMQ persists queues and messages durably, so the broker renders as a stateful workload with a per-replica persistent volume mounted at the RabbitMQ data directory. RabbitMQ always authenticates: the application user comes from the developer value and its password from the optional `password` credential slot when attached (an env secretKeyRef), falling back to a deterministic inline development default otherwise so the render stays reproducible. The image is management-enabled, so the broker exposes both the AMQP port and the management dashboard. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The broker's metrics are on in the image, on port 15692. With `values.metrics_enabled` true that port, which checks no caller, joins the ClusterIP Service and its address is published as provides.metrics_endpoint; false, that key is empty and the port stays off the Service.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/rabbitmq:1.0.0 - v1.0.0
Recipe
rabbitmq-cluster
RabbitMQ realized as a RabbitmqCluster custom resource the RabbitMQ Cluster Operator reconciles, rather than a directly synthesized StatefulSet — the controller-driven realization of the same RabbitMQ component, selectable in the Manifest alongside the synthesis recipe. The recipe emits one RabbitmqCluster conformant to the rabbitmq.com/v1beta1 API; the controller turns it into a running broker, its client Service and its credential Secret. The RabbitmqCluster is written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation, where one is set. provides.host is the client Service the controller creates for the cluster, so a consumer wires to it exactly as it would to the synthesized realization. Credential custody is the operator's choice: infra.credentials generated leaves the password with the controller, and supplied points it at an existing Secret named in infra.credentials_secret. The RabbitMQ Cluster Operator is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. Reach for this realization where the control plane is already in place; reach for the synthesis recipe where it is not. The controller switches the broker's metrics on and puts their port, 15692, which checks no caller, on the client Service. With `values.metrics_enabled` true their address is published as provides.metrics_endpoint while the cluster runs one node; with more, provides.metrics_service names that client Service for a scraper to discover every node behind it. With the value false, those keys are empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/rabbitmq-cluster:1.0.0 - v1.0.0
Recipe
rabbitmq-cluster-operator
Installs the RabbitMQ Cluster Operator as an explicit component: a ServiceAccount, the cluster RBAC the controller needs to reconcile RabbitmqCluster custom resources, and the controller Deployment on the upstream image. Emitting the control plane as a first-class component lets a controller-driven RabbitMQ ship its runtime prerequisite alongside the RabbitmqCluster custom resource it emits (the bundled-deploy prerequisite mode), so the custom resource and the controller that reconciles it are deployed together. This is a minimal control-plane bootstrap; a production install typically sources the project's upstream release manifest, which also carries the custom-resource definition, the leader-election role and the admission webhooks with their serving certificates. The controller serves its own metrics at /metrics on port 8080, which checks no caller. With `values.metrics_enabled` true a ClusterIP Service, `<instance>-metrics` in infra.namespace, carries that port and provides.metrics_endpoint publishes its address with the namespace; false, that key is empty and no Service is rendered.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/rabbitmq-cluster-operator:1.0.0 - v1.0.0
Recipe
qdrant
Qdrant on Kubernetes via the helm provider's Workload synthesis, running the official upstream qdrant/qdrant image (no external Helm chart). Vector data is the point of this store rather than a cache of something else, so the workload renders as a StatefulSet with a persistent volume mounted at the server's storage path, alongside the governing headless Service the stateful form requires and a ClusterIP Service publishing both the HTTP and the gRPC port. Authentication is off by default; when the optional api_key credential slot is attached the value is delivered as an env secretKeyRef and the server reads it as its service API key, so the plaintext key never appears in the rendered manifest, and the optional read_only_api_key slot reaches the server the same way as its read-only key. provides.host is the bare in-cluster Service name; a consumer in another namespace qualifies it on its own side. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `qdrant-chart`, which renders the product's own Helm chart, where the server should carry the maintainers' own start-up script, volume ownership step and pre-stop pause. That realization additionally renders the chart's interaction check, a Pod annotated as a test hook whose script deletes and recreates a collection on the running server; this realization renders no such Pod. provides.metrics_endpoint names the server's metrics on the REST port; while api_key is attached a scrape presents the read-only key where read_only_api_key is attached, and the api_key otherwise.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/qdrant:1.0.0 - v1.0.0
Recipe
pydantic-ai
A Pydantic AI agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as PYDANTIC_AI_RESULT_TYPE, PYDANTIC_AI_RETRIES and PYDANTIC_AI_SYSTEM_PROMPT (while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/pydantic-ai:1.0.0 - v1.0.0
Recipe
qdrant-chart
Qdrant vector database on Kubernetes via the product's own Helm chart, served from its project's HTTPS chart repository. The chart is configured for the shape this Component describes: one server with a persistent volume for its vector data, publishing the REST and gRPC listeners. It also renders the chart's own interaction check, a Pod carrying the `helm.sh/hook: test` annotation, whose script deletes the collection named `test_collection` on the running server, then recreates it, inserts points into it and queries them back: anything held under that collection name is destroyed. Deploying the Release does not run it, but an output applied wholesale, with the hook annotation disregarded, does, and the chart offers no setting to decline it. Where a Release must not carry it, apply the rendered files selectively, or reach for `qdrant`, which renders no such Pod. What it carries that a hand-assembled workload does not is the start-up machinery the product's own maintainers keep in step with it: joining the peer group, taking ownership of the data volume, a readiness check and a pre-stop pause. Reach for this realization where a chart repository is reachable at build time; reach for `qdrant`, the workload-synthesis realization, where the build must complete with nothing fetched. provides.host is the bare in-cluster Service name; a consumer in another namespace qualifies it on its own side. provides.metrics_endpoint names the server's metrics on the REST port; while api_key is attached a scrape presents the read-only key where read_only_api_key is attached, and the api_key otherwise.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/qdrant-chart:1.0.0 - v1.0.0
Recipe
promptlayer
The hosted prompt registry as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the registry's base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a service its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here: a consumer that needs the registry's API key declares the `credential` slot on the socket that reaches the registry and delivers it from its own container, because the environment-variable name a credential arrives under belongs to the consumer rather than to the endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/promptlayer:1.0.0 - v1.0.0
Recipe
promptflow
A Prompt flow application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the flow talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. Its own value flow_entrypoint reaches the container as PROMPTFLOW_FLOW_ENTRYPOINT by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/promptflow:1.0.0 - v1.0.0
Recipe
prometheus
Prometheus metrics server on Kubernetes via the helm provider's Workload synthesis, running the official prom/prometheus image with no external chart. It renders a StatefulSet with a data volume for the time-series database, a Headless Service, and a ClusterIP Service for queries. Each connection on the `scrape_target` uses-socket becomes one job in a minimal prometheus.yml the container writes at start-up, its label and cadence set by the connection's own `values`. A target publishing `metrics_endpoint` is scraped at that address, its host qualified with the target's namespace where that differs from Prometheus's. A target publishing `metrics_service` or `metrics_pod_selector` is discovered from the cluster instead, in the namespace its connection's `aggrega.ai/namespace` annotation names, or in every namespace for KEDA and Knative Serving; while one is, the server runs under a ServiceAccount of its own, granted read access where it discovers, and the build halts when Prometheus declares no namespace. A target publishing none halts the build. A connection's `credential` is mounted as a file, by reference, and its job presents it as a bearer token. With `infra.external_base_url` set, the server takes it as its external URL and is kept serving at its root. With `values.otlp_receiver` true, it also runs the OTLP receiver and publishes its address as provides.otlp_http_endpoint. Reach for this realization where the build must complete with nothing fetched; reach for `prometheus-chart` where the server is also expected to scrape whatever the cluster advertises through scrape annotations.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/prometheus:1.0.0 - v1.0.0
Recipe
prometheus-chart
Prometheus metrics server on Kubernetes via the prometheus-community Helm chart, configured as one server: the chart's alerting, exporter and push-gateway subcharts are switched off. It renders a Deployment, a ConfigMap carrying the scrape configuration, a PersistentVolumeClaim for the time-series database, a ClusterIP Service, and the ServiceAccount, ClusterRole and ClusterRoleBinding the server reads the cluster's own endpoints under. The server image and the chart version are pinned. That cluster-scoped read grant also serves the chart's own jobs, which scrape whatever the cluster advertises through scrape annotations. Each connection on the `scrape_target` uses-socket becomes one scrape job, as on the workload-synthesis realization: a target publishing `metrics_endpoint` is scraped at that address; one publishing `metrics_service` or `metrics_pod_selector` is discovered in the namespace its connection's `aggrega.ai/namespace` annotation names, in every namespace for KEDA and Knative Serving, or in the server's own where neither declares one; a target publishing none halts the build. A connection's `credential` is mounted as a file through `server.extraSecretMounts`, and its job presents it as a bearer token. Reach for it where a chart repository is reachable and the server should discover targets beyond those wired, and for `prometheus` where the build must fetch nothing. With `infra.external_base_url` set, the server takes it as its external URL through the chart's `server.baseURL` and is kept serving at its root. With `values.otlp_receiver` true, `server.extraArgs` also turns on the OTLP receiver, whose address it publishes as provides.otlp_http_endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/prometheus-chart:1.0.0 - v1.0.0
Recipe
presidio-anonymizer
The Presidio anonymizer on Kubernetes via the helm provider's Workload synthesis, running the vendor's published image directly: a stateless Deployment plus a ClusterIP Service on the port the service answers on. Synthesized rather than wrapping a chart because the project publishes no maintained one — the chart in its repository is a sample that floats its image tag and declares no health checks. The service keeps no state, so it renders no volume. The listen port is handed to the image through the PORT variable it reads, from the same infra.port the Service and the published address read. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The image is published under the MIT licence.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/presidio-anonymizer:1.0.0 - v1.0.0
Recipe
postgres
PostgreSQL on Kubernetes via the helm provider's Workload synthesis (Deployment + ClusterIP Service). provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Reach for this realization where no PostgreSQL control plane runs in the target cluster: it needs no controller installed and no chart repository, so it builds with nothing fetched. Reach for the controller-driven recipe where that control plane is already in place, and for the reference recipe where the server already exists outside the deployment.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/postgres:1.0.0 - v1.0.0
Recipe
presidio-analyzer
The Presidio analyzer on Kubernetes via the helm provider's Workload synthesis, running the vendor's published image directly: a stateless Deployment plus a ClusterIP Service on the port the service answers on. Synthesized rather than wrapping a chart because the project publishes no maintained one — the chart in its repository is a sample that floats its image tag and declares no health checks. The service keeps no state, so it renders no volume. The listen port is handed to the image through the PORT variable it reads, from the same infra.port the Service and the published address read. The image loads its English language model before it answers anything, which is why readiness is first asked only after the grace period the image itself declares. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The image is published under the MIT licence.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/presidio-analyzer:1.0.0 - v1.0.0
Recipe
postgres-external
A PostgreSQL server that already exists outside this deployment, reached as a reference realization: nothing is provisioned and no compute runs. helm emits an ExternalName Service that maps a stable in-cluster name onto the operator-supplied endpoint. The same provides contract the provisioning Recipes publish is published here, so a consumer's wiring is unchanged — only the pointer string behind it differs. No credential is delivered by this recipe: the credential for the referenced server is a referenceable secret attached to the Component's optional password slot, because the environment-variable name belongs to the consumer rather than to the endpoint. Reach for this realization where the server already exists outside the deployment and nothing is to be provisioned. Reach for the synthesis recipe to stand a server up on a cluster that runs no PostgreSQL control plane, and for the controller-driven recipe where that control plane is already in place.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/postgres-external:1.0.0 - v1.0.0
Recipe
pgvector
pgvector on Kubernetes via the helm provider's Workload synthesis: a StatefulSet running the pgvector project's own image with a persistent volume, a ClusterIP Service, and the headless Service the stateful form requires. A ConfigMap carries the script that enables the extension in the database the server creates on its first start. The project publishes no Helm chart, so the workload is synthesized rather than sourced. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Reach for this realization on Kubernetes: it needs no control plane and no chart repository, and it builds with nothing fetched. The Compose Recipe of the same name is the local-development realization.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/pgvector:1.0.0 - v1.0.0
Recipe
postgres-cloudnative-pg
PostgreSQL realized as a Cluster custom resource the CloudNativePG control plane reconciles, rather than a directly synthesized StatefulSet — the controller-driven realization of the same PostgreSQL component, selectable in the Manifest alongside the synthesis recipe and the reference recipe. The recipe emits one Cluster conformant to the postgresql.cnpg.io/v1 API; the controller turns it into a running server, the Services that reach it and the Secret holding the database owner's password, and creates the initial database and its owning role from the developer's own values. provides.host is the read-write Service, so a consumer wires to it exactly as it would to the synthesized realization. The Cluster, and the ConfigMap carrying any bootstrap scripts, are written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation, where one is set. Credential custody is the operator's choice: infra.credentials generated leaves the owner's password with the controller, and supplied points it at an existing Secret named in infra.credentials_secret. The CloudNativePG control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. Reach for this realization where the control plane is already in place; reach for the synthesis recipe where it is not, and for the reference recipe where the server already exists outside the deployment. A core/backup attachment bound to helm/backup-cloudnative-pg is realized through the Barman Cloud plugin.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/postgres-cloudnative-pg:1.0.0 - v1.0.0
Recipe
phidata
A Phidata agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL; the datastore wired into `storage` contributes the coordinates for session history and the store wired into `knowledge_base` those for retrieved reference material; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. A store served over TLS alone or over HTTPS halts the build on a store connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as PHIDATA_ADD_HISTORY_TO_MESSAGES and PHIDATA_MARKDOWN by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/phidata:1.0.0 - v1.0.0
Recipe
otel-collector-chart
The OpenTelemetry Collector, contrib distribution, on Kubernetes via the project's own opentelemetry-collector Helm chart, served from its HTTPS chart repository, as one Deployment. It renders the Deployment, a ClusterIP Service publishing the OTLP gRPC and HTTP ports and no other, the ServiceAccount the pod runs under, and the collector's configuration as a ConfigMap emitted beside the chart. The configuration is the OtelCollectorConfig bundle, which the compose realization imports too, with one exporter for each backend wired to the traces, metrics and logs sockets. The chart is told to create no configuration of its own and to mount that ConfigMap instead, so the receivers and ports the chart configures by default are absent from the render. The pod carries a digest of the infra and the wired backends, so a changed address rolls it. The image tag follows infra.engine_version and the chart version is pinned; the chart repository must be reachable at build time. With `values.metrics_enabled` true the collector serves its own telemetry metrics at port 8888 on every interface, which checks no caller, the chart puts that port on the Service, and provides.metrics_endpoint publishes the address; false, that key is empty. The build then halts on a Service type other than ClusterIP.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/otel-collector-chart:1.0.0 - v1.0.0
Recipe
opensearch-chart
OpenSearch on Kubernetes via the product's own Helm chart, published by OpenSearch's maintainers in their HTTPS chart repository. The chart is configured for one node with its indexes on its own volume. It renders a StatefulSet, a ClusterIP Service publishing the REST API, the transport and the performance analyzer's ports, a headless Service, a ConfigMap, a Secret holding the security configuration and a PodDisruptionBudget. The security plugin stays on. The image's demonstration setup is replaced by one account, admin, whose password is read from the attached credential's Secret by name, so it never appears in the rendered files; with none attached it is a development default. A credential that names no Secret halts the build. An init container generates the node's transport key each time the pod starts. The REST API is served over plain HTTP, not encrypted, and is reachable only inside the Release. Reach for this realization on Kubernetes where a chart repository is reachable at build time; for local development reach for the Compose Recipe opensearch, which runs the same image with the same security setup.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/opensearch-chart:1.0.0 - v1.0.0
Recipe
openhands
An OpenHands agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as OPENHANDS_TASK (while set) and OPENHANDS_MAX_ITERATIONS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openhands:1.0.0 - v1.0.0
Recipe
openbao
OpenBao secrets-management server on Kubernetes via the official OpenBao Helm chart (openbao/openbao). The chart is configured for a lean, single-workload deployment: the agent injector is disabled, the server runs in standalone mode with a single replica, and its state is kept in the built-in file storage backend on a persistent volume. Under this configuration the chart renders a StatefulSet for the server, a ConfigMap carrying the server configuration, the main and internal (headless) Services, a ServiceAccount, a ClusterRoleBinding, and the chart's own test Pod, marked as a helm test hook. OpenBao is initialized and unsealed at runtime inside the cluster, so no credential is wired at render time and no secret is materialized. With `infra.external_base_url` set, the server advertises it as its API address through the chart's `server.ha.apiAddr`, which the chart reads in every mode; unset, the chart advertises the pod's own address. The server image is pinned to a resolvable release tag for reproducible renders. Default chart version in source.version; override via infra.source_version. provides.host is the bare, self-contained in-cluster Service name — it carries no namespace, so the recipe is independent of where it is deployed; a consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). OpenBao is distributed under the MPL-2.0 license. With `values.metrics_enabled` true the server configuration gains a telemetry stanza turning on its Prometheus metrics, served at /v1/sys/metrics on the API port to a token, and provides.metrics_endpoint publishes the address; false writes nothing and leaves that key empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openbao:1.0.0 - v1.0.0
Recipe
openclaw
An OpenClaw agent on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied image, which carries both the agent and the JSON configuration it reads. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef, which the agent reads directly. The model endpoint published by the component wired into the `model` socket is delivered under a name this recipe defines, because the agent resolves its endpoint from its configuration file rather than from the environment. The endpoint and the model name are delivered as OPENCLAW_MODEL_BASE_URL and OPENCLAW_MODEL by default; a connection writing an `inject:` block receives the names it dictates in their place, as variables or as files mounted at the paths it names, while the API key stays. The agent's own values reach the container as OPENCLAW_EXECUTION_MODE, OPENCLAW_MAX_EXECUTION_TIME and OPENCLAW_MAX_STEPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openclaw:1.0.0 - v1.0.0
Recipe
openai
OpenAI's managed chat-completions API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openai:1.0.0 - v1.0.0
Recipe
openapi-knative
The OpenAPI-driven tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The location of the OpenAPI document is delivered into the container, so the operations the tool offers are read from the specification the developer named rather than fixed by this recipe. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, so a consumer wires to the tool's declared contract rather than to anything Knative-specific. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openapi-knative:1.0.0 - v1.0.0
Recipe
openai-agents-sdk
An application built on the OpenAI Agents SDK, on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the image the Release supplies. The model wired into the `model` socket reaches the application as the variables the SDK reads: its endpoint as OPENAI_BASE_URL, its model name as OPENAI_DEFAULT_MODEL, and the connection's credential as OPENAI_API_KEY through a secretKeyRef, so the key never renders into the manifest. While the component's upload_traces_to_openai value is false, its default, the application also receives OPENAI_AGENTS_DISABLE_TRACING set to 1, which turns off the SDK's trace upload to OpenAI; set true, the variable is absent. Each tool wired into the `tools` socket reaches the application under variables named for the connection's alias, upper-cased with hyphens as underscores: its address as AGGREGA_TOOL_<ALIAS>_URL, and, where the connection carries a credential, that credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL through a secretKeyRef, with the header the tool publishes for it as AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection holding the tool's whole address as a secret, for a server that reads its key from its address, delivers that secret as AGGREGA_TOOL_<ALIAS>_URL through a secretKeyRef instead. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names: a tools connection in place of its address and header variables, while what its attachments deliver stays, and the model connection beside the variables the SDK reads. The component's own `inject:` block delivers its entries as a connection's are.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openai-agents-sdk:1.0.0 - v1.0.0
Recipe
openai-embedding
OpenAI's managed embeddings API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openai-embedding:1.0.0 - v1.0.0
Recipe
openai-compatible
An OpenAI-compatible broker gateway as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied gateway endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/openai-compatible:1.0.0 - v1.0.0
Recipe
open-webui
A self-hosted Open WebUI installation on Kubernetes via the helm provider's Workload synthesis: a single-replica Deployment, a ClusterIP Service on the listen port, and one persistent volume holding the accounts, the conversations, the prompt library and the document index. The model edge's endpoint becomes OPENAI_API_BASE_URL and its model name the model a new conversation opens on; the platform's Ollama client is switched off. The listen port is emitted as PORT as well as published on the container port and the Service, because the platform reads its bind port from the environment. infra.external_base_url, when set, reaches the platform as WEBUI_URL. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY, and with none attached no key is emitted. The session secret is delivered as WEBUI_SECRET_KEY; it has no default, and the Component requires its slot. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `open-webui-chart`, which renders the platform's own Helm chart, where single sign-on or per-component log levels are wanted — that realization carries the integration surface this one has nowhere to declare.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/open-webui:1.0.0 - v1.0.0
Recipe
ollama
Ollama as a self-hosted inference server on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). Renders a StatefulSet: a weights-pull initContainer resolves the model pointer into a per-replica model-cache volume, and the serving container mounts the same volume and serves the OpenAI-compatible API behind a ClusterIP Service. The model value is a deterministic pointer the initContainer resolves at pod start; the compiler never fetches weights. When the optional model_pull_token credential slot is attached the token reaches the initContainer as an env secretKeyRef, so it never appears in the rendered manifest; unattached, weights are pulled anonymously, byte-identically to a recipe with no slot. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Readiness and liveness probes ask the server's root path, which answers once the server listens. Reach for this realization where the build must complete with nothing fetched, or where more than one serving replica is wanted — each replica here gets its own model cache; reach for `ollama-chart`, which renders the community-maintained Ollama chart, where the replica bounds should render a horizontal autoscaler, which this realization does not emit.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/ollama:1.0.0 - v1.0.0
Recipe
open-webui-chart
A self-hosted Open WebUI installation on Kubernetes via the platform's own Helm chart, served from its project's HTTPS chart repository: one platform process against one durable volume, reaching its model over the `model` uses-edge. It renders a StatefulSet whose init container seeds the data directory from the image, a ClusterIP Service on the listen port, a PersistentVolumeClaim for the platform's state and a ServiceAccount; the chart's ingress, route and certificate resources are left switched off. The platform image and the chart version are pinned. What it carries that a hand-assembled workload does not is the platform's own integration surface, of which this Recipe passes single sign-on against several identity providers and per-component log levels through infra.overrides. Reach for this realization where a chart repository is reachable at build time; reach for `open-webui`, the workload-synthesis realization, where the build must complete with nothing fetched. It declines the chart's bundled model runner, extension server and session cache, and empties the chart's shipped literal API key. infra.external_base_url, when set, reaches the platform as WEBUI_URL. The session secret has no default: the Component requires its slot, and the chart reads it from the Secret that slot's credential names.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/open-webui-chart:1.0.0 - v1.0.0
Recipe
ollama-chart
Ollama as a self-hosted inference server on Kubernetes via the community-maintained Ollama Helm chart, served from that project's HTTPS chart repository. It renders a Deployment, a ClusterIP Service on the model port, a PersistentVolumeClaim for the model cache and the pod's ServiceAccount, plus a horizontal autoscaler when the Release declares replica bounds; the chart's Ingress, gateway and route surfaces, its Knative option and its connection-test Pod are left switched off. Over a hand-assembled workload it adds that autoscaler, and it resolves the model pointer through the runtime's own pull step after the server is up. Reach for this realization where a chart repository is reachable at build time; reach for `ollama`, the workload-synthesis realization, where the build must complete with nothing fetched, or where more than one serving replica is wanted — that realization gives each replica its own model cache, which this one cannot. The runtime image and the chart version (source.version, overridden via infra.source_version) are pinned. provides.host is the bare in-cluster Service name; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/ollama-chart:1.0.0 - v1.0.0
Recipe
nvidia-nim
NVIDIA Inference Microservice as a self-hosted, GPU-native inference server on Kubernetes via the helm provider's Workload synthesis, running the NIM container directly with no chart. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: the serving image comes from infra.image, a NIM image the reader must be entitled to pull, and the catalog names none. Renders a StatefulSet whose container resolves the model profile into a per-replica model-cache volume at start and serves the OpenAI-compatible API behind a ClusterIP Service. It always requests a GPU (nvidia.com/gpu), and schedules the pod under infra.gpu.runtime_class when one is named. An attached ngc_api_key reaches the container as an env secretKeyRef, so the key never appears in the rendered manifest. provides.host is the bare in-cluster Service name. Reach for this realization where no NIM Operator runs and the build must fetch nothing. Reach for `nvidia-nim-chart` where NVIDIA's nim-llm chart is wanted and its repository is reachable from the build, and for `nvidia-nim-service` where the NIM Operator runs or is declared beside the model.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nvidia-nim:1.0.0 - v1.0.0
Recipe
nvidia-nim-chart
NVIDIA Inference Microservice as a self-hosted, GPU-native inference server on Kubernetes through NVIDIA's own `nim-llm` Helm chart, pulled from NVIDIA's chart repository at helm.ngc.nvidia.com. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: the chart's licence file names LicenseRef-NvidiaProprietary, and it runs the NIM image infra.image names, which the reader must be entitled to pull. The catalog names no image, and the build halts until infra.image is set. The chart renders a StatefulSet giving each replica its own model-cache claim, its Service on infra.port and the headless Service beside it. An attached ngc_api_key reaches the chart as the name and data key of its Secret, so no key is rendered. The chart's connection-test Pods are left out, and its autoscaler, ingress and monitor stay off. Reach for this realization where NVIDIA's chart is wanted and its repository is reachable from the build. Reach for `nvidia-nim`, the workload synthesis, where the build must fetch nothing or a GPU runtime class is named, which this chart cannot set; reach for `nvidia-nim-service` where the NIM Operator runs.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nvidia-nim-chart:1.0.0 - v1.0.0
Recipe
nvidia-api
The managed NVIDIA API as a reference realization: no compute runs in- cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nvidia-api:1.0.0 - v1.0.0
Recipe
nvidia-nim-operator
Installs NVIDIA's NIM Operator through NVIDIA's own `k8s-nim-operator` Helm chart, pulled from NVIDIA's chart repository at helm.ngc.nvidia.com, at the chart release whose application version is the tag of the operator image the Component defaults to; the image is infra.image. The chart is official, published by NVIDIA; complete, carrying the operator, its custom-resource definitions and every role it binds; and maintained, each of its releases carrying the operator release of the same version. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: NVIDIA's documentation lists the chart and the operator image under its own licence agreement, and every NIMService the operator reconciles pulls a NIM image the reader must be entitled to pull. The cluster must already run NVIDIA's GPU Operator, which no build checks. The chart renders the operator Deployment, its account, its cluster and leader-election roles, its metrics Service and the nine definitions of its `apps.nvidia.com` API, NIMService among them. Its admission webhook, which needs cert-manager, stays off, as do its definitions-upgrade hook and its Dynamo components. Nothing is generated when the build runs. The chart places its namespaced objects in the namespace the Release, or else the Aggregate, declares for the component, so the build halts unless one is declared.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nvidia-nim-operator:1.0.0 - v1.0.0
Recipe
nvidia-nim-service
NVIDIA Inference Microservice realized as a `NIMService` custom resource that NVIDIA's NIM Operator reconciles, rather than a directly synthesized StatefulSet, selectable in the Manifest beside the synthesis and chart realizations. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: the operator runs the NIM image infra.image names, which the reader must be entitled to pull, and reads the key from the Secret the ngc_api_key slot names, so the build halts while that slot is unattached. The catalog names no image. The resource is authored against the operator's 3.1.2 release at the `apps.nvidia.com/v1alpha1` API, and written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation. The operator runs one serving replica on one model-cache claim it creates, requests the GPU on the container, and creates the Service provides names. The NIM Operator is a runtime prerequisite: the build halts unless the operator affirms it with infra.control_plane_present. Reach for this realization where the operator runs, or is declared beside the model as `core/ai-model/nvidia-nim-operator`; reach for `nvidia-nim`, the synthesis, where it is not, or where more than one replica is wanted.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nvidia-nim-service:1.0.0 - v1.0.0
Recipe
nginx
NGINX reverse-proxy gateway on Kubernetes via the helm provider's Workload synthesis, running the official upstream nginx image (no external Helm chart). NGINX holds no persistent state, so it renders a stateless Deployment plus a ClusterIP Service and declares no storage. The upstream host and port arrive on .context.inject.env under the two variables the locals name, which the `upstream` edge's inject block names, and are written into a minimal NGINX server configuration whose proxy_pass forwards matching requests to that upstream; the config is written at container start and nginx is then started in the foreground. With an edge wired on any socket but guard and either variable missing from the pipe, the build halts naming the edge and its socket; with neither on the pipe and no such edge, the location answers 503. A core/prompt-injection-shield attachment's realization (prompt-injection-shield-nginx) writes the two variables with the address of the guard wired to the guard socket. Each core/traffic-shadowing attachment adds a mirror to that location: an upstream group naming the Service the attachment's infra names, and an internal location sending a copy of every forwarded request there with its original URI, whose response NGINX discards. NGINX copies every request it forwards and cannot copy a share of them, so the FeatureRecipe realizing the attachment (traffic-shadowing-nginx) halts on a share below 100, and this Recipe halts on an attachment to a gateway that forwards nothing. Each core/concurrency-limit attachment adds a cap, one total across NGINX's worker processes, and a request past it is answered 503; concurrency-limit-nginx refuses scope model.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nginx:1.0.0 - v1.0.0
Recipe
nemoguard-jailbreak-detect
NVIDIA's NemoGuard JailbreakDetect NIM on Kubernetes via the helm provider's Workload synthesis, running the NIM container directly with no chart, since NVIDIA publishes none for it. Using it requires an NVIDIA AI Enterprise licence or NVIDIA Developer Program membership, and an NGC API key: the image comes from infra.image, a NIM image the reader must be entitled to pull, and the catalog names none. Renders a StatefulSet of one replica whose container fetches its model into a model-cache volume at start and serves the classify route behind a ClusterIP Service. It always requests a GPU (nvidia.com/gpu), schedules the pod under infra.gpu.runtime_class when one is named, and pulls with the Secret infra.image_pull_secret names. An attached ngc_api_key reaches the container as an env secretKeyRef, so the key never appears in the rendered manifest. provides.host is the bare in-cluster Service name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nemoguard-jailbreak-detect:1.0.0 - v1.0.0
Recipe
nemo-guardrails
The NeMo Guardrails server as a guard in a model's path, on Kubernetes via the helm provider's Workload synthesis, running the image in infra.image directly: a stateless Deployment, a ClusterIP Service on the port the guard answers on, and a ConfigMap carrying the guard's configuration as the product's own two files. Synthesized rather than wrapping a chart because the project publishes none. The ConfigMap's config.yaml names the Hugging Face classifier rail on the guard's input, pointed at the endpoint and model the detector wired on the detector socket publishes, with values.threshold and values.blocked_labels, and passes each admitted request through to the upstream unchanged; its rails.co carries the flow that keeps a request body from switching that rail off. Each file is mounted on its own at the path the image's start-up reads, so the guard starts in the product's single-configuration mode and refuses a request naming any other configuration. A NemoGuard model wired on content_safety or jailbreak_detection adds its own input rail. The upstream address is handed to the guard through the MAIN_MODEL_BASE_URL variable it reads, from infra.upstream_endpoint. When the optional upstream_key slot is attached, config.yaml names the variable the credential is delivered under as an env secretKeyRef, so its value never appears in the rendered manifest; unattached, config.yaml names no credential and the guard presents none. The guard answers its health route as soon as it serves, so readiness and liveness both ask it there. provides.host is the bare, self-contained in-cluster service name of the synthesized Service; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nemo-guardrails:1.0.0 - v1.0.0
Recipe
nats
NATS on Kubernetes via the helm provider's Workload synthesis, running the official upstream nats image (no external Helm chart). Used as an in-memory messaging server with JetStream file storage disabled, so it renders a stateless Deployment plus a ClusterIP Service and declares no storage. Authentication is off by default; when the optional password credential slot is attached the value is delivered as an env secretKeyRef and the server is started with user/password authentication flags referencing it, so the plaintext password never appears in the rendered manifest. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Its readiness and liveness probes open a connection to the client port. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `nats-chart`, which renders the product's own chart, where in-place configuration reload, probes against the monitoring endpoint, a graceful client drain on shutdown or the monitoring endpoint itself are wanted — this realization carries none of them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nats:1.0.0 - v1.0.0
Recipe
nats-chart
NATS messaging server on Kubernetes via the product's own Helm chart, served from its project's HTTPS chart repository. The chart is configured for the lean shape this Component describes: one server routing subjects in memory, with stream persistence, clustering, gateway, leaf-node, MQTT and websocket left off and the chart's administrative client Deployment declined. It renders a StatefulSet, a ClusterIP Service on the client port, a Headless Service, a ConfigMap holding the server configuration, and a disruption budget. What it carries that a hand-assembled workload does not is the server's own operational machinery: a sidecar that reloads the configuration in place, probes against the monitoring endpoint on the port the Component declares, and a graceful shutdown that drains connected clients. Reach for this realization where a chart repository is reachable at build time; reach for `nats`, the workload-synthesis realization, where the build must complete with nothing fetched. Every resource it renders lands in the namespace the Release declares for the component, or in `default` where none is declared. Both images it runs are pinned to resolvable release tags, and the chart version in source.version. provides.host is the bare in-cluster Service name; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/nats-chart:1.0.0 - v1.0.0
Recipe
mysql
MySQL on Kubernetes via the helm provider's Workload synthesis (StatefulSet + ClusterIP Service + Headless Service) on the official mysql image — no external chart. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mysql:1.0.0 - v1.0.0
Recipe
mysql-operator
Installs the MySQL Operator for Kubernetes control plane as an explicit component: a ServiceAccount, the cluster RBAC the controller needs to reconcile InnoDBCluster custom resources, and the controller Deployment on the pinned controller image. Emitting the control plane as a first-class component lets a controller-driven MySQL InnoDB Cluster ship its runtime prerequisite alongside the custom resource it emits, so the custom resource and the controller that reconciles it are deployed together. This is a minimal control-plane bootstrap installing no custom-resource definitions, and at the image the Component pins its role falls short of what the controller needs: it grants no pod status, disruption budgets, cron jobs, namespaces or definition reads, and renders none of the sidecar and switchover roles the controller binds a cluster's servers to. Reach for `mysql-operator-chart`, which renders the project's own chart.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mysql-operator:1.0.0 - v1.0.0
Recipe
mysql-operator-chart
Installs the MySQL Operator for Kubernetes control plane through the project's own Helm chart, at the chart release whose application version is the controller image's. The chart is official, published by the MySQL team from the operator's repository; complete, carrying the controller, its definitions and its roles; and maintained, released at each operator release. The project's chart repository serves its newest release alone, so a chart release it has replaced resolves only from a cache that already holds it. Beside what the minimal bootstrap installs — the controller, its service account and a cluster role — the chart renders the five custom-resource definitions the controller reads, the controller's full cluster role, the roles it binds the database servers' sidecars and switchover jobs to, the peering object its replicas coordinate through and the controller's Service. The controller image is pinned from infra.image and nothing is generated when the build runs. Reach for this realization; its chart repository must be reachable from the build. The alternative, `mysql-operator`, installs no definitions and at the image the Component pins grants less than the controller needs. The chart places every namespaced object in the namespace the Release, or else the Aggregate, declares for the component, so the build halts unless that namespace is infra.namespace.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mysql-operator-chart:1.0.0 - v1.0.0
Recipe
mysql-innodb-cluster
A MySQL InnoDB Cluster realized as an InnoDBCluster custom resource the MySQL Operator for Kubernetes reconciles. The recipe emits one InnoDBCluster conformant to the mysql.oracle.com/v2 API, requesting a count of MySQL server instances and a count of MySQL Router instances; the controller acts on that request inside the target cluster. provides.host is the Service the controller creates for the cluster, so a consumer wires to it by name. Credential custody is the operator's choice: infra.credentials supplied points the custom resource at an existing Secret named in infra.credentials_secret, and generated renders that Secret beside the custom resource from the credential attached to the root_password slot. The custom resource, and the Secret where one is rendered, are written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation, where one is set. The MySQL Operator control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. Reach for this realization where the control plane is already in place; reach for core/datastore/mysql where it is not, or where the application wants a logical database and a login role created for it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mysql-innodb-cluster:1.0.0 - v1.0.0
Recipe
model-invocation-knative
The model-invocation tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The model the tool calls arrives on the tool's model socket, and the recipe delivers that model's endpoint and model identifier — and its in-cluster host, port and protocol where the model publishes them — into the container alongside the developer's task type, so the running tool calls the model the topology resolved rather than one named in its own configuration. A connection writing an `inject:` block receives instead the names it dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/model-invocation-knative:1.0.0 - v1.0.0
Recipe
milvus-chart
Milvus vector database on Kubernetes via the product's own Helm chart, published by Milvus's maintainers in their HTTPS chart repository. The chart is configured for one standalone server that holds its metadata store and its data on its own volume, so it runs one image and deploys no etcd, object store or message-queue service. It renders a Deployment, a ClusterIP Service publishing the API and the metrics port, a PersistentVolumeClaim and a ConfigMap. Authentication is off unless a password is attached; attached, it is switched on for the API and for the management endpoints the metrics port serves, whose health check and metrics stay open, and the password is read from the attached credential's Secret by name, so it never appears in the rendered files. A credential that names no Secret halts the build. Reach for this realization on Kubernetes where a chart repository is reachable at build time; for local development reach for the Compose Recipe milvus, which runs the same image in the same single-server mode.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/milvus-chart:1.0.0 - v1.0.0
Recipe
mistral
Mistral AI's managed chat API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mistral:1.0.0 - v1.0.0
Recipe
metagpt
A MetaGPT agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as METAGPT_ROLE and METAGPT_OBJECTIVE by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/metagpt:1.0.0 - v1.0.0
Recipe
memcached
Memcached on Kubernetes via the helm provider's Workload synthesis, running the official upstream memcached image directly (no external Helm chart). Used as an in-memory cache with no persistence, so it renders a stateless Deployment plus a ClusterIP Service and declares no storage. The cache runs without authentication, and a password attached to the password slot halts the build naming the slot: the memcached image enforces no password, and its two authentication modes, SASL over the binary protocol and an ASCII authentication file, are not spoken by the catalog's Memcached clients, so the cache would accept every client whatever the attachment. Access to it is the network's to restrict. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). With infra.client_tls set to true the cache serves TLS alone on its port: the Secret attached to server_certificate is mounted as files, the server is started with -Z and presents its tls.crt and tls.key, requests no client certificate, and protocol is published as memcached-tls; the TCP probes are unchanged, since a connection opens before any handshake. client_tls true with nothing attached to server_certificate halts the build naming the slot, a certificate attached while client_tls is false halts it naming the setting, and an attachment publishing no Secret name halts it naming the slot.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/memcached:1.0.0 - v1.0.0
Recipe
mastra
A server built by Mastra, on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the image the Release supplies, and a ClusterIP Service fronting infra.port, which the server receives as PORT beside MASTRA_HOST set to 0.0.0.0. A readiness probe reads Mastra's /health route, so the Service sends no traffic until the server answers; provides is that Service's host, port, endpoint and protocol, and, where values.agent_id is set, the address of that agent's Agent2Agent card. The model wired into the `model` socket reaches the server as the variables Mastra reads for an openai/ model: its endpoint as OPENAI_BASE_URL, and the connection's credential as OPENAI_API_KEY through a secretKeyRef, so the key never renders into the manifest. Each tool wired into the `tools` socket reaches the server under variables named for the connection's alias, upper-cased with hyphens as underscores: its address as AGGREGA_TOOL_<ALIAS>_URL, and, where the connection carries a credential, that credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL through a secretKeyRef, with the header the tool publishes for it as AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names: a tools connection in place of its address and header variables, while what its attachments deliver stays, and the model connection beside the variables Mastra reads. So does the component's own block. A set values.agent_id reaches the server as MASTRA_AGENT_ID by default; an entry of the component's own block reading it alone replaces that name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mastra:1.0.0 - v1.0.0
Recipe
mariadb
MariaDB on Kubernetes via the helm provider's Workload synthesis, running the official mariadb image directly with no external chart (StatefulSet + ClusterIP Service + Headless Service). Nothing has to be installed in the cluster first and no chart repository has to be reachable, so this is the realization to reach for where no MariaDB control plane runs and where the build has to complete without network access. It honors the developer's first-initialization scripts, which the controller-driven realization cannot. It runs one server, so infra.ha.enabled and a non-zero infra.ha.read_replicas halt the build naming the key; the controller-driven realization honors both. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mariadb:1.0.0 - v1.0.0
Recipe
mariadb-operator
Installs the mariadb-operator control plane as an explicit component: a ServiceAccount, the namespaced Role and cluster RBAC the controller needs to reconcile MariaDB custom resources, the configuration the controller reads at startup, and the controller Deployment on the upstream image. Emitting the control plane as a first-class component lets a controller-driven MariaDB ship its runtime prerequisite alongside the MariaDB custom resource it emits (the bundled-deploy prerequisite mode), so the custom resource and the controller that reconciles it are deployed together. This is a minimal control-plane bootstrap. It installs no custom-resource definitions, no certificate controller and no admission webhook, so reach for it only where the definitions are already installed in the cluster and the build must fetch nothing; for a full install reach for `mariadb-operator-chart`, which renders the project's own chart with all three. The controller serves its own metrics at /metrics on port 8080, which checks no caller. With `values.metrics_enabled` true a ClusterIP Service, `<instance>-metrics` in infra.namespace, carries that port and provides.metrics_endpoint publishes its address with the namespace; false, that key is empty and no Service is rendered.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mariadb-operator:1.0.0 - v1.0.0
Recipe
mariadb-operator-chart
Installs the mariadb-operator control plane through the project's own Helm chart, at the chart release whose application version is the controller image's. The chart is official, published by the mariadb-operator maintainers from the project's repository; complete, carrying the controller, its definitions and its admission webhook; and maintained, released with every controller release. Beside what the minimal bootstrap installs — the controller, its configuration and its permissions — the chart renders the twelve custom-resource definitions, the certificate controller that issues and rotates the webhook's serving certificate in the cluster, and the validating admission webhook with its Deployment and Service. Every image is pinned, the controller's three from infra.image, and nothing is generated when the build runs. Reach for this realization for a full install, where the chart repository is reachable from the build; reach for `mariadb-operator` where the custom-resource definitions are already installed in the cluster and the build must fetch nothing. The chart places every namespaced object in the namespace the Release, or else the Aggregate, declares for the component, so the build halts unless that namespace is infra.namespace. With `values.metrics_enabled` true the chart's `<name>-metrics` Service carries the controller's metrics port and provides.metrics_endpoint publishes its address; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mariadb-operator-chart:1.0.0 - v1.0.0
Recipe
loki
Grafana Loki log-aggregation server on Kubernetes via the official Grafana Loki Helm chart (grafana/loki), driven in single-binary mode. Every Loki subsystem runs inside one process backed by a filesystem object store, so the chart renders one StatefulSet with a per-replica data volume, its read/write/backend scaling paths held at zero replicas. The auxiliary services the chart bundles by default — the ingress gateway, the chunk and result caches, and the canary probe DaemonSet — are turned off so this Recipe renders a single workload; those are separate services modelled on their own. A memcached wired to the chunk_cache socket is exactly that case: the chart keeps its own bundled cache off and the recipe points the chunk-cache client at the wired memcached instead, with the batch size, parallelism, timeout and entry lifetime taken from that edge's own values, over TLS to one publishing memcached-tls. Unwired, the client stays off and every chunk is read from the data volume. The Loki server image and the config-reload sidecar image are pinned to resolvable tags for reproducible renders. The developer's retention_period flows into the in-process compactor, which deletes log streams once they age past it. provides.host is the bare in-cluster Service name this Recipe pins the chart's resource names to, with no namespace; a consumer in another namespace qualifies it on its own side. provides.otlp_http_endpoint is that Service's /otlp address, for OpenTelemetry logs, and provides.metrics_endpoint its /metrics address, where the server serves its own metrics in the Prometheus format. The packaged Loki server is AGPL-3.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/loki:1.0.0 - v1.0.0
Recipe
mariadb-cluster
MariaDB realized as a MariaDB custom resource the mariadb-operator control plane reconciles, rather than a directly synthesized StatefulSet — the controller-driven realization of the same MariaDB component, selectable in the Manifest alongside the synthesis recipe. The recipe emits one MariaDB conformant to the k8s.mariadb.com/v1alpha1 API; the controller turns it into a running server, its client Service and its credential Secrets, and creates the initial database and application user from the developer's own values. The MariaDB is written into the namespace the Release, or else the Aggregate, names through the aggrega.ai/namespace annotation, where one is set. provides.host is the client Service the controller creates for the instance and provides.port is the port the custom resource carries, so a consumer wires to it exactly as it would to the synthesized realization. infra.ha turns on replication: infra.ha.read_replicas replicas, automatic failover while infra.ha.enabled, and provides.host then <name>-primary. Credential custody is the operator's choice: infra.credentials generated leaves both passwords with the controller, and the required root_password slot is bound to helm/generated-secret; supplied points it at an existing Secret named in infra.credentials_secret. The control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource reaches a cluster that cannot reconcile it. Reach for this realization where the control plane is in place, and for the synthesis recipe where it is not. A core/backup attachment bound to helm/backup-mariadb-operator is realized through a mariadb-operator Backup.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/mariadb-cluster:1.0.0 - v1.0.0
Recipe
llamaindex
A LlamaIndex agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as LLAMAINDEX_AGENT_MODE, LLAMAINDEX_SYSTEM_PROMPT (while set), LLAMAINDEX_MAX_FUNCTION_CALLS and LLAMAINDEX_LOG_LEVEL by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/llamaindex:1.0.0 - v1.0.0
Recipe
livekit-sip
A LiveKit SIP bridge on Kubernetes via the helm provider's Workload synthesis, running the upstream sip image (no external Helm chart, because the project publishes none for this service). It renders a stateless Deployment plus a ClusterIP Service and declares no volume, because the bridge keeps the state of a call in progress in the shared store rather than on disk. The Service publishes the signalling port over datagrams and over a stream, and fronts signalling and not media: the RTP media range reaches a peer only where the pod uses the node's own network. The bridge reaches the media server and the store over its two edges, the store over TLS when it publishes valkey-tls or redis-tls, checked against the connection's server_ca, mounted, or the image's trust store; the token-signing secret arrives by reference from the credential on the media_server connection, so the plaintext never appears in the rendered manifest. It publishes an in-cluster address only, with external-IP discovery off. With `values.metrics_enabled` true the bridge's metrics are switched on, on port 6789, which joins the container and the ClusterIP Service and checks no caller, and their address is published as provides.metrics_endpoint; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/livekit-sip:1.0.0 - v1.0.0
Recipe
livekit-server
A LiveKit media server on Kubernetes via the helm provider's Workload synthesis, running the upstream livekit-server image (no external Helm chart). It renders a stateless Deployment plus a ClusterIP Service and declares no volume, because the server keeps room and participant state in memory rather than on disk. It is synthesized because LiveKit's own Helm chart fails the maintained test: the chart its published repository serves packages a server older than the one this Component pins. Media travels over one multiplexed UDP port by default. Setting both range bounds selects a contiguous UDP port range instead, which is emitted as no container or service port and carries client traffic only where the pod uses the node's own network. Binding the state_store edge lets several replicas coordinate through the shared store, over TLS to one publishing valkey-tls or redis-tls, checked against the connection's server_ca, mounted, or the image's trust store. The token-signing secret arrives from the api_secret credential slot by reference, so the plaintext never appears in the rendered manifest. It publishes an in-cluster address only: external-IP discovery is off, and reaching clients outside the cluster is the ingress tier's concern. With `values.metrics_enabled` true the server's metrics are switched on, on port 6789, which joins the container and the ClusterIP Service and checks no caller, and their address is published as provides.metrics_endpoint; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/livekit-server:1.0.0 - v1.0.0
Recipe
livekit-agent-worker
A LiveKit agent worker on Kubernetes via the helm provider's Workload synthesis, running an image the Release supplies (no external Helm chart, because the project publishes none for this service, and no pinned image, because it publishes none of those either — the agent framework ships as a language package). It renders a stateless Deployment, and a ConfigMap only for files a connection or the component's own `inject:` block dictates: no Service, because the worker serves no application traffic, and no volume for state, because it keeps a session's state in memory. The worker is configured entirely on the environment, in the variables the framework itself reads: the media server's WebSocket address and API key identifier, composed from the host, port and api_key the media server publishes, the dispatch name, and the log level. The one port declared is the framework's own health endpoint, which is a container port the probes reach on the pod address and which no Service publishes, so the worker exposes nothing to the rest of the Release. The token-signing secret arrives from the credential on the media_server connection as an env secretKeyRef, so the plaintext secret never appears in the worker's own manifest. The public key identifier travels over the same edge, so one key pair is authored once. A media_server connection writing an `inject:` block, or the component's own block, also delivers the names it dictates, as variables or as files mounted at the paths it names, beside the framework's variables. provides.agent_name is the dispatch name the Aggregate authored, republished so a consumer writing a dispatch rule reads it rather than repeating it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/livekit-agent-worker:1.0.0 - v1.0.0
Recipe
litellm
LiteLLM as a self-hosted, OpenAI-compatible model proxy on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). Renders a Deployment, a ClusterIP Service and a ConfigMap carrying the proxy's configuration, which publishes the alias in values.model and points it at values.upstream_model. An attached api_key, provider_key or database_url reaches the proxy as an env secretKeyRef, so no credential value appears in the proxy's own resources. The configuration contract is authored against LiteLLM v1.103.1. On every Release a second ConfigMap carries a request guard, which forwards only the request fields the proxy declares for the route, plus any the Release admits in infra.admitted_request_fields, and refuses the rest naming the field. The same ConfigMap configures a path allowlist, a second container running the official nginx image (BSD-2-Clause) in front of the proxy on infra.port, which answers 403 to every path but the model routes, the model listing, the health routes and key management. Both come from the LiteLLMGuard bundle. PROXY_BASE_URL is emitted only when infra.external_base_url is set, the build halting on one carrying a path. With `values.metrics_enabled` true the proxy serves its Prometheus metrics on its loopback address, and the path allowlist forwards a GET of /metrics alone there from port 9090, on the ClusterIP Service, published as provides.metrics_endpoint; false leaves that key empty. Reach for this realization where the build must fetch nothing; reach for `litellm-chart`, BerriAI's own chart, where its registry is reachable at build time.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/litellm:1.0.0 - v1.0.0
Recipe
litellm-gateway
An AI gateway on Kubernetes via the helm provider's Workload synthesis, running the official upstream proxy image directly. Renders a Deployment, a ClusterIP Service and a ConfigMap carrying the routing table. Every credential reaches the gateway as an env secretKeyRef, so no credential value appears in its own resources. It is proven at LiteLLM v1.103.1 and halts on any other engine_version. On every Release a second ConfigMap carries a request guard, which forwards only the request fields the proxy declares for the route, plus any the Release admits in infra.admitted_request_fields, and refuses the rest naming the field. The same ConfigMap configures a path allowlist, a container running the official nginx image (BSD-2-Clause) on infra.port, which answers 403 to every path but the routes the gateway serves. A core/context-propagation attachment, a core/data-masking one listing what to mask, or values.log_masking listing it, adds the module each relies on. A core/prompt-injection-shield attachment sends every map entry to the guard on the guard socket. PROXY_BASE_URL and the Presidio and cache store addresses appear only when infra.external_base_url and their sockets are set. A store publishing redis-tls is reached over TLS, its password inside the cache's URL: keep that password to letters and digits. With `values.metrics_enabled` true the path allowlist forwards a GET of /metrics alone from port 9090, on the ClusterIP Service, to the gateway's metrics, published as provides.metrics_endpoint; false leaves that key empty. Reach for this realization where the build must fetch nothing; reach for `litellm-gateway-chart`, BerriAI's own chart, where its registry is reachable at build time.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/litellm-gateway:1.0.0 - v1.0.0
Recipe
litellm-gateway-chart
An AI gateway on Kubernetes via BerriAI's own litellm-helm chart, pulled from its OCI registry, running the official upstream proxy image at infra.engine_version. The chart's README calls it community maintained; it was checked for being official, complete and maintained against the chart itself. Its bundled PostgreSQL and Redis stay off, and its schema-migration Job renders only with database_url attached. It is proven at LiteLLM v1.103.1 and halts on any other engine_version. The gateway's configuration, its routing table and the modules Features rely on, comes from the LiteLLMGatewayConfig bundle, and the request guard and the path allowlist in front of the gateway from the LiteLLMGuard bundle, exactly as on the synthesis, the allowlist as a second container of the chart's pod. Every credential reaches the gateway as a secretKeyRef, so no credential value is rendered; unattached, api_key leaves the gateway with no master key, as on the synthesis. A cache store publishing redis-tls is reached over TLS, its certificate checked against the connection's server_ca or the image's trust store and for the store's published host; its password is expanded into the cache's connection URL in the container, so keep it to letters and digits. With `values.metrics_enabled` true the allowlist forwards a GET of /metrics alone from port 9090, on the chart's metrics Service, published as provides.metrics_endpoint; false leaves that key empty. Reach for this realization where a chart registry is reachable at build time; reach for `litellm-gateway`, the workload synthesis, where the build must complete with nothing fetched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/litellm-gateway-chart:1.0.0 - v1.0.0
Recipe
litellm-chart
LiteLLM as a self-hosted, OpenAI-compatible model proxy on Kubernetes via BerriAI's own litellm-helm chart, pulled from its OCI registry, running the official upstream image at infra.engine_version. The chart's README calls it community maintained; it was checked for being official, complete and maintained against the chart itself. Its bundled PostgreSQL and Redis stay off, and its schema-migration Job renders only with database_url attached. A ConfigMap of the Recipe's own, mounted by the chart, carries the proxy's configuration, publishing the alias in values.model and pointing it at values.upstream_model. An attached api_key, provider_key or database_url reaches the proxy by a secretKeyRef to its slot's Secret, so no credential value is rendered. Unattached, api_key leaves the proxy with no master key, as on the synthesis. The request guard and the path allowlist in front of the proxy come from the LiteLLMGuard bundle, exactly as on the synthesis, the allowlist as a second container of the chart's pod. With `values.metrics_enabled` true the allowlist forwards a GET of /metrics alone from port 9090, on the chart's metrics Service, published as provides.metrics_endpoint; false leaves that key empty. Reach for this realization where a chart registry is reachable at build time; reach for `litellm`, the workload synthesis, where the build must complete with nothing fetched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/litellm-chart:1.0.0 - v1.0.0
Recipe
lindy
A Lindy assistant on Lindy's managed platform as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the operator-supplied platform endpoint. provides.endpoint is that host and provides.lindy_id is the developer-selected assistant. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/lindy:1.0.0 - v1.0.0
Recipe
letta
The Letta stateful-agent platform as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the platform's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to the component's slot is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the platform. The credential slot is required here: a build leaving it empty halts naming it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/letta:1.0.0 - v1.0.0
Recipe
langfuse
Langfuse as a self-hosted tracing and prompt-registry platform on Kubernetes via the helm provider's Workload synthesis, running the official upstream images directly (no external Helm chart). Renders one Deployment carrying the console and ingestion API and the asynchronous worker, and a ClusterIP Service fronting the console. The configuration contract is authored against Langfuse 4.22.0. The relational backend, the cache and the object store arrive as topology edges; the columnar analytics store is supplied as operator-set pointer strings, and the build halts until both of its addresses are set. infra.external_base_url reaches the console as NEXTAUTH_URL. An attached credential, on a connection or in a Component slot, reaches the containers as an env secretKeyRef, never as a value in the manifest. The encryption key, the session secret, the salt and the object store's secret key have no default: the first three slots are required, and an event_store connection with no credential halts the build unless its values set unauthenticated: true. The backend password, left unattached, renders a development literal; the cache and analytics-store passwords render nothing. A cache publishing valkey-tls or redis-tls is reached over TLS, its certificate checked against the cache connection's server_ca, mounted, or the runtime's trust store; another TLS protocol on the cache halts the build. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `langfuse-chart`, which renders the platform maintainers' own chart in separate Deployments, where the console and the worker will be scaled independently.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/langfuse:1.0.0 - v1.0.0
Recipe
langfuse-chart
Langfuse as a self-hosted tracing and prompt-registry platform on Kubernetes via the platform maintainers' own Helm chart, served from their HTTPS chart repository. The chart's bundled PostgreSQL, Valkey, object-store and ClickHouse parts are switched off: the backend, the cache and the object store arrive as topology edges, and the analytics store as operator-set pointer strings. It renders separate Deployments for the console and the asynchronous worker, a ClusterIP Service fronting the console, a ServiceAccount and the Secret holding the chart's own keys. Separate Deployments let an operator scale or restart the console or the worker without the other, and the chart carries the platform's schema-migration switches. Reach for this realization where a chart repository is reachable at build time and the console and the worker will be scaled independently; reach for `langfuse`, the workload-synthesis realization, where the build must complete with nothing fetched. The platform image is pinned on the chart's shared image block, so the console and the worker run one version, and the chart version in source.version. infra.external_base_url reaches the console as the chart's `langfuse.nextauth.url`. The configuration contract is authored against the chart's own application version. A cache publishing valkey-tls or redis-tls is reached through the chart's own Redis TLS setting, checked against the cache connection's server_ca, mounted, or the runtime's trust store; another TLS protocol on the cache halts the build. The object store's secret key has no default: an event_store connection with no credential halts the build unless its values set unauthenticated: true.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/langfuse-chart:1.0.0 - v1.0.0
Recipe
kubestash-chart
Installs KubeStash through AppsCode's own Helm chart, from AppsCode's chart repository, at the chart release infra.source_version names; the build halts when it is unset. It renders the operator and its webhook server, their permissions, KubeStash's custom-resource definitions, and the addons and functions the backup and restore jobs run. The licence is read from the Secret the license slot names, under the data key key.txt, and the build halts on any other key. The webhook server's certificate is issued by cert-manager, which must already run in the cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kubestash-chart:1.0.0 - v1.0.0
Recipe
kong
Kong, a database-backed API gateway, on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the proxy port). The PostgreSQL backend is the system of record, so the gateway renders stateless. Its host, port, user and database arrive on .context.inject.env under the four KONG_PG_* variables the `backend` edge's inject block names; an edge leaving any missing halts the build naming the edge and the variables. A migrations step runs to completion before the gateway serves. With no credential on the backend connection the Recipe uses an inline development default; with one, KONG_PG_PASSWORD comes from the credential subsystem. Each model wired to the models socket gets a chat and a completion route through Kong's AI proxy, written as Kong's declarative configuration in a ConfigMap that a second step imports into the database before the gateway starts. A connection's credential arrives as an env secretKeyRef and the routes name it through Kong's environment vault, so no credential value is in the rendered output. Reach for this realization where the build must complete with nothing fetched; reach for `kong-chart` where the gateway will be upgraded in place, since it migrates the schema on either side of an upgrade. With `values.metrics_enabled` true the Service also carries the Status API port, 8007, and provides.metrics_endpoint publishes /metrics on it; false, that key is empty. That port, open on the pod for the probes, answers only a GET of /status, /status/ready or /metrics from another address: the KongStatusGuard bundle refuses the rest, POST /status/ready among them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kong:1.0.0 - v1.0.0
Recipe
kong-enterprise-chart
Kong Gateway Enterprise on Kubernetes via the official Kong Helm chart, served from that project's HTTPS chart repository, with the chart's enterprise settings on. The gateway's configuration lives in the PostgreSQL backend wired on the backend socket, so the chart's bundled database, ingress controller, manager console and developer portal are off. It renders a Deployment, a ServiceAccount, ClusterIP Services for the proxy data plane and the in-cluster Admin API, the chart's Kong custom-resource definitions, and the three schema-migration Jobs Kong sequences around a release. The image is the one infra.image names and no other. infra.image_pull_secret becomes the chart's image pull secret. The licence reaches every Kong container as KONG_LICENSE_DATA, from a secretKeyRef to the Secret the license slot names, so no licence material is in the rendered output. Each model wired to the models socket gets a chat and a completion route through Kong's AI proxy, the routes core/gateway/kong serves, in a ConfigMap that a step in the gateway's own pod imports into the database. It is the component's only realization. Any of the six AI Features halts the build until infra.ai_gateway_entitled affirms the licence's AI entitlement. With `values.metrics_enabled` true a ClusterIP Service, `<rendered_name>-status`, carries the chart's status port, 8100, and provides.metrics_endpoint publishes /metrics on it; false, neither. That port, open on the pod for the probes, answers only a GET of /status, /status/ready or /metrics from another address: the KongStatusGuard bundle refuses the rest, POST /status/ready and any route the image adds among them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kong-enterprise-chart:1.0.0 - v1.0.0
Recipe
kong-chart
Kong, a database-backed API gateway, on Kubernetes via the official Kong Helm chart, served from that project's HTTPS chart repository. The gateway's configuration lives in the PostgreSQL backend wired on the `backend` uses-socket, so the chart's bundled database, ingress controller and manager console are switched off. It renders a Deployment, a ServiceAccount, ClusterIP Services for the proxy data plane and the in-cluster Admin API, the chart's Kong custom-resource definitions, and the three schema-migration Jobs the product's maintainers sequence around a release: a bootstrap, an up-migration before an upgrade and a finishing migration after one. Reach for this realization where a chart repository is reachable at build time and the gateway will be upgraded in place; reach for `kong`, the workload-synthesis realization, where the build must complete with nothing fetched. Each model wired to the models socket gets a chat and a completion route through Kong's AI proxy, in a ConfigMap beside the chart that an init container imports into the database, retried on a first install until the bootstrap Job provisions the schema. A connection's credential arrives as an env secretKeyRef the routes name through Kong's environment vault, so no credential value is in the rendered output. With `values.metrics_enabled` true a ClusterIP Service, `<rendered_name>-status`, carries the chart's status port, 8100, and provides.metrics_endpoint publishes /metrics on it; false, neither. That port, open on the pod for the probes, answers only a GET of /status, /status/ready or /metrics from another address: the KongStatusGuard bundle refuses the rest, POST /status/ready among them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kong-chart:1.0.0 - v1.0.0
Recipe
kong-ai-pii-service
Kong's AI PII service on Kubernetes via the helm provider's Workload synthesis, running the image infra.image names: a stateless Deployment plus a ClusterIP Service on the port the service answers on. Synthesized because Kong publishes no chart for the service. It keeps no state, so it renders no volume. The image comes from infra.image alone, and the build halts naming infra.image when it is unset or empty; infra.image_pull_secret becomes the pod's image pull secret. The service listens on 8080 over plain HTTP, the port Kong's own run command publishes. It serves HTTPS on 8443 only when started with a certificate and a key, which this Recipe does not mount, so it publishes http. Readiness is asked of GET /llm/v1/status, which the service answers once it is up; it loads its language model as it starts, fetching one its image does not carry. Memory is requested and limited at infra.compute.memory, 1Gi unless the Release sets another, above the 600MB Kong asks to be free for the one model the service loads by default. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kong-ai-pii-service:1.0.0 - v1.0.0
Recipe
keycloak
Keycloak, a database-backed identity and access-management server, on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the HTTP port). Keycloak holds no local state — the PostgreSQL backend is the system of record — so it renders stateless, and it provisions and migrates its own schema on startup. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `keycloak-chart`, which renders the codecentric Keycloak.X chart, where the server will be scaled out — that realization carries the ordered pods, the headless Service and the replicated cache stack a single synthesized Deployment does not. With `infra.external_base_url` set, the server's hostname is that full URL, so the issuer and every address the server advertises follow it; unset, the server takes its address from each incoming request. With no credential attached — on the backend connection for the database password, on the admin_password slot for the bootstrap admin password — each takes an inline development default; with one attached, each is supplied via the credential subsystem. With `values.metrics_enabled` true, the server's metrics are switched on and the management port joins the Service, so its metrics and health endpoints, which check no caller, answer at the Service name; the address is published as provides.metrics_endpoint. False, that key is empty and the port stays off the Service.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/keycloak:1.0.0 - v1.0.0
Recipe
knative-serving
Installs the Knative Serving control plane as an explicit component: two ServiceAccounts, the cluster permissions each of them needs, the configuration entry carrying the request-queue sidecar image, and the four control-plane Deployments with the in-cluster Services through which they reach one another. Emitting the control plane as a first-class component lets a realization that scales a workload to zero ship its runtime prerequisite alongside the custom resources it emits (the bundled-deploy prerequisite mode), so the custom resource and the controllers that reconcile it are deployed together. This is a minimal control-plane bootstrap. It does not install the custom-resource definitions, the ingress and networking layer, the admission-webhook registrations, or the configuration bodies the remaining defaults are edited in; a production install sources the project's own release manifests, which carry all of those. With `values.metrics_enabled` true it also emits the observability configuration naming the Prometheus format, so each of the four processes serves its own metrics, and provides publishes the pod label, port name and path a scraper discovers them by; false, neither.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/knative-serving:1.0.0 - v1.0.0
Recipe
keycloak-chart
Keycloak, a database-backed identity and access-management server, on Kubernetes via the codecentric Keycloak.X Helm chart, served from that project's HTTPS chart repository. What it encodes that a hand-assembled workload does not is the cluster-forming configuration: a StatefulSet with ordered pod management behind a headless Service, the database-backed cache stack the server replicates sessions over, and probes aimed at the server's own management endpoint. The chart's Ingress, Route, autoscaler, network policy and monitoring resources are left switched off, and the server image and the chart version are pinned. Reach for this realization where a chart repository is reachable at build time and the server will be run as more than one replica; reach for `keycloak`, the workload-synthesis realization, where the build must complete with nothing fetched. With `infra.external_base_url` set, the server's hostname is that full URL, delivered through the chart's `extraEnv`, so the issuer and every address the server advertises follow it; unset, the server takes its address from each incoming request. The chart serves the server under the path `/auth`, so an ingress in front of it maps the address's path onto that one. The chart serves the server's metrics, at /auth/metrics on the management port its console Service carries, whatever `values.metrics_enabled` says; that port checks no caller. True publishes the address as provides.metrics_endpoint; false leaves that key empty. That Service, and the management port with it, stays inside the cluster whatever `infra.overrides.service_type` says; LoadBalancer or NodePort publishes the console alone on `<rendered_name>-external`.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/keycloak-chart:1.0.0 - v1.0.0
Recipe
keda
Installs the KEDA control plane as an explicit component: a ServiceAccount, the namespaced and cluster permissions the operator needs to reconcile the scaling custom resources and the adapter needs to answer for them, the operator and metrics-adapter Deployments with the Services through which they reach one another, and the API-service registration that puts the adapter behind the cluster's external-metrics API. Emitting the control plane as a first-class component lets an event-driven realization ship its runtime prerequisite alongside the custom resources it emits (the bundled-deploy prerequisite mode), so the custom resource and the operator that reconciles it are deployed together. This is a minimal control-plane bootstrap installing no custom-resource definitions and no admission webhooks, and at the images the Component pins it does not run: the operator and the metrics adapter both stop at start because neither Deployment sets the WATCH_NAMESPACE variable they require. Reach for `keda-chart`, which renders the project's own chart. provides publishes the names of the two Services, their metrics port name and the path a scraper discovers the operator's and the adapter's metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/keda:1.0.0 - v1.0.0
Recipe
keda-chart
Installs the KEDA control plane through the project's own Helm chart, at the chart release whose application version is the one the operator and metrics-adapter images come from. The chart is official, published by the KEDA maintainers; complete, carrying everything the project's install needs; and maintained, released with every KEDA release. Beside what the minimal bootstrap installs — the operator, the metrics adapter, their Services, their permissions and the external-metrics API registration — the chart renders the six KEDA custom-resource definitions, the admission-webhook Deployment, its Service and the validating webhook configuration that checks scaling resources as they are written. The operator issues the certificates the adapter and the webhooks serve on when it starts, so nothing is generated when the build runs. Reach for this realization; its chart repository must be reachable from the build. The alternative, `keda`, installs no definitions and at the images the Component pins does not start. The chart places every namespaced object in the namespace the Release, or else the Aggregate, declares for the component, so the build halts unless that namespace is infra.namespace. provides publishes the names of the operator's and the adapter's Services, their metrics port name and the path a scraper discovers their metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/keda-chart:1.0.0 - v1.0.0
Recipe
kasten-chart
Installs Veeam Kasten through Kasten's own Helm chart, k10, from Kasten's chart repository, at the chart release infra.source_version names, 9.0.9 by default. It renders Kasten's services, their permissions and volume claims, its aggregated APIs, its bundled Prometheus, and the free edition licence the chart carries; Kasten registers its policy and profile resource definitions itself when it starts. It renders no policy and no profile: helm/backup-kasten writes those for each core/backup attachment. The chart places its objects in the namespace the Release, or else the Aggregate, declares for the component, and Kasten reads its profiles and their Secrets only from its own namespace, so the build halts unless that namespace is kasten-io. The chart generates a new certificate for Kasten's data mover on every render, so the two Secrets it would write are left out, and the Secrets the data_mover_cert and data_mover_key slots name take their place: the build halts unless they name kopia-tls-cert under tls.crt and kopia-tls-key under tls.key, the names Kasten reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kasten-chart:1.0.0 - v1.0.0
Recipe
kafka
Apache Kafka single-node KRaft on Kubernetes via the helm provider's Workload synthesis (a StatefulSet with a per-replica data volume, a ClusterIP Service exposing the client bootstrap port, and a headless governing Service). Nothing has to be installed in the cluster first and no chart repository has to be reachable, so this is the realization to reach for where no Kafka control plane runs and where the build has to complete without network access; reach for the controller-driven recipe where the Strimzi Cluster Operator is already in place. The node runs the broker and controller roles in one process (no ZooKeeper) and stores its commit log on the persistent volume at /var/lib/kafka/data so log data survives restarts. The cluster identity is supplied verbatim through CLUSTER_ID from the developer value, and the image reads it as-is with no random generation, so the rendered manifest is byte-identical across builds. KAFKA_ADVERTISED_LISTENERS advertises the in-cluster Service name so clients reconnect through the stable address rather than the pod's own hostname. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The image serves no Prometheus metrics, so the metrics keys are published empty and values.metrics_enabled true halts the build. The Java heap is half infra.compute.memory, in whole Mi or Gi. With infra.client_tls true the client port serves TLS alone from the Secret attached to server_certificate, mounted as files, and protocol is kafka-ssl.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kafka:1.0.0 - v1.0.0
Recipe
kafka-strimzi
Apache Kafka realized as a Strimzi custom resource the Strimzi Cluster Operator reconciles, rather than a directly synthesized StatefulSet — the operator-driven realization of the same Kafka component, selectable in the Manifest alongside the synthesis recipe. The recipe emits a Kafka custom resource (and its node-pool) in the KRaft shape at the Strimzi `v1` API, authored against the Strimzi Cluster Operator's 1.3.0 release; the operator turns it into a running broker. The Strimzi Cluster Operator is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. Reach for this realization where the control plane is already in place and the cluster should own broker upgrades, rolling restarts and partition rebalancing; reach for the synthesis recipe where it is not. It runs Kafka 4.2.0 to 4.3.1, so the build halts on any other infra.engine_version, and on a values.cluster_id or infra.controller_port other than the default, which the operator decides itself. With values.metrics_enabled true the Kafka resource turns on the Strimzi metrics reporter on every node, which serves /metrics on port tcp-prometheus and checks no caller, and provides publishes the pod label, port name and path a scraper discovers the nodes by; false, those keys are empty and nothing else changes. With infra.client_tls true the client listener, named tls, serves TLS alone and protocol is kafka-ssl. A consumer trusts ca.crt of the Secret <instance>-cluster-ca-cert through server_ca; one in another namespace holds a copy of it there.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/kafka-strimzi:1.0.0 - v1.0.0
Recipe
jaeger
Jaeger single-node tracing on Kubernetes via the helm provider's Workload synthesis (a StatefulSet with a per-replica data volume, plus a ClusterIP Service exposing the OTLP receivers and the query UI). Jaeger v2 is an OpenTelemetry Collector distribution configured by a YAML config file rather than the legacy environment variables. That config is delivered as a file: the recipe emits it as a ConfigMap via artifacts: and mounts it into the container through a configmap volume at /etc/jaeger, and the container reads it with --config /etc/jaeger/config.yaml. The embedded key-value store writes to the persistent volume at /badger (keys under /badger/key, values under /badger/data) so trace data survives restarts. The OTLP receiver endpoints in the config track the recipe's OTLP port sizing so the process listens on the same ports the Service publishes. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). It is synthesized because Jaeger's own Helm chart fails the complete test: it keeps no trace store on a volume, which the on-disk store this Component describes needs. With `values.metrics_enabled` true the config has Jaeger serve its own metrics at port 8888 on every interface, which checks no caller and joins the ClusterIP Service, and provides.metrics_endpoint publishes the address; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/jaeger:1.0.0 - v1.0.0
Recipe
k8up-chart
Installs K8up through the project's own Helm chart, at the chart release whose image is the one the Component pins. The chart is official, published by the K8up maintainers from the project's repository; complete, carrying the custom-resource definitions, the operator and its permissions; and maintained, its newest release carrying K8up 2.16.1. It renders the operator Deployment, its metrics Service, the custom-resource definitions, the cluster roles the operator and its backup jobs run under, and the chart's post-install cleanup Job. It renders no schedule: helm/backup-k8up writes one beside each datastore carrying a core/backup attachment. The operator acts on K8up resources in every namespace, so the Release may place it in any namespace. Its chart repository must be reachable from the build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/k8up-chart:1.0.0 - v1.0.0
Recipe
huggingface
Hugging Face's managed Inference API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/huggingface:1.0.0 - v1.0.0
Recipe
harbor-chart
A Harbor container registry on Kubernetes via the product's own Helm chart, served from its project's HTTPS chart repository. The chart deploys the whole registry as one unit, in the order its services need, with the wiring the product's own maintainers keep in step with it. Its Ingress and transport security are left switched off, and the reverse proxy is published as one cluster-internal Service. `values.metrics_enabled` turns the chart's metrics on and publishes the four Services serving them. The same chart setting opens the blob store's debug listener — profiling, execution traces and runtime variables, with no caller check — to every in-cluster caller. Every image the chart runs is pinned to a resolvable release tag — one tag, which infra.engine_version selects for all of them — and the chart version is pinned in source.version. Every secret the chart would otherwise generate at render time is supplied instead, so the output is byte-identical across renders: the token-signing key and its certificate, the services' three shared secrets, the request-forgery key and the key stored credentials are encrypted with always from the attached credential by name, the administrator's password by name where attached, and every password the Release leaves unset from a fixed development value. No key is written into the rendered artifact. The address the registry reports as its own — the token-service address it returns to a container client and the commands its console shows — is the chart's `externalURL`: `infra.external_base_url` when set, and the build halts when that address carries a path; unset, the published in-cluster Service address.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/harbor-chart:1.0.0 - v1.0.0
Recipe
haystack
A Haystack agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. Nothing is delivered from the `document_store` connection by default. A connection writing an `inject:` block receives the names it dictates instead of the defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as HAYSTACK_PIPELINE_YAML_PATH and HAYSTACK_MAX_LOOPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/haystack:1.0.0 - v1.0.0
Recipe
higress-chart
Higress on Kubernetes via the project's own `higress-core` Helm chart: the controller with its discovery service and the gateway, each a Deployment behind a ClusterIP Service, their access control (a cluster-wide grant to read Secrets), the IngressClass the controller watches and the custom-resource definitions Higress is configured through. Neither the console nor the bundled Redis is rendered. The controller watches only the gateway's namespace, under an IngressClass named for the gateway and that namespace, so Higress gateways in different namespaces stay apart; two in one namespace collide, because Higress names some of its resources itself. With a model wired it emits the gateway's registry of the models' hosts, an ingress per alias and the AI proxy, pinned by tag and digest, with one rule per connection naming its credential's Secret, never the key; each ingress drops the AI proxy's copy of a caller's key. Over https the gateway checks that a model's certificate chains to the system's trust anchors and names the model's host; an alias mixing http and https connections halts the build. While core/data-masking, core/period-budget or core/audit-logging is attached, it adds Higress's plugin for each onto the model routes: request masking, with what it cannot read refused unless a client crafts the body to get past it, a token limit counted in the store on the cache socket, reached in plaintext, so a store served over TLS halts the build, and statistics in the access log. While core/retry-policy is attached, it writes the route's own retry onto each model route as Ingress annotations. provides also publishes the pod label, port name and path a scraper discovers the gateway's metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/higress-chart:1.0.0 - v1.0.0
Recipe
gumloop
A Gumloop automation pipeline on Gumloop's managed platform as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the managed endpoint. provides.endpoint is the managed API host and provides.pipeline_id is the developer-selected pipeline. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/gumloop:1.0.0 - v1.0.0
Recipe
griptape
A Griptape agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL; the store wired into the `task_memory` socket contributes its host and port, so off-prompt tool output has somewhere to live; a model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. A task-memory store served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as GRIPTAPE_OFF_PROMPT and GRIPTAPE_STREAM by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/griptape:1.0.0 - v1.0.0
Recipe
grpc-knative
The gRPC-calling tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The location of the protobuf definition is delivered into the container, so the methods the tool offers are read from the service's own contract rather than fixed by this recipe. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, and it describes how a caller reaches the tool rather than how the tool reaches the service it calls. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/grpc-knative:1.0.0 - v1.0.0
Recipe
grafana
Grafana dashboards server on Kubernetes via the helm provider's Workload synthesis, running the official upstream grafana/grafana image directly (no external Helm chart). Grafana keeps its own state — users, saved dashboards, preferences — under /var/lib/grafana, so the workload mounts a persistent volume at that path; the process itself is a single stateless replica, so it renders a Deployment plus a ClusterIP Service. The admin username is a developer value. The admin password is wired through the optional `admin_password` credential slot: when a credential is attached it renders as an env secretKeyRef (GF_SECURITY_ADMIN_PASSWORD), so the plaintext never lands in the Deployment manifest; when none is attached the Recipe falls back to a deterministic development default so the render stays byte-identical across runs. With `infra.external_base_url` set, it reaches the server as its root URL (GF_SERVER_ROOT_URL); unset, that variable is not emitted and the server keeps its default. The server listens on `infra.port`, which also carries its own metrics in the Prometheus format at /metrics, published as provides.metrics_endpoint. The packaged Grafana server is AGPL-3.0. Reach for this realization where the build must complete with nothing fetched, or where no chart repository is reachable from it; reach for `grafana-chart`, which renders the community-maintained Grafana chart, where data sources or dashboards are to be provisioned with the server — that realization carries the provisioning surface this one has nowhere to declare.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/grafana:1.0.0 - v1.0.0
Recipe
grafana-chart
Grafana dashboards server on Kubernetes via the community-maintained Grafana Helm chart, served from that project's HTTPS chart repository. It renders one server with a persistent volume for its own state — a Deployment, a ClusterIP Service, a PersistentVolumeClaim, the server configuration, the admin account's Secret and the account it runs under — with the chart's ingress, route, autoscaling, network-policy, monitoring and image-renderer resources left switched off. The server image and the chart version are pinned. What it carries that a hand-assembled workload does not is the provisioning surface: data sources and dashboards declared through infra.overrides, so an operator hands the server one without an application change. Reach for this realization where a chart repository is reachable at build time and data sources or dashboards are to be provisioned with the server; reach for `grafana`, the workload-synthesis realization, where the build must complete with nothing fetched. With `infra.external_base_url` set, it becomes the server's root URL through the chart's own `grafana.ini` value; unset, that value is not set. The server's own metrics, in the Prometheus format at /metrics, are served through the same Service and port and published as provides.metrics_endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/grafana-chart:1.0.0 - v1.0.0
Recipe
gemini
Google's managed Gemini API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/gemini:1.0.0 - v1.0.0
Recipe
gemini-agent
A Gemini assistant on Google's generative-language API as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the managed endpoint. provides.endpoint is the managed API host and provides.model is the developer-selected model string. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/gemini-agent:1.0.0 - v1.0.0
Recipe
gcp-credentialed-service
Runs the service on Kubernetes via the helm provider's Workload synthesis and delivers its Google Cloud credential as a mounted file. The `gcp_key` secret slot mounts as a read-only file under /var/secrets/google — the whole Secret mounts as a directory, so the credential's file name is the Secret's data key — and GOOGLE_APPLICATION_CREDENTIALS points at that file, the path the Google client libraries read the service-account key from. The credential path is derived from the slot's published data key, so it tracks whatever key the attached secret publishes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/gcp-credentialed-service:1.0.0 - v1.0.0
Recipe
function-knative
The generic event-driven compute primitive realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the function occupies no compute between the calls that invoke it. This is the request-paced of the primitive's two realizations: work arrives as an inbound call. Where work arrives on a queue, the Manifest maps the component to the queue-paced realization instead. No Service definition and no provides contract are emitted, because the component declares none: the addressable name belongs to the control plane that creates it, and this primitive publishes nothing across an edge. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. An edge's inject block dictates what reaches the container, since the image is the Release's own code: each env entry a variable under that name, beside FUNCTION_RUNTIME, which no edge may dictate while it is the default, and each file entry a read-only file from a ConfigMap emitted beside the Service. The component's own inject block delivers env entries too, and one reading runtime alone replaces FUNCTION_RUNTIME, its default name. An arg entry, or any other mechanism, is refused by name: the recipe renders no command line. A credential published for the pod to reference, such as core/referenced-secret, halts the build by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/function-knative:1.0.0 - v1.0.0
Recipe
function-keda
The generic event-driven compute primitive realized as a Deployment paired with a KEDA scaling resource the KEDA control plane reconciles. The recipe synthesizes the workload and emits one ScaledObject beside it; the control plane reads the depth of the event source the Release names and sets the replica count from it, zero included, so the function occupies no compute while the source is empty. This is the queue-paced of the primitive's two realizations: work arrives on an event source the Release names, rather than as an inbound call. Where work arrives as an inbound call, the Manifest maps the component to the request-paced realization instead. The two are chosen there and never both at once. No Service definition and no provides contract are emitted, because the component declares none and because the event source invokes the function rather than a caller reaching it. The KEDA control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. An edge's inject block dictates what reaches the container, since the image is the Release's own code: each env entry a variable under that name, beside FUNCTION_RUNTIME, which no edge may dictate while it is the default, and each file entry a read-only file from a ConfigMap. The component's own inject block delivers env entries too, and one reading runtime alone replaces FUNCTION_RUNTIME, its default name. An arg entry, or any other mechanism, is refused by name: the recipe renders no command line. A credential published for the pod to reference, such as core/referenced-secret, halts the build by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/function-keda:1.0.0 - v1.0.0
Recipe
fms-guardrails-orchestrator
The FMS Guardrails Orchestrator on Kubernetes via the helm provider's Workload synthesis, running the image in infra.image directly: a stateless Deployment, a ClusterIP Service on the port its API answers on, and a ConfigMap carrying its configuration file. Synthesized rather than wrapping a chart because the project publishes none. The configuration names the upstream server in infra.upstream_host and infra.upstream_port as the one chat requests are forwarded to, and registers the detector wired on the detector socket, at the host and port it publishes, under values.detector_id with values.default_threshold. The file is mounted on its own, at a path the image does not use, and the ORCHESTRATOR_CONFIG variable points the orchestrator at it; the ports reach it through the HTTP_PORT and HEALTH_HTTP_PORT variables it reads. Its health route answers as soon as it serves, on the health port, so readiness and liveness both ask it there. provides.host is the bare, self-contained in-cluster service name of the synthesized Service; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/fms-guardrails-orchestrator:1.0.0 - v1.0.0
Recipe
fluentd
Fluentd log aggregator on Kubernetes via the helm provider's Workload synthesis, running the official upstream fluent/fluentd image directly (no external Helm chart). Fluentd is the receiving end: a minimal fluent.conf with a `forward` input source bound to the infra port is written at start-up, and records whose tag matches the developer's output match are routed to standard output. Fluentd is then exec'd against that config. The buffer is in-memory by default, so the aggregator renders as a stateless Deployment with no persistent volume; when infra.storage.size is set a durable volume backs a file buffer mounted at /fluentd/log. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed; a consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). provides.port is the forward input port upstream forwarders ship records to. Reach for this realization where the build must complete with nothing fetched; reach for `fluentd-chart`, which renders the Fluent project's own Helm chart as the same aggregator, where a chart repository is reachable at build time. The Fluentd server is Apache-2.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/fluentd:1.0.0 - v1.0.0
Recipe
fluent-bit
Fluent Bit log forwarder on Kubernetes via the helm provider's Workload synthesis, running the official upstream fluent/fluent-bit image directly (no external Helm chart). The sink host and port resolved from the `forward_target` uses-edge arrive on .context.inject.env and are written into a minimal fluent-bit.conf `forward` output at start-up, so the forwarder ships records to that sink; Fluent Bit is then exec'd against that config with its monitoring endpoint bound to the infra port, which also serves its metrics in the Prometheus format at /api/v2/metrics/prometheus, published as provides.metrics_endpoint. Fluent Bit holds no persistent state, so it renders as a stateless Deployment. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed; a consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). Reach for this realization where the build must complete with nothing fetched, or where a single forwarding pod with its own declared input is what is wanted; reach for `fluent-bit-chart`, which renders the official Fluent chart, where a chart repository is reachable and records are to be collected from the cluster's own nodes — that realization carries the per-node collector shape and the cluster read grants this one does not carry. The Fluent Bit server is Apache-2.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/fluent-bit:1.0.0 - v1.0.0
Recipe
fluentd-chart
Fluentd log aggregator on Kubernetes via the Fluent project's own fluentd Helm chart, served from that project's HTTPS chart repository. The chart deploys a log-collecting DaemonSet by default; here its values configure it as the aggregator the Component describes: one replica running the official fluent/fluentd image at infra.engine_version, a `forward` input on the infra port, the records the developer's output match selects routed to standard output, no host log mounts and no cluster read grants. The chart was checked for being official, complete and maintained against the chart itself. The buffer is in-memory by default, so the chart renders a Deployment with no persistent volume. When infra.storage.size is set the chart renders a StatefulSet whose claim of that size backs a file buffer mounted at /fluentd/log, as on the synthesis. provides.host is the chart's Service, with no namespace; it carries the forward port beside the chart's own metrics port. Reach for this realization where a chart repository is reachable at build time; reach for `fluentd`, the workload-synthesis realization, where the build must complete with nothing fetched. Default chart version in source.version; override via infra.source_version. The Fluentd server is Apache-2.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/fluentd-chart:1.0.0 - v1.0.0
Recipe
fluent-bit-chart
Fluent Bit log processor and forwarder on Kubernetes via the official Fluent Helm chart, served from that project's HTTPS chart repository. The chart deploys Fluent Bit in its node-collector shape: one pod per node reading the container log files that node writes, a filter that enriches every record with the pod, namespace and label metadata it reads back from the cluster, and the ServiceAccount, ClusterRole and ClusterRoleBinding that filter needs to do so — a grant a forwarder assembled as a plain workload lacks, so it can neither read the node's log files nor enrich a record with the workload it came from. The sink wired on the `forward_target` uses-socket becomes the `forward` output, exactly as in the workload-synthesis realization. Reach for this realization where a chart repository is reachable at build time and log records are to be collected from the cluster itself; reach for `fluent-bit`, the workload-synthesis realization, where the build must complete with nothing fetched, or where a single forwarding pod with its own inputs is what is wanted rather than a per-node collector. The forwarder image tag is pinned. Default chart version in source.version; override via infra.source_version. provides.host and provides.metrics_service name the in-cluster Service, with no namespace, for a scraper to discover every node's collector behind it; metrics_endpoint stays empty, as one address reaches one node per scrape. That Service carries the monitoring port alone, which checks no caller, so it stays inside the cluster: an `infra.overrides.service_type` other than ClusterIP halts the build. The Fluent Bit server is Apache-2.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/fluent-bit-chart:1.0.0 - v1.0.0
Recipe
file-parser-knative
The document-parsing tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The developer's chosen engine and extraction mode reach the running tool as environment variables, so how much structure survives a parse is fixed at authoring time rather than at call time. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, so a consumer wires to the tool's declared contract rather than to anything Knative-specific. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/file-parser-knative:1.0.0 - v1.0.0
Recipe
filesystem
The file-tree tool realized on Kubernetes by the helm provider's Workload synthesis, rendered offline as plain manifests. The Workload becomes a StatefulSet whose tree lives on a per-instance persistent volume, mounted where the Release's image reads and writes it, with a ClusterIP Service fronting the port the tool listens on and a headless Service carrying the instance's stable network identity. It runs one instance, a condition of the tool being correct rather than a starting size: two instances behind one address would each hold their own copy of the tree, so a file written through one would be missing from the other. The count is pinned here for that reason and is not exposed as an operational knob. A read-only access mode reaches the mount as well as the tool, so a tree the Aggregate declared read-only cannot be written even by an image that would try. The developer's access mode, declared effect and timeout reach the running tool as environment variables, together with the directory the tree is mounted at. The image is the Release's, because this tool packages no upstream product: a build with none set halts and names the key. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the StatefulSet, never inside the tree. provides is the synthesized ClusterIP Service, so a consumer wires to a host, port, endpoint and protocol and never to the file tree behind them. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/filesystem:1.0.0 - v1.0.0
Recipe
ferretdb
A MongoDB-wire-compatible document-database proxy on Kubernetes via the helm provider's Workload synthesis. The proxy process holds no local state, so it renders a stateless Deployment plus a ClusterIP Service on the wire port. The PostgreSQL backend identity resolved from the `backend` uses-edge (host, port, user, database) arrives on .context.inject.env and is composed into the proxy's PostgreSQL connection URL; the URL is rendered password-free and the password is supplied separately on the standard libpq PGPASSWORD environment variable, which the driver merges into the connection. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). FerretDB's debug listener, port 8088, checks no caller and serves /debug/pprof, /debug/vars, /debug/archive and /debug/graphs beside its metrics at /debug/metrics. While `values.metrics_enabled` is false the Recipe binds it to the pod's loopback address, which no other pod reaches, and provides.metrics_endpoint is empty. True, it binds it on every interface of the pod, its port joins the ClusterIP Service, and provides.metrics_endpoint publishes its /debug/metrics there, so every in-cluster caller that reaches the pod reaches the whole listener. Neither probe uses it: readiness runs `ferretdb ping` inside the container, and liveness checks that the wire port accepts connections.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/ferretdb:1.0.0 - v1.0.0
Recipe
event-publisher-knative
The event-publishing tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The broker the tool publishes to arrives on the tool's messaging socket, and the recipe delivers that broker's host, port and protocol into the container alongside the developer's topic, so the running tool addresses the broker the topology resolved rather than one named in its own configuration. A connection writing an `inject:` block receives instead the names it dictates, as variables or as files from a ConfigMap emitted beside the Service. A server_ca on the connection is mounted read-only and named as EVENT_TOOL_BROKER_SERVER_CA_FILE, and is refused on a broker served without TLS. provides is the addressable name the control plane gives the Service. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/event-publisher-knative:1.0.0 - v1.0.0
Recipe
external-secrets
external-secrets operator on Kubernetes via the official external-secrets Helm chart (external-secrets/external-secrets). The chart deploys the controller that reconciles ExternalSecret resources and synchronizes secret material from external secret-management systems into native Kubernetes Secrets, a webhook that validates and defaults external-secrets' custom resources, and a certificate controller that manages the webhook's serving certificate, together with their RBAC, the webhook Service, and the admission-webhook configurations. When the Component's install_crds value is on, the chart also installs the external-secrets CustomResourceDefinitions. The three container images share one repository; the chart defaults each tag to its appVersion, which this Recipe re-pins explicitly for reproducible renders. The bundled provider subchart is off, so a single coherent operator renders with no foreign datastore, cache, or broker. Default chart version in source.version; override via infra.source_version. external-secrets is distributed under the Apache-2.0 license. provides publishes the pod label, port name and path a scraper discovers the three processes' metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/external-secrets:1.0.0 - v1.0.0
Recipe
etcd
Single-node etcd on Kubernetes via the helm provider's Workload synthesis (a StatefulSet with a persistent volume, plus a ClusterIP Service and a headless Service), running the official upstream image (no external Helm chart). etcd holds its data on the persistent volume and serves the v3 client API over HTTP without authentication. Its own metrics, in the Prometheus format at /metrics on the client port, are published as provides.metrics_endpoint. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/etcd:1.0.0 - v1.0.0
Recipe
envoy-gateway
An Envoy Gateway ingress gateway via the project's own Helm chart, from a container registry: the controller and its access control, the certificate-generation job and — unless infra.install_crds is false, when infra.control_plane_present must be true or the build halts — the routing API's experimental-channel definitions. Beside the chart this Recipe emits the gateway class, claimed under the controller name the control plane answers for, and the gateway with one HTTP listener. Where infra.tls_secret names a certificate Secret, the gateway gains an `https` listener on port 443 terminating TLS with it; the Secret is referenced and never emitted. It emits neither the data plane, which the control plane creates, nor a namespace. Wired to the AI gateway controller on the extension socket, the control plane names it as its extension server, and the gateway must sit in envoy-gateway-system. Each model on the models socket gets a rule of one AI route and the resources reaching its published endpoint, its credential named only by its Secret; the project's client traffic policy raises the gateway's buffer to 50Mi. With a model wired, provides.host is the data-plane Service Envoy Gateway creates and provides.endpoint its /v1 base URL, qualified with envoy-gateway-system. The route's one traffic policy holds core/period-budget as a rate limit on the route's token cost, counted in the cache socket's store, over TLS to one publishing valkey-tls or redis-tls, core/retry-policy as a retry count, and core/concurrency-limit as a per-model cap on requests in flight. core/audit-logging is the gateway's access log, adding the model and token counts.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/envoy-gateway:1.0.0 - v1.0.0
Recipe
email-knative
The mailbox tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The developer's chosen action reaches the running tool as an environment variable, so what the agent may do with the mailbox is fixed at authoring time rather than at call time. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, so a consumer wires to the tool's declared contract rather than to anything Knative-specific. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/email-knative:1.0.0 - v1.0.0
Recipe
envoy-ai-gateway-operator
Installs the Envoy AI Gateway control plane through the project's own Helm chart, referenced as an artifact in a container registry at the release the controller image comes from. The chart's render carries the controller Deployment and its arguments, the Service Envoy Gateway, the cluster's API server and a metrics scraper reach it on, its service account, the cluster role and binding granting what the project's chart grants, a cluster role letting Envoy Gateway read inference pools, and the mutating webhook configuration that adds the AI external processor to each Envoy Gateway data-plane pod. That Service's /metrics address on 8080 is provides.metrics_endpoint. Beside the chart it emits the six AI Gateway custom-resource definitions, carried beside this Recipe exactly as the project's definitions chart publishes them at the same release. The webhook's serving certificate comes from cert-manager by default: the chart's Certificate and self-signed Issuer are rendered, so cert-manager must already run in the cluster. With infra.webhook_tls_secret set it is the certificate in the Secret that key names, which the operator creates beforehand: no Certificate or Issuer is rendered, cert-manager is not needed, and the Secret the chart would otherwise generate is left out. Every namespaced resource lands in the namespace the Release declares for the component, or in `default` where none is declared; the controller expects the Envoy Gateway it extends in `envoy-gateway-system`, the chart's own default. The controller's MCP session-encryption seed is read when it starts from the Secret infra.mcp_session_seed_secret names, and is otherwise the chart's published default.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/envoy-ai-gateway-operator:1.0.0 - v1.0.0
Recipe
dynamiq
A Dynamiq agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied image, which carries both the framework and the workflow it executes. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_URL — the variable Dynamiq's own OpenAI connection reads, which is not the one the bare client SDK would consult. The datastore wired into the `state_store` socket contributes its host and port. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_URL and the API key stay. A state store served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as DYNAMIQ_WORKFLOW_ENTRYPOINT and DYNAMIQ_STREAM_EVENTS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/dynamiq:1.0.0 - v1.0.0
Recipe
elasticsearch
Elasticsearch on Kubernetes via the helm provider's Workload synthesis: a StatefulSet running Elastic's own image as one node, with its indexes on a persistent volume, a ClusterIP Service publishing the REST API, and the headless Service the stateful form requires. Elastic's own Helm chart for Elasticsearch has had no release since version 8.5.1 and its repository is archived, so the server is synthesized rather than sourced. Authentication stays on with the built-in superuser, elastic, whose password is read from the attached credential's Secret by name, so it never appears in the rendered files; with none attached it is a development default. A credential that names no Secret halts the build. The image's first-start setup, which generates certificates and prints a password, is switched off. The REST API is served over plain HTTP, not encrypted, and is reachable only inside the Release; the transport layer listens inside the pod only. Reach for this realization on Kubernetes: it needs no control plane and no chart repository, and it builds with nothing fetched. The Compose Recipe of the same name is the local-development realization.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/elasticsearch:1.0.0 - v1.0.0
Recipe
documentdb
PostgreSQL 17 with the DocumentDB extension on Kubernetes via the helm provider's Workload synthesis (StatefulSet + ClusterIP Service + headless Service + PVC). provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/documentdb:1.0.0 - v1.0.0
Recipe
docker-model-runner
Docker Model Runner as a self-hosted inference server on Kubernetes, synthesized as a StatefulSet running the runner's own image. It serves the OCI model reference the Aggregate names, the same reference the Docker Compose realization hands the host's runner, so one Aggregate asks both targets for the same model. A pull step running the same image puts the model into a per-replica model cache before the server starts. The server applies the context size and runtime flags each time it starts, and serves the OpenAI-compatible API under /engines/v1 behind a ClusterIP Service. The compiler never fetches the model. This realization is synthesized because the project's own Helm chart fails the maintained and complete tests: it lags the runner's releases, and it exposes no setting for the context size, runtime flags, compute resources or listen port. It is also published to no chart repository. provides.host is the bare in-cluster name of the Service, with no namespace, so a consumer in another namespace qualifies it itself. infra.endpoint, the host runner's address on Docker Compose, is refused here.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/docker-model-runner:1.0.0 - v1.0.0
Recipe
dify-ssrf-proxy
The Dify egress proxy on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the proxy port). The proxy caches in memory only and keeps nothing across a restart, so it renders stateless. The whole Squid configuration is emitted as a ConfigMap via artifacts: and mounted whole-directory, and the container is started against that file directly. Every value the configuration needs is resolved when the stack is compiled, so the rendered file is the final configuration rather than a template a container entrypoint has to expand at startup — which is what lets a reviewer read the effective access rules out of the manifest. The access rules are ordered so that the refusal comes before the permissions: any destination in private, loopback, link-local, shared-CGN or multicast address space is refused, and only then are the named domains and the calling workloads permitted. That ordering is the security property — a supplied address reaches the public internet and cannot reach the cluster's own services or the node's metadata endpoint — so a private range becomes reachable only by being named in allowed_private_networks, which renders ahead of the refusal. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing).
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/dify-ssrf-proxy:1.0.0 - v1.0.0
Recipe
dify-sandbox
The Dify code sandbox on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment and a ClusterIP Service on the execution port. Each execution is independent and its working directory is discarded, so the sandbox holds nothing between calls and renders stateless. Code the sandbox runs was supplied by an application author, so its egress leaves through the proxy wired on the ssrf_proxy edge, which refuses private destinations. The execution key comes from a credential when one is attached and otherwise from a pinned development literal, for evaluation and local development: while it is in force, any workload that can reach this Service can execute code through it. Python requirements the developer declared are mounted from a ConfigMap. provides.host is the bare in-cluster Service name, with no namespace; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/dify-sandbox:1.0.0 - v1.0.0
Recipe
dify-plugin-daemon
The Dify plugin runtime on Kubernetes via the helm provider's Workload synthesis. Installed plugin packages and the Python environments prepared for them live on the runtime's own volume, so the workload renders as a StatefulSet with a persistent claim, alongside the governing headless Service the stateful form requires and a ClusterIP Service publishing the dispatch port. One replica is rendered: a second would hold a different plugin set and answer differently depending on which one a call reached. It keeps its plugin catalogue in a database of its own on the backend's server, named by plugin_database, and calls the API server back on its internal API. The backend password and the two keys it shares with the API server and the worker each come from a credential when one is attached and otherwise from a pinned development literal, for evaluation and local development; the two it presents, to the backend and to the API server, are held on those connections, and the two keys are attached on all three or on none. A cache publishing valkey-tls or redis-tls is reached over TLS, its certificate checked against the cache connection's server_ca, mounted, or the system roots; another TLS protocol on the cache halts the build. The runtime's own metrics, in the Prometheus format at /metrics/ on the dispatch port, asking no key, are published as provides.metrics_endpoint. provides.host is the bare in-cluster Service name, with no namespace; a consumer in another namespace qualifies it on its own side.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/dify-plugin-daemon:1.0.0 - v1.0.0
Recipe
datastore-query-knative
The datastore-query tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. The store the tool queries arrives on the tool's datastore socket, and the recipe delivers that store's host and port — and the user and database name where the store publishes them — into the container alongside the developer's dialect, so the running tool addresses the store the topology resolved rather than one named in its own configuration; on such a connection a store served over TLS alone halts the build, naming it. A connection writing an `inject:` block receives instead the names it dictates, as variables or as files from a ConfigMap emitted beside the Service, and is not refused for TLS. provides is the addressable name the control plane gives the Service. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/datastore-query-knative:1.0.0 - v1.0.0
Recipe
databricks
A Databricks model serving endpoint as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/databricks:1.0.0 - v1.0.0
Recipe
computer-use
The virtual-desktop tool realized on Kubernetes by the helm provider's Workload synthesis, rendered offline as plain manifests. The Workload becomes a Deployment that keeps the desktop session resident and a ClusterIP Service that fronts the port its control API listens on, because an agent screenshots and drives the same desktop across turns rather than starting a new one per call. It runs a single replica for that reason: two replicas behind one Service are two desktops, and a screenshot taken from the second would not show what the pointer input did to the first. The developer's screen geometry, declared effect and timeout reach the running session as environment variables. The image is the Release's, because this tool packages no upstream product: a build with none set halts and names the key. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Deployment. provides is the synthesized Service, so a consumer wires to a host, port, endpoint and protocol and never to the desktop behind them. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/computer-use:1.0.0 - v1.0.0
Recipe
contour
Contour ingress gateway via the official project-contour Helm chart, served from the project's HTTPS chart repository. The chart deploys the Contour control plane, which reconciles the cluster's Ingress and HTTPProxy resources, alongside an Envoy data plane that serves client traffic. The chart manages the Contour CRDs and creates the IngressClass named from the Component's ingress_class_name value, so applications routing to that class are programmed into Envoy. The Envoy service is exposed as a LoadBalancer by default; set infra.overrides.service_type to change it. This Recipe pins the Contour and Envoy container images to explicit release tags for reproducible renders. Default chart version in source.version; override via infra.source_version. provides also publishes the pod label, port name and path a scraper discovers the Envoy pods' metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/contour:1.0.0 - v1.0.0
Recipe
cohere
Cohere's managed Command API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/cohere:1.0.0 - v1.0.0
Recipe
cognee
A semantic-memory store on Kubernetes via the helm provider's Workload synthesis, running the official upstream image directly (no external Helm chart). The graph, the embeddings and the bookkeeping tables all live in files under one persistent volume, so the workload renders as a StatefulSet with that volume mounted at the state root, alongside the governing headless Service the stateful form requires and a ClusterIP Service publishing the API port. This realization is where the component's engine value is earned: the container's configuration names the graph engine the entities and relations are written into, the store the embeddings are written into, the models the ingested material is read and encoded with, and the ontology the extracted names are resolved against — so the rendered artifact shows how memory is built rather than leaving it to the running process to reveal. A store that only indexed vectors would carry none of those settings. A credential attached on a model connection reaches the server as an env secretKeyRef — LLM_API_KEY for the extraction model, EMBEDDING_API_KEY for the embedding model — so its value never appears in the rendered manifest; the component's model_api_key is read only for a socket left unwired, under the same names. With nothing attached, the server holds no model credential. provides.host is the bare, self-contained in-cluster service name of the synthesized ClusterIP Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing). The configuration contract is authored against cognee 1.5.3.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/cognee:1.0.0 - v1.0.0
Recipe
code-interpreter-knative
The code-executing tool realized as a Knative Service the Knative Serving control plane reconciles. The recipe emits a single Service custom resource; the control plane turns it into a revision, admits an inbound call while that revision runs no replicas, and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. The replica floor and ceiling carry through as the revision's own scaling bounds. Each revision serves whichever runtime the image carries, and the runtime the developer selected is delivered into the container so the tool can refuse work it cannot evaluate. Nothing is written back, and nothing is mounted beyond the read-only files a connection dictates, which is what keeps one execution from reaching the next. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Service. provides is the addressable name the control plane gives the Service, so a consumer wires to the tool's declared contract rather than to anything Knative-specific. The Knative Serving control plane is a runtime prerequisite: the build halts unless the operator affirms it is installed (infra.control_plane_present), so no custom resource is emitted into a cluster that cannot reconcile it. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/code-interpreter-knative:1.0.0 - v1.0.0
Recipe
chroma
Chroma on Kubernetes via the helm provider's Workload synthesis: a StatefulSet running the official chromadb/chroma image with a persistent volume at the server's data path, a ClusterIP Service on the HTTP port, and the headless Service the stateful form requires. The project's one Helm chart is for its distributed version, and its maintainers mark it as not ready for use, so the single server is synthesized rather than sourced. The server checks no credential, so the Recipe publishes it on a ClusterIP Service and nothing else: it is reachable from inside the cluster, and a Release that exposes it further puts something that authenticates clients in front of it. provides.host is the bare in-cluster service name, with no namespace; a consumer in another namespace qualifies it on its own side. Reach for this realization on Kubernetes: it needs no control plane and no chart repository, and it builds with nothing fetched. The Compose Recipe of the same name is the local-development realization.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/chroma:1.0.0 - v1.0.0
Recipe
claude-agent-sdk
An application built on the Claude Agent SDK, on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the image the Release supplies. The model wired into the `model` socket reaches the command line each session runs in as the variables it reads: its endpoint as ANTHROPIC_BASE_URL, its model name as ANTHROPIC_MODEL, and the connection's credential as ANTHROPIC_API_KEY through a secretKeyRef, so the key never renders into the manifest. Each tool wired into the `tools` socket reaches the application under variables named for the connection's alias, upper-cased with hyphens as underscores: its address as AGGREGA_TOOL_<ALIAS>_URL, and, where the connection carries a credential, that credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL through a secretKeyRef, with the header the tool publishes for it as AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection holding the tool's whole address as a secret, for a server that reads its key from its address, delivers that secret as AGGREGA_TOOL_<ALIAS>_URL through a secretKeyRef instead. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names: a tools connection in place of its address and header variables, while what its attachments deliver stays, and the model connection beside the variables the command line reads. The component's own `inject:` block delivers its entries as a connection's are.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/claude-agent-sdk:1.0.0 - v1.0.0
Recipe
cloudflare-ai-gateway
References a Cloudflare AI Gateway that already exists, as a reference realization: nothing runs in the cluster and nothing in the account is created or changed. helm emits an ExternalName Service mapping an in-cluster name onto the gateway's host. Reach for it where the gateway is created and configured outside the Release; reach for the terraform Recipe where the Release is to create it. It carries no Feature, because every control the gateway keeps is a setting of the gateway itself, which this realization does not write, so the build halts on any attachment. No credential is delivered: a consumer that needs one declares it as `credential` on the socket that reaches the gateway and delivers it from its own container.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/cloudflare-ai-gateway:1.0.0 - v1.0.0
Recipe
cert-manager
cert-manager TLS-certificate lifecycle controller via the official cert-manager Helm chart, sourced from the jetstack OCI registry. The chart deploys the controller that reconciles Certificate and Issuer resources, a webhook that validates and defaults cert-manager's custom resources, and a CA injector that wires the webhook's CA bundle into the cluster, together with their RBAC, the webhook Service, the admission-webhook configurations, and a post-install startup check that confirms the certificate API is serving. When the Component's install_crds value is on, the chart also installs the cert-manager CustomResourceDefinitions. This Recipe pins the controller, webhook, CA-injector, startup-check, and ACME-solver container images to explicit release tags for reproducible renders. Default chart version in source.version; override via infra.source_version. Every namespaced resource lands in the namespace the Release declares for the component, or in `default` where none is declared, except two Roles and their RoleBindings: they let the controller and the CA injector hold their leader-election lease, which the chart keeps in `kube-system`, so they land there. provides publishes the pod label, port name and path a scraper discovers the three processes' metrics by.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/cert-manager:1.0.0 - v1.0.0
Recipe
browser
The browser-driving tool realized on Kubernetes by the helm provider's Workload synthesis, rendered offline as plain manifests. The Workload becomes a Deployment that stays up between calls and a ClusterIP Service that fronts the port it listens on, because an agent reaches this tool by connecting to it rather than by triggering it. It runs a single replica: a browsing session's cookies and open pages live in the process serving them, so a second replica behind one Service would scatter one session's steps across two browsers. The developer's automation engine, browser build, headless choice, declared effect and timeout reach the running tool as environment variables. The image is the Release's, because this tool packages no upstream product: a build with none set halts and names the key. An edge from the tool, on a socket the author names, delivers the names its `inject:` block dictates, as variables or as files from a ConfigMap emitted beside the Deployment. provides is the synthesized Service, so a consumer wires to a host, port, endpoint and protocol and never to the browser behind them. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value of the tool alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/browser:1.0.0 - v1.0.0
Recipe
bedrock
Amazon Bedrock's managed model API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied regional endpoint. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/bedrock:1.0.0 - v1.0.0
Recipe
bland
The Bland voice-agent platform as a reference realization on Kubernetes: no compute runs in-cluster and nothing is provisioned. The helm provider emits an ExternalName Service mapping an in-cluster name onto the platform's API base address, so a consumer may reach it under a stable in-cluster name whichever address the operator pointed it at. This recipe pins no artifact. It names no image, no chart and no module — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here. A credential attached to either of the component's slots is materialized by that credential's own realization, and the consuming application references the material under the name it was given, because the environment-variable name a credential arrives under belongs to the consumer rather than to the platform. The api_key slot is required here: a build leaving it empty halts naming it. The byo_telephony_key slot stays optional.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/bland:1.0.0 - v1.0.0
Recipe
bedrock-agent
An Amazon Bedrock agent as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the operator-supplied regional agent-runtime endpoint. No credential is delivered by this recipe — consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials. provides.endpoint is the regional host and provides.agent_name is the developer-selected agent.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/bedrock-agent:1.0.0 - v1.0.0
Recipe
bedrock-embedding
Amazon Bedrock's managed embedding API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied regional endpoint. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/bedrock-embedding:1.0.0 - v1.0.0
Recipe
barman-cloud-plugin-chart
Installs the Barman Cloud plugin for CloudNativePG through the chart the CloudNativePG project publishes, at the chart release whose application version is the image the Component pins. The chart is official, from CloudNativePG's own chart repository; complete, carrying the plugin's custom-resource definition, its Deployment, the Service the operator finds it by, its permissions and its certificates; and maintained, its newest release carrying the plugin's newest, 0.15.1. The certificates are cert-manager Certificates from an Issuer the chart renders, so cert-manager must already run in the cluster. It renders no object store and no backup: helm/backup-cloudnative-pg writes those for each core/backup attachment. The build halts unless the Release, or else the Aggregate, declares a namespace for the component, which must be the CloudNativePG operator's. Its chart repository must be reachable from the build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/barman-cloud-plugin-chart:1.0.0 - v1.0.0
Recipe
azure-openai-embedding
Azure OpenAI's managed embedding API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer. Reach for it where the Azure OpenAI resource already exists outside the Release; reach for terraform/azure-openai-embedding where the Release is to create the resource and its model deployment.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/azure-openai-embedding:1.0.0 - v1.0.0
Recipe
azure-openai
Azure OpenAI's managed chat API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer. Reach for it where the Azure OpenAI resource already exists outside the Release; reach for terraform/azure-openai where the Release is to create the resource and its model deployment.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/azure-openai:1.0.0 - v1.0.0
Recipe
azure-api-management
References an API already served by an Azure API Management instance, as a reference realization: nothing runs in the cluster and nothing in Azure is created or changed. helm emits an ExternalName Service mapping an in-cluster name onto the gateway's host. Reach for it where the API is configured outside the Release; reach for the terraform Recipe where the instance exists and the Release is to configure the API. It carries no Feature, because every control the gateway keeps is a policy in the API's own configuration, which this realization does not write, so the build halts on any attachment. No credential is delivered: a consumer that needs one declares it as `credential` on the socket that reaches the gateway and delivers it from its own container.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/azure-api-management:1.0.0 - v1.0.0
Recipe
azure-managed
An Azure managed online endpoint as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the operator-supplied endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/azure-managed:1.0.0 - v1.0.0
Recipe
autogpt
An AutoGPT agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as AUTOGPT_AI_NAME, AUTOGPT_AI_ROLE, AUTOGPT_AI_GOALS (as JSON text), AUTOGPT_CONTINUOUS_MODE and AUTOGPT_CONTINUOUS_LIMIT (while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/autogpt:1.0.0 - v1.0.0
Recipe
autogen
An AutoGen agent application on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment running the operator-supplied application image. The model endpoint published by the component wired into the `model` socket is injected as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads, so the application talks to the wired model without hardcoding an endpoint. A model API key attached on the `model` connection is delivered as OPENAI_API_KEY via a secretKeyRef. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as AUTOGEN_AGENT_NAME and AUTOGEN_SYSTEM_MESSAGE by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/autogen:1.0.0 - v1.0.0
Recipe
argo-workflows
Argo Workflows via the project's own Helm chart, served from the argoproj HTTPS chart repository. The chart deploys the workflow controller, which watches Workflow custom resources and starts a pod per step, together with the API server that serves the submission and query API and the web console, the role and role binding for a workflow pod's service account in each namespace the controller manages, and the controller's ConfigMap. When the Component's install_crds value is on, the chart also installs the Argo Workflows CustomResourceDefinitions. This Recipe pins the controller, API server, workflow executor and definition-installer container images to explicit release tags for reproducible renders, and pins the chart's resource names to the release name the helm provider gives the chart, so the API server's Service is named predictably, in the alphabet a Kubernetes name allows. Default chart version in source.version; override via infra.source_version. Every namespaced resource lands in the namespace the Release declares for the component, or in `default` where none is declared. The workflow role and role binding are also rendered in each namespace workflow_namespaces names, `default` unless the Aggregate sets it, because workflows run there. With `values.metrics_enabled` true the chart serves the workflow controller's metrics over plain HTTP and adds its ClusterIP Service, `<rendered_name>-workflow-controller`, port 8080, which checks no caller; provides.metrics_endpoint publishes the address. False, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/argo-workflows:1.0.0 - v1.0.0
Recipe
apisix
Apache APISIX on Kubernetes via the helm provider's Workload synthesis (a stateless Deployment plus a ClusterIP Service on the proxy port), running the official upstream image with no Helm chart. APISIX holds no local state — the etcd store is the system of record — so it renders stateless. The Admin API is reachable only from inside the cluster — no Service publishes its port — and is authenticated by the key the `admin_key` credential slot names, which the Deployment delivers from the Secret the attached credential publishes, so no rendered file carries the key. Its client allowlist, `allow_admin`, admits every address, so narrow it wherever the admin port is routable from outside the cluster. It renders no route: routes live in etcd, written through the Admin API at run time, so it serves no model, publishes an empty endpoint and model, and halts on a connection to the models socket — `apisix-chart` serves those. At its default image, APISIX 3.17.0, the gateway has no semantic cache plugin (`ai-cache`), and its token rate limiting counts per node; it halts on a core/period-budget, core/semantic-caching or core/prompt-deny-patterns attachment, and on infra.callers, which only a routes file declares. With `values.metrics_enabled` true, the gateway's metrics server moves off the loopback address and its port, 9091, which checks no caller, joins the Service, its address published as provides.metrics_endpoint; false, that key is empty. Reach for this realization where the build must complete with nothing fetched; reach for `apisix-chart`, which renders the project's own chart from named settings, where the configuration is expected to follow the product or a model is served.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/apisix:1.0.0 - v1.0.0
Recipe
apisix-chart
Apache APISIX on Kubernetes via the project's own Helm chart, its bundled etcd and ingress controller switched off: a Deployment, a ClusterIP Service for the proxy data plane and the configuration map the gateway reads. With nothing wired to the models socket the configuration store is the etcd component wired on the `etcd` socket, and a ClusterIP Service fronts the Admin API, authenticated by the key the `admin_key` slot names: the chart reads it from the Secret the attached credential publishes, so no rendered file carries a key. Its client allowlist, `apisix.admin.allow.ipList`, admits every address, so narrow it wherever the admin port is routable from outside the cluster. With a model wired the chart runs in its standalone mode: each connection's routes, through APISIX's AI proxy, sit in a ConfigMap named after the gateway, which reads them with no etcd and no Admin API. A connection's credential, and each key of the callers infra.callers declares, arrives as an env secretKeyRef the routes file names through APISIX's $env:// reference, so no key is in the rendered output; each caller is a consumer whose key every route checks. Written onto the routes: core/period-budget as token rate limiting, with one replica; core/semantic-caching as a semantic cache; core/prompt-deny-patterns as a prompt guard — no completion route beside the budget or the guard. With `values.metrics_enabled` true, the chart serves the gateway's metrics on its own Service, port 9091, which checks no caller, published as provides.metrics_endpoint; false, that key is empty. Reach for it where a chart repository is reachable at build time or a model is served, and for `apisix` where the build must fetch nothing.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/apisix-chart:1.0.0 - v1.0.0
Recipe
anthropic
Anthropic's managed Claude API as a reference realization: no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in- cluster DNS name onto the managed endpoint. provides.endpoint is the managed API host and provides.model is the developer-selected model string. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/anthropic:1.0.0 - v1.0.0
Recipe
anthropic-agent
A Claude assistant on Anthropic's managed API as a reference realization, so no compute runs in-cluster. The helm provider emits an ExternalName Service that maps an in-cluster DNS name onto the managed endpoint. provides.endpoint is the managed API host and provides.model is the developer-selected model string. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/anthropic-agent:1.0.0 - v1.0.0
Recipe
ai21
AI21 Labs' managed Jamba API as a reference realization: no compute runs in-cluster. helm emits an ExternalName Service mapping an in-cluster name onto the managed endpoint. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/ai21:1.0.0 - v1.0.0
Recipe
agentgateway-chart
agentgateway on Kubernetes via the project's own standalone Helm chart, fetched from the project's container registry: a Deployment, a ClusterIP Service for the data plane, a service account and the ConfigMap holding the configuration file the gateway reads, which this Recipe writes through the chart's `config` value. Each model wired to the models socket is one entry of that file, named by the connection's alias and sent to the endpoint and model name the model publishes. A connection's credential, and each key of the callers infra.callers declares, arrives as an environment variable by secretKeyRef, and the file names the variable, so no key is in the rendered output; each caller is an API key checked in strict mode. Written into the file: core/data-masking and core/prompt-deny-patterns as agentgateway's regular-expression guards, core/period-budget as a token budget, core/audit-logging as a record of each model request past its key, budget and authorization checks, and core/retry-policy as a retry policy on the listener model requests arrive on. The per-caller budget and the records are kept in a SQLite file on a volume this Recipe renders while either needs it, with one replica. Reach for it on Kubernetes, where the build can reach the chart's registry; reach for `agentgateway` under Docker Compose. The image is pinned by digest, and infra.source_version accepts only the chart versions this configuration is proven at. The gateway's metrics are always on, on port 15020, which checks no caller. With `values.metrics_enabled` true a ClusterIP Service of its own, `<rendered_name>-metrics`, carries that port and provides.metrics_endpoint publishes its address; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agentgateway-chart:1.0.0 - v1.0.0
Recipe
agenta-worker-streams
The agent workspace's stream-consuming worker on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment and no Service. There is no Service because nothing connects to this tier — work reaches it by being written to the streams it reads — and emitting one would advertise an endpoint that answers nothing. Each backing system's identity arrives on .context.inject.env and is composed here into the single connection string its client reads, under the same variable names the control-plane API's Recipe uses, because the two run the same image and read the same configuration. The relational connections are rendered password-free: the password arrives beside them as an env secret reference and is named inside each URL in the platform's own runtime-substitution form. A cache or queue publishing valkey-tls or redis-tls is reached over TLS, its certificate checked against that connection's server_ca, mounted, or the image's trust store; another TLS protocol halts the build. The stream selector is emitted from the component's own value, empty by default, which is what makes one container run every family. This Recipe publishes no provides contract, because the component declares none.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-worker-streams:1.0.0 - v1.0.0
Recipe
agenta-worker-queues
The agent workspace's queue-consuming worker on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment and no Service. There is no Service because nothing connects to this tier — work reaches it by being enqueued — and emitting one would advertise an endpoint that answers nothing. Each backing system's identity arrives on .context.inject.env and is composed here into the single connection string its client reads, under the same variable names the control-plane API's Recipe uses, because the two run the same image and read the same configuration. The model-facing service's in-cluster base URL, resolved from the `services` edge, is what a job that runs an agent turn or scores an evaluation calls. The relational connections are rendered password-free: the password arrives beside them as an env secret reference and is named inside each URL in the platform's own runtime-substitution form. A cache or queue publishing valkey-tls or redis-tls is reached over TLS, its certificate checked against that connection's server_ca, mounted, or the image's trust store; another TLS protocol halts the build. The queue selector is emitted from the component's own value, empty by default, which is what makes one container run every family. This Recipe publishes no provides contract, because the component declares none.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-worker-queues:1.0.0 - v1.0.0
Recipe
agenta-web
The agent workspace's browser console on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port the ingress tier routes the browser path to. The console holds no state and owns no datastore, so it renders stateless. Two kinds of address are emitted, and keeping them apart is what makes the console work. The in-cluster base URLs of the control-plane API and the model-facing service, resolved from the `api` and `services` edges, are what the server-side rendering path calls: a page assembled inside the cluster reaches those services by their Service names. The externally reachable base URL the operator supplied is what the browser calls, since a Service name does not resolve outside the cluster, and the two path suffixes on it match the routes the ingress tier assembles. The authentication core's own base URL, resolved from the `auth_core` edge, is emitted for the sign-in flow. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-web:1.0.0 - v1.0.0
Recipe
agenta-web-mobile
The agent workspace's small-screen front end on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port the ingress tier routes its own path to. The front end holds no state and owns no datastore, so it renders stateless. It makes the same split between in-cluster and browser-facing addresses the full console's Recipe makes, for the same reason: the base URLs of the control-plane API and the model-facing service resolved from the `api` and `services` edges are what the server-side rendering path calls, while the externally reachable base URL the operator supplied is what the browser calls. The authentication core's own base URL, resolved from the `auth_core` edge, is emitted for the sign-in flow. One variable differs from the full console's rendering and is worth naming: this front end declares itself the small-screen destination, so a visitor it serves is not redirected onward. That is what keeps the pair from bouncing a visitor between them. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-web-mobile:1.0.0 - v1.0.0
Recipe
agenta-services
The agent workspace's model-facing service on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port the console's playground and the queue-consuming worker call. The service keeps no durable state of its own, so it renders stateless, with no persistent volume and no headless Service. Each backing system's identity arrives on .context.inject.env and is composed here into the single connection string its client reads, exactly as the control-plane API's Recipe does — the two services read one model, so they are given the same connections under the same variable names. The relational connections are rendered password-free: the password arrives beside them as an env secret reference and is named inside each URL in the platform's own runtime-substitution form, so the plaintext never appears in the rendered manifest. The object store's secret key has no default, so a store connection carrying no credential halts the build, naming it, unless its values set unauthenticated: true, and then no secret key is sent. A cache or queue publishing valkey-tls or redis-tls gets a rediss URL, naming that connection's server_ca, mounted, where one is held; another TLS protocol halts the build. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-services:1.0.0 - v1.0.0
Recipe
agenta-runner
The agent runner on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port the model-facing service calls. The local sandbox provider is the only one enabled — an agent's harness runs as a child process in this container — so the deployment needs nothing else present for a run to start. What this Recipe deliberately does not render is the point of it. A locally sandboxed harness shares this container, so every variable here is readable by an agent's own code; the Recipe therefore emits the runner's own configuration, the platform's in-cluster address, and the runner's shared token, and nothing else. There is no database URL, no cache or queue address, no object-store key and no encryption key in the rendered container, because the component declares no socket that would carry one. The single replica is deliberate rather than a default left in place: a local sandbox belongs to the container that created it, so a second replica would answer for sessions it cannot reach. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-runner:1.0.0 - v1.0.0
Recipe
agenta-migrations
The agent workspace's schema-migration step on Kubernetes via the helm provider's Workload synthesis, rendered in the run-to-completion form: the workload runs once, applies the migration history to both of the platform's databases, and exits. It is the only tier of this platform that is not a Deployment, and the form is what makes that true rather than a comment claiming it — a long-running controller shape would restart the run forever. No Service is emitted, and no port: nothing connects to a run that exits. Nor is a deadline or a retry count emitted: the workload contract this Recipe authors against expresses neither for a run-to-completion form, so the run takes whatever the target platform's own defaults are. An installation that needs a bounded window for the long case — a first run against an empty database, which applies the whole history — sets it on the emitted resource at the target. The backend identity resolved from the `backend` uses-edge arrives on .context.inject.env and is composed here into the two connection strings the migration runner reads, one per database named on that edge. The connections are rendered password-free: the password arrives beside them as an env secret reference and is named inside each URL in the platform's own runtime-substitution form. This Recipe publishes no provides contract, because the component declares none.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-migrations:1.0.0 - v1.0.0
Recipe
agenta-cron
The agent workspace's scheduler on Kubernetes via the helm provider's Workload synthesis: a single-replica stateless Deployment and no Service. The controller shape is a deliberate choice rather than a default left in place. The scheduler is a resident process carrying its own schedule table, which is how the platform's own published manifests run it, so this Recipe renders it as a Deployment rather than as a scheduled workload. A scheduled form would hand the timing decision to the platform underneath and fire the container on a fixed expression, which is not what this tier does — it wakes on its own clock and reads schedules a user authored in the platform. One replica, and that too is deliberate: two schedulers reading one schedule table would each enqueue the same due item. There is no Service because nothing connects to this tier. Each backing system's identity arrives on .context.inject.env and is composed here into the single connection string its client reads, under the same variable names the control-plane API's Recipe uses. The relational connections are rendered password-free: the password arrives beside them as an env secret reference and is named inside each URL in the platform's own runtime-substitution form. A cache or queue publishing valkey-tls or redis-tls gets a rediss URL, naming that connection's server_ca, mounted, where one is held; another TLS protocol halts the build. This Recipe publishes no provides contract, because the component declares none.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-cron:1.0.0 - v1.0.0
Recipe
agent-invocation-knative
The agent-delegation tool as one Knative Service custom resource, which Knative Serving runs as a revision scaled on observed concurrency, to zero replicas between calls yet admitting each. The container receives the address keys the wired agent publishes and the developer's protocol and execution mode, so the tool reaches the agent the topology resolved, not one it names. A connection writing an `inject:` block receives instead the names it dictates, as variables or as files from a ConfigMap emitted beside the Service, beside what its attachments deliver. The component's own `inject:` block adds `env` and `file` entries beside them, and one reading a tool value alone replaces that value's default name. The build halts unless infra.control_plane_present affirms Knative Serving is installed, so no Service reaches a cluster unable to reconcile it. A credential or client_secret on the connection reaches the container by reference to its Secret: a credential beside the header the agent publishes, the Secret holding the whole header value; a client_secret beside the client id, scopes and token or discovery URL the container obtains a token with by the OAuth 2.0 client-credentials grant, directly or through OpenID Connect. Either halts the build on an agent labelled external publishing protocol http; to one inside the Release, http carries the credential or token unencrypted on the cluster network. A client_certificate, for mutual TLS, and a server_ca are each mounted read-only from their Secret, variables naming the files. Each Secret is one already in the namespace, named through helm/k8s-existing-secret; a realization naming it after the tool halts the build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agent-invocation-knative:1.0.0 - v1.0.0
Recipe
agenta-api
The agent workspace's control-plane API on Kubernetes via the helm provider's Workload synthesis: a stateless Deployment plus a ClusterIP Service on the HTTP port other workloads of the platform call. The service keeps no durable state of its own — the relational backend, the cache, the queue and the object store are the systems of record — so it renders stateless, with no persistent volume and no headless Service. Each backing system's identity arrives on .context.inject.env and is composed here into the single connection string its client reads. The two relational connections are rendered password-free: the password is delivered beside them as an env secret reference and named inside each URL in the platform's own runtime-substitution form, so the plaintext never appears in the rendered manifest even though the client receives one URL. With no credential attached the Recipe uses inline development defaults, and with one attached the value is supplied through the credential subsystem. A cache or queue publishing valkey-tls or redis-tls is reached over TLS, trusting that connection's server_ca or the image's store. The object store's secret key is the exception: it has no default, so a store connection carrying no credential halts the build, naming it, unless its values set unauthenticated: true, and then no secret key is sent. provides.host is the bare, self-contained in-cluster service name of the synthesized Service — it carries no namespace, so the recipe is independent of where it is deployed. A consumer in another namespace qualifies the host with the target namespace on its own side (caller-side routing), and assembles its own URL rather than reusing provides.endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/agenta-api:1.0.0 - v1.0.0
Recipe
adminer
Adminer on Kubernetes, rendered offline as plain manifests by the helm provider's Workload synthesis. The Workload becomes a Deployment and a NodePort Service so the UI is reachable. The database host resolved from the `databases` uses-edge arrives on .context.inject.env and renders as ADMINER_DEFAULT_SERVER.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/adminer:1.0.0 - v1.0.0
Recipe
a2a-remote
An agent another party runs over the Agent2Agent (A2A) protocol, reached at the address infra.endpoint names, as a reference realization: nothing is provisioned and no compute runs. The helm provider emits an ExternalName Service mapping an in-cluster name onto the agent's host. provides.endpoint is the address as written, protocol is its scheme, agent_card_url is where the agent's card is read, and credential_header is the request header the agent reads a caller's credential from, Authorization unless infra.credential_header names another. token_url, oidc_discovery_url and scopes are where and for what a caller obtains a token under the card's OAuth 2.0 or OpenID Connect scheme, as infra names them, and empty where it names none. No credential is held here: a consumer that needs one holds it on its own connection to the agent and presents it under that header, because the credential belongs to the consumer rather than to the address. infra.endpoint is the address without any part that carries a key, at least a scheme and a host. An address carrying a user name or password is refused, for infra.endpoint and infra.agent_card_url alike. infra.token_url and infra.oidc_discovery_url, where set, are each refused unless they are an https URL naming a host with no user name or password, since OAuth 2.0 sends a client secret only over TLS. Reach for this realization on Kubernetes; compose/a2a-remote is the same reference under Docker Compose.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/helm/a2a-remote:1.0.0 - v1.0.0
Recipe
weaviate
Weaviate for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, ports, environment). The rendered service always states its access posture: anonymous access is enabled where no credential is attached, and where an api_key credential is attached anonymous access is turned off, API-key authentication is turned on and the key arrives on the env pipe. Storage sizing and the persistent volume are a synthesized-Helm concern, so this local-development flavor writes its data inside the container and declares no data volume, mirroring how the Compose postgres flavor leaves storage to the helm target. provides.host is the bare service name, resolved on the compose network. With `values.metrics_enabled` true, PROMETHEUS_MONITORING_ENABLED turns on Weaviate's metrics at port 2112, which every peer on the compose network reaches, and provides.metrics_endpoint publishes the address; false, that key is empty. The port checks no caller and also answers /tenant-activity with each collection's tenant names.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/weaviate:1.0.0 - v1.0.0
Recipe
vonage
The Vonage voice carrier as a reference realization for local development with Docker Compose: no service is defined for it and nothing is provisioned, because the carrier runs outside the composition. What a consumer needs is the base address, both webhook addresses, the modality and the direction, and those reach the consuming service through the injection context rather than through a service definition — compose has no counterpart to an in-cluster alias, so the address travels to the caller instead of being aliased for it. The two webhook addresses are published to consumers exactly as the operator supplied them. They are opaque pointer strings: this recipe pipes each through unchanged, and nothing here resolves an address, contacts one, or derives one from the state of a cluster. This recipe pins no artifact. It names no image — the Release calls a service its own operator contracts for directly. No credential is delivered here, and none can be: with no service defined, a credential attached to one of the component's slots would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the carrier, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vonage:1.0.0 - v1.0.0
Recipe
vllm
vLLM as a self-hosted inference server for local development via Docker Compose. Synthesized-mode Workload: the image, the command line, the environment, the model-cache volume and the weights-pull step are translated from the Workload kind by the compose provider. The weights pull is an init container, so it becomes a run-to-completion service that the serving service waits on before it starts. The model value is a deterministic pointer that step resolves at start; the compiler never fetches weights. Both services mount the same named model-cache volume, so the weights outlive the containers and a restart re-runs the pull step against an already-populated cache — how the downloader reconciles that is the downloader's own behaviour. When the optional model_pull_token credential slot is attached its value arrives on the inject pipe and is rendered as the HF_TOKEN environment variable on the weights-pull service, which is therefore also the key the attached credential must publish under; unattached, no such variable is emitted and weights are pulled anonymously. An attached api_key is rendered the same way as VLLM_API_KEY on the serving service, and the server then requires it of every caller under /v1. A GPU count is reserved as that many NVIDIA GPUs for the serving service, which the Docker host must provide. Compose is single-host, so the replica count and the GPU runtime class the Kubernetes realization honours are not honoured here; the build reports each as an unsupported-field warning naming the setting rather than dropping it in silence. provides.host is the compose service name, which a consumer on the same compose network resolves directly.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vllm:1.0.0 - v1.0.0
Recipe
vertex-ai
Google Vertex AI's managed model API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied regional endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: consumers authenticate with ambient GCP workload identity or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vertex-ai:1.0.0 - v1.0.0
Recipe
vertex-ai-embedding
Google Vertex AI's managed embedding API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied regional endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: consumers authenticate with ambient GCP workload identity or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vertex-ai-embedding:1.0.0 - v1.0.0
Recipe
vertex-ai-agent
A Vertex AI agent as a reference realization for local development, so no service runs. The compiler carries the operator-supplied regional endpoint and the agent identifier to consumers through the injection contract, and delivers no credential — consumers authenticate with their own Google credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vertex-ai-agent:1.0.0 - v1.0.0
Recipe
valkey
Valkey for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, port, exposed service). Used as an in-memory cache with no persistence, so no data volume is declared. Authentication is off by default. An attached password is enforced: the server is started with --requirepass through the image's VALKEY_EXTRA_FLAGS variable, which keeps the value off the container's command line. The value is read under the variable the attachment's values.key names, or else VALKEY_PASSWORD, and an attachment delivering it under neither halts the build naming the slot. The image expands that variable unquoted, so a password holding a space stops the server at start, and one holding *, ? or [ is matched against file names in the data directory, so write the password without either. provides.host is the bare service name, resolved on the compose network. Encrypted client connections are not served here: infra.client_tls set to true, or a certificate attached to server_certificate, halts the build by name, because the server reads its certificate from files and no credential Feature realized on Compose names a secret a service can mount. The helm realizations valkey and valkey-chart serve TLS from that slot.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/valkey:1.0.0 - v1.0.0
Recipe
twilio
The Twilio messaging and voice carrier as a reference realization for local development with Docker Compose: no service is defined for it and nothing is provisioned, because the carrier runs outside the composition. What a consumer needs is the base address, the account coordinate, the modality and the direction, and those reach the consuming service through the injection context rather than through a service definition — compose has no counterpart to an in-cluster alias, so the address travels to the caller instead of being aliased for it. This recipe pins no artifact. It names no image — the Release calls a service its own operator contracts for directly. No credential is delivered here, and none can be: with no service defined, a credential attached to one of the component's slots would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the carrier, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/twilio:1.0.0 - v1.0.0
Recipe
vapi
The Vapi voice-agent platform as a reference realization for local development: no service is defined for it and nothing is provisioned. The compiler carries the platform's API base address and the assistant coordinates to consumers through the injection contract, so an application developed locally reaches the same platform it will reach in a cluster. This recipe pins no artifact. It names no image and no chart — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here, and none can be: with no service defined, a credential attached to the component's slot would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the platform, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/vapi:1.0.0 - v1.0.0
Recipe
tyk
Tyk Gateway for local development via Docker Compose, synthesized from the Workload kind by the compose provider. There is no chart to point at here, so this Recipe assembles the container over the same published image the Kubernetes realization's chart runs and sets the same gateway settings as environment overrides; everything else is left to the configuration file the image ships. The gateway holds no local state — the Redis-protocol store wired on the `cache` edge holds its keys, counters and sessions — so it renders stateless with no volume. With no credential on the cache connection the gateway connects to the store with an empty password, as the store components for this target expect; an attached one arrives under TYK_STORE_PASSWORD, the key it must publish under, and is rendered as a plain environment variable. A cache publishing valkey-tls or redis-tls is reached over TLS, checked against the image's trust store; a server_ca, or another TLS protocol, halts. A dependency derived from a uses-edge waits on the store having started rather than reporting healthy, so on a cold host the gateway may start first and retries; no readiness wait is hidden in a command line. Nothing is published to the host, so the data plane is reachable from the compose network only. The contract is key-for-key the Kubernetes realization's; only host differs, being the compose service name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/tyk:1.0.0 - v1.0.0
Recipe
trustyai-guardrails-gateway
TrustyAI's guardrails gateway in a model's path, for local development via Docker Compose: one service running the image in infra.image, synthesized from the Workload kind by the compose provider. A compose config carries its configuration file, pointing the gateway at the orchestrator wired on the orchestrator socket and declaring one route, values.route, that applies the detector that orchestrator registered — by the detector_id it publishes — with values.threshold and answers a detection with values.fallback_message; it is mounted as one file, which the GATEWAY_CONFIG variable names. The product serves no health route, so the gateway's one healthcheck opens a connection to its port with the bash the image carries. The service publishes no port on the host: callers sit on the same compose network, and a gateway that authenticates no caller is kept off the host's interfaces. provides.host is the compose service name, and the published contract is key for key the one the Kubernetes realization publishes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/trustyai-guardrails-gateway:1.0.0 - v1.0.0
Recipe
trustyai-huggingface-detector
TrustyAI's Hugging Face detector for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates it into one service running the image in infra.image, preceded by a weights-pull step. The pull step is an init container, so it becomes a run-to-completion service the detector waits on before it starts; it runs the same image and the Hugging Face command-line tool the image carries, and fetches the model named in values.model at values.model_revision into a named volume the detector reads through the MODEL_DIR variable. The compiler never fetches weights. The volume outlives the containers, so a restart re-runs the pull step against weights already in place, and how the tool reconciles that is its own behaviour. When the optional model_pull_token slot is attached its value arrives on the inject pipe and is rendered as the HF_TOKEN environment variable on the pull step alone, which is therefore also the key the attached credential must publish under; unattached, no such variable is emitted and weights are pulled anonymously. The detector's one healthcheck asks its health route with the curl the image carries, which answers only once the model is loaded. The service entry publishes no port on the host: callers sit on the same compose network, and a service that authenticates no caller is kept off the host's interfaces. provides.host is the compose service name, and the published contract is key for key the one the Kubernetes realization publishes; provides.metrics_endpoint is where the service serves its own metrics, in the Prometheus format at /metrics on its port.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/trustyai-huggingface-detector:1.0.0 - v1.0.0
Recipe
tgi
Text Generation Inference as a self-hosted inference server for local development via Docker Compose, synthesized from the Workload kind by the compose provider. A weights-pull step, running the runtime's own downloader from the serving image at the same tag, resolves the model pointer into a named model-cache volume before the server starts; the compiler never fetches weights. The server is started under the weights identifier, which is the model name published to consumers. An attached model_pull_token is rendered as the HF_TOKEN environment variable on the pull step, which is also the key the credential must publish under; unattached, weights are pulled anonymously. An attached api_key is rendered the same way as API_KEY on the serving service, and the server then requires it of every caller on its generation routes. A GPU count is honoured: the compose provider reserves that many NVIDIA GPUs for the serving service, which the Docker host must be able to hand out. Compose is single-host: the replica count and the GPU runtime class are authored all the same, and the build reports each as an unsupported-field warning naming the setting. provides.host is the compose service name, which a consumer on the same compose network resolves directly.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/tgi:1.0.0 - v1.0.0
Recipe
timescaledb
TimescaleDB for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry (image, port, exposed service). provides.host is the bare service name, resolved on the compose network. This is the local-development realization; on Kubernetes the component is realized by the synthesized timescaledb recipe, which needs nothing installed in the target cluster, and by timescaledb-cloudnative-pg, which emits a custom resource for the CloudNativePG control plane to reconcile.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/timescaledb:1.0.0 - v1.0.0
Recipe
tempo
Grafana Tempo distributed-tracing server for local development via Docker Compose, synthesized from the Workload kind by the compose provider: one service running the same grafana/tempo image, tagged from infra.engine_version, and the same single-binary configuration, the TempoConfig bundle, as the workload-synthesis realization on Kubernetes, delivered as a compose config mounted as the file the server reads, with the trace store on a named volume. The OpenTelemetry receivers are on, gRPC on 4317 and HTTP on 4318, beside the HTTP query API on 3200, and nothing is published to the host: a trace producer on the compose network reaches the gRPC receiver at the service name, provides.host, on provides.port. The server's own metrics, in the Prometheus format at /metrics on the HTTP port 3200, are published as provides.metrics_endpoint. The developer's retention_period drives the in-process compactor. Docker Compose sets no size on a volume, so infra.storage.size reaches nothing here. The healthcheck asks the server's readiness endpoint with the wget the image's busybox carries. Reach for it under Docker Compose; on Kubernetes, reach for `tempo` or `tempo-chart`.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/tempo:1.0.0 - v1.0.0
Recipe
telnyx
The Telnyx messaging and voice carrier as a reference realization for local development with Docker Compose: no service is defined for it and nothing is provisioned, because the carrier runs outside the composition. What a consumer needs is the base address, the modality and the direction, and those reach the consuming service through the injection context rather than through a service definition — compose has no counterpart to an in-cluster alias, so the address travels to the caller instead of being aliased for it. This recipe pins no artifact. It names no image — the Release calls a service its own operator contracts for directly. No credential is delivered here, and none can be: with no service defined, a credential attached to the component's slot would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the carrier, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/telnyx:1.0.0 - v1.0.0
Recipe
tei
Text Embeddings Inference as a self-hosted embedding server for local development via Docker Compose, synthesized from the Workload kind by the compose provider. A weights-pull step resolves the model pointer into a named model-cache volume before the server starts; the compiler never fetches weights. The serving image carries no download tool, so that step runs a general-purpose Python image pinned by this Recipe, and engine_version moves the serving image only. The server is started under the weights identifier, which is the model name published to consumers, with values.pooling as its pooling strategy. An attached model_pull_token is rendered as the HF_TOKEN environment variable on the pull step, which is also the key the credential must publish under; unattached, weights are pulled anonymously. An attached api_key is rendered the same way as API_KEY on the serving service, and the server then requires it of every caller, a scrape of provides.metrics_endpoint included. A GPU count is honoured: the compose provider reserves that many NVIDIA GPUs for the serving service, which the Docker host must be able to hand out. Compose is single-host: the replica count and the GPU runtime class are authored all the same, and the build reports each as an unsupported-field warning naming the setting. provides.host is the compose service name, which a consumer on the same compose network resolves directly.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/tei:1.0.0 - v1.0.0
Recipe
smolagents
A smolagents agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as SMOLAGENTS_AGENT_TYPE and SMOLAGENTS_MAX_STEPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/smolagents:1.0.0 - v1.0.0
Recipe
semantic-kernel
A Semantic Kernel agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; the model name is delivered as OPENAI_CHAT_MODEL_ID, the variable Semantic Kernel's OpenAI chat service reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. A connection writing an `inject:` block also receives the names it dictates, as variables or as files mounted at the paths it names, beside these. The agent's own values reach the container as SEMANTIC_KERNEL_PLANNER_TYPE and SEMANTIC_KERNEL_EXECUTION_SETTINGS (as JSON text, while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/semantic-kernel:1.0.0 - v1.0.0
Recipe
search
The managed web-search API as a reference realization for local development with Docker Compose: no service is defined for it and nothing is provisioned, because the search API runs outside the composition. What a consumer needs is the base address, the provider whose dialect it speaks and how many results to ask for, and those reach the consuming service through the injection context rather than through a service definition — compose has no counterpart to an in-cluster alias, so the address travels to the caller instead of being aliased for it. This recipe pins no artifact. It names no image — the deployment calls a service its own operator contracts for directly. No credential is delivered here: a consumer that needs the search API's key declares the `credential` slot on the socket that reaches the search API and delivers it from its own container, because the environment-variable name a credential arrives under belongs to the consumer rather than to the endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/search:1.0.0 - v1.0.0
Recipe
sagemaker
Amazon SageMaker's managed model serving endpoint as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied regional endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/sagemaker:1.0.0 - v1.0.0
Recipe
rivet
A Rivet agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The image carries both the Rivet runtime and the project file it executes. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as RIVET_PROJECT_FILE, RIVET_GRAPH_NAME and RIVET_INPUTS (as JSON text, while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/rivet:1.0.0 - v1.0.0
Recipe
s3-object-store
An S3-compatible object store on a Compose host: SeaweedFS 4.48, Apache-2.0, from its project's published chrislusf/seaweedfs image, run by the compose provider's Workload synthesis as one weed server process — master, volume server, filer and S3 gateway — with a named volume for its data and the S3 port published on the host. An init service creates the bucket values.bucket names before the server serves, and leaves an existing one as it is. Reach for it where the store runs on a Compose host with nothing else in place; for helm/s3-object-store in a Kubernetes cluster; for terraform/s3-object-store-amazon-s3 or cloudformation/s3-object-store-amazon-s3 in an AWS account; and for terraform/s3-object-store-gcs in a Google Cloud project. The access_key_id and secret_access_key slots are the server's S3 account, and both are required here: each value arrives on the inject pipe under AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY and is written into the service's environment as delivered, so a Compose variable reference such as ${S3_SECRET_ACCESS_KEY} keeps the key out of the file. An empty slot halts naming it. The server serves plain HTTP on the Compose network and the host; only the S3 gateway listens beyond the container, on its S3 port and the gRPC port beside it, and each refuses a caller without the key. It publishes the bucket, the endpoint http://<service>:8333, its host, port 8333, protocol s3 and region us-east-1, the one S3 clients read from a server reporting none. infra.region, any infra.overrides key and a bucket name SeaweedFS does not take halt naming the key.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/s3-object-store:1.0.0 - v1.0.0
Recipe
rabbitmq
RabbitMQ for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, ports, exposed service). RabbitMQ always authenticates, so the application user and password are always emitted: the user comes from the developer value and the password from the optional `password` credential slot when attached, falling back to a deterministic inline development default otherwise so the render stays reproducible. The image is management-enabled, so the broker exposes both the AMQP port and the management dashboard. provides.host is the bare service name, resolved on the compose network. The broker's metrics are on in the image, on port 15692, which checks no caller and which any peer on the compose network reaches. With `values.metrics_enabled` true their address is published as provides.metrics_endpoint; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/rabbitmq:1.0.0 - v1.0.0
Recipe
retell
The Retell AI voice-agent platform as a reference realization for local development: no service is defined for it and nothing is provisioned. The compiler carries the platform's API base address and the agent coordinates to consumers through the injection contract, so an application developed locally reaches the same platform it will reach in a cluster. This recipe pins no artifact. It names no image and no chart — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here, and none can be: with no service defined, a credential attached to the component's slot would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the platform, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/retell:1.0.0 - v1.0.0
Recipe
replicate
Replicate's managed prediction API as a reference realization: no compute runs in-cluster. The compiler carries the endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/replicate:1.0.0 - v1.0.0
Recipe
redis
Redis for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, port, exposed service). Used as an in-memory cache with no persistence, so no data volume is declared. Authentication is off by default. An attached password is enforced: the server is started with --requirepass in its arguments, since the image reads no variable that would carry it, so the value shows in the service's command. The value is read under the variable the attachment's values.key names, or else REDIS_PASSWORD, and an attachment delivering it under neither halts the build naming the slot. provides.host is the bare service name, resolved on the compose network. Encrypted client connections are not served here: infra.client_tls set to true, or a certificate attached to server_certificate, halts the build by name, because the server reads its certificate from files and no credential Feature realized on Compose names a secret a service can mount. The helm realization redis serves TLS from that slot. The packaged Redis server is AGPL-3.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/redis:1.0.0 - v1.0.0
Recipe
qdrant
Qdrant for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, ports, environment). Authentication is off by default; when an api_key credential is attached its value arrives on the env pipe and is rendered as the server's service API key, and a read_only_api_key credential is rendered the same way as its read-only key. Storage sizing and the persistent volume are a synthesized-Helm concern, so this local-development flavor is ephemeral and declares no data volume, mirroring how the Compose postgres flavor leaves storage to the helm target. provides.host is the bare service name, resolved on the compose network. provides.metrics_endpoint names the server's metrics on the REST port; while api_key is attached a scrape presents the read-only key where read_only_api_key is attached, and the api_key otherwise.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/qdrant:1.0.0 - v1.0.0
Recipe
pydantic-ai
A Pydantic AI agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as PYDANTIC_AI_RESULT_TYPE, PYDANTIC_AI_RETRIES and PYDANTIC_AI_SYSTEM_PROMPT (while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/pydantic-ai:1.0.0 - v1.0.0
Recipe
promptflow
A Prompt flow application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. Its own value flow_entrypoint reaches the container as PROMPTFLOW_FLOW_ENTRYPOINT by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/promptflow:1.0.0 - v1.0.0
Recipe
promptlayer
The hosted prompt registry as a reference realization for local development with Docker Compose: no service is defined for it and nothing is provisioned, because the registry runs outside the composition. What a consumer needs is the base address and the release label, and those reach the consuming service through the injection context rather than through a service definition — compose has no counterpart to an in-cluster alias, so the address travels to the caller instead of being aliased for it. This recipe pins no artifact. It names no image — the deployment calls a service its own operator contracts for directly. No credential is delivered here: a consumer that needs the registry's API key declares the `credential` slot on the socket that reaches the registry and delivers it from its own container, because the environment-variable name a credential arrives under belongs to the consumer rather than to the endpoint.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/promptlayer:1.0.0 - v1.0.0
Recipe
prometheus
Prometheus metrics server for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload into a service entry. Each connection on the `scrape_target` uses-socket becomes one job in a minimal prometheus.yml the container writes at start-up: the `metrics_endpoint` that connection's target publishes is the job's address, and the connection's own `values` set that job's label and cadence. A target publishing none halts the build, naming it, and so does one publishing `metrics_service` or `metrics_pod_selector`: its metrics are found by discovering it in a Kubernetes cluster, which Docker Compose does not have. A connection's `credential` arrives as a container variable, which the start-up command writes to the file the job presents as a bearer token. The time-series database is persisted to a named volume. With `infra.external_base_url` set, the server takes it as its external URL and is kept serving at its root; unset, neither flag is emitted. provides.host is the bare service name, resolved on the compose network. The server's port is also published on the host: on its loopback address, 127.0.0.1, only, or on every host interface while `infra.publish_on_all_interfaces` is true. With `values.otlp_receiver` true, the server runs the OTLP receiver, which checks no caller, on that same port, so anything that reaches the port can write metrics into it, and it publishes the receiver's address as provides.otlp_http_endpoint; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/prometheus:1.0.0 - v1.0.0
Recipe
presidio-anonymizer
The Presidio anonymizer for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates it into one service running the vendor's published image, with the listen port handed to the image through the PORT variable it reads. The service keeps no state, so it mounts no volume. Its one healthcheck is the image's own — curl against /health, which the image carries — with the timings the image declares. The service entry publishes no port on the host: callers sit on the same compose network, and a service that authenticates no caller is kept off the host's interfaces. provides.host is the compose service name, which a consumer on that network resolves directly, and the published contract is key for key the one the Kubernetes realization publishes. The image is published under the MIT licence.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/presidio-anonymizer:1.0.0 - v1.0.0
Recipe
presidio-analyzer
The Presidio analyzer for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates it into one service running the vendor's published image, with the listen port handed to the image through the PORT variable it reads. The service keeps no state, so it mounts no volume. Its one healthcheck is the image's own — curl against /health, which the image carries — with the timings the image declares. The service entry publishes no port on the host: callers sit on the same compose network, and a service that authenticates no caller is kept off the host's interfaces. provides.host is the compose service name, which a consumer on that network resolves directly, and the published contract is key for key the one the Kubernetes realization publishes. The image is published under the MIT licence.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/presidio-analyzer:1.0.0 - v1.0.0
Recipe
postgres
PostgreSQL for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry (image, port, exposed service). provides.host is the bare service name, resolved on the compose network. Bootstrap scripts (init_scripts) and a custom server configuration block (custom_config) are both realized: each script mounts under /docker-entrypoint-initdb.d and runs once when the database first initializes, and the configuration block mounts as the postgresql.conf the server's config_file argument points at.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/postgres:1.0.0 - v1.0.0
Recipe
phidata
A Phidata agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL; the datastore wired into `storage` contributes the coordinates for session history and the store wired into `knowledge_base` those for retrieved reference material; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. A store served over TLS alone or over HTTPS halts the build on a store connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as PHIDATA_ADD_HISTORY_TO_MESSAGES and PHIDATA_MARKDOWN by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/phidata:1.0.0 - v1.0.0
Recipe
otel-collector
The OpenTelemetry Collector, contrib distribution, for local development via Docker Compose, synthesized from the Workload kind by the compose provider: one stateless service running the same image as the Kubernetes realization with the same configuration, the OtelCollectorConfig bundle, delivered as a compose config mounted over the file the image starts the collector with. The bundle writes one exporter for each backend wired to the traces, metrics and logs sockets, and the service carries a digest of the infra and those backends, so a changed address re-creates it. The service publishes nothing to the host: a producer on the compose network reaches it at the service name, on the OTLP gRPC and HTTP ports. The image carries no shell and no client to ask the collector's health check from inside the container, so the service declares no healthcheck. provides.host is the compose service name. With `values.metrics_enabled` true the collector serves its own telemetry metrics at port 8888 on every interface, which every peer on the compose network reaches and which checks no caller, and provides.metrics_endpoint publishes the address; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/otel-collector:1.0.0 - v1.0.0
Recipe
pgvector
pgvector for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry running the pgvector project's own image, with a named data volume and the script that enables the extension in the database the server creates on its first start. provides.host is the bare service name, resolved on the compose network, and the port is also published on the host, so a client on the developer's own machine connects to it. Reach for this realization for local development; on Kubernetes the component is realized by the Helm Recipe of the same name, which synthesizes a StatefulSet and needs nothing installed in the target cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/pgvector:1.0.0 - v1.0.0
Recipe
opensearch
OpenSearch for local development via Docker Compose. Synthesized-mode Workload: one service running the product's own image as one node, with its indexes on a named volume. provides.host is the bare service name, resolved on the compose network; the port is published on no host interface, so only the stack's own services reach the server. The security plugin stays on. The image's demonstration setup is replaced by one account, admin, whose password is the value the attached credential delivers, written into the compose file as delivered — a Compose variable reference such as ${OPENSEARCH_ADMIN_PASSWORD} keeps the material out of the file; with none attached it is a development default. The service generates its transport key each time it starts. The REST API is served over plain HTTP, not encrypted, and is reachable only inside the Release. Reach for this realization for local development; on Kubernetes the component is realized by the Helm Recipe opensearch-chart.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/opensearch:1.0.0 - v1.0.0
Recipe
openhands
An OpenHands agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as OPENHANDS_TASK (while set) and OPENHANDS_MAX_ITERATIONS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openhands:1.0.0 - v1.0.0
Recipe
openai
OpenAI's managed API as a reference realization for local development: no service runs. The compiler carries the endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openai:1.0.0 - v1.0.0
Recipe
openclaw
An OpenClaw agent for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The image carries both the agent and the JSON configuration it reads. A model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY, which the agent reads directly. The model endpoint published by the component wired into the `model` socket is rendered under a name this recipe defines, because the agent resolves its endpoint from its configuration file rather than from the environment. The endpoint and the model name are delivered as OPENCLAW_MODEL_BASE_URL and OPENCLAW_MODEL by default; a connection writing an `inject:` block receives the names it dictates in their place, as variables or as files mounted at the paths it names, while the API key stays. The agent's own values reach the container as OPENCLAW_EXECUTION_MODE, OPENCLAW_MAX_EXECUTION_TIME and OPENCLAW_MAX_STEPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openclaw:1.0.0 - v1.0.0
Recipe
openai-embedding
OpenAI's managed embeddings API as a reference realization for local development: no service runs. The compiler carries the endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openai-embedding:1.0.0 - v1.0.0
Recipe
openai-agents-sdk
An application built on the OpenAI Agents SDK, for local development via Docker Compose. Synthesized-mode Workload: the image the Release supplies and its environment are translated from the Workload kind by the compose provider. The model wired into the `model` socket reaches the application as the variables the SDK reads: its endpoint as OPENAI_BASE_URL, its model name as OPENAI_DEFAULT_MODEL, and the connection's credential as OPENAI_API_KEY. While the component's upload_traces_to_openai value is false, its default, it also receives OPENAI_AGENTS_DISABLE_TRACING set to 1, which turns off the SDK's trace upload to OpenAI; set true, the variable is absent. Each tool wired into the `tools` socket reaches it under the variables helm/openai-agents-sdk names: its address as AGGREGA_TOOL_<ALIAS>_URL, or the whole address the connection holds as a secret in its place, and the connection's credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header the tool publishes, AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block, or the component's own, delivers the names it dictates, as on helm/openai-agents-sdk, a file as a compose config mounted at the path it names. Compose has no secret reference, so a secret held on a connection is written into the file as it was delivered; a Compose variable reference such as ${MODEL_API_KEY} is written as that reference, and Docker Compose substitutes it when the stack starts. Two secrets the Release alone attaches, a model credential beside a tool credential, both arrive under the key value and halt the build naming both; an Aggregate naming values.key on one of them avoids it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openai-agents-sdk:1.0.0 - v1.0.0
Recipe
openai-compatible
An OpenAI-compatible broker gateway as a reference realization for local development: no service runs. The compiler carries the operator-supplied gateway endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/openai-compatible:1.0.0 - v1.0.0
Recipe
ollama
Ollama as a self-hosted inference server for local development via Docker Compose, synthesized from the Workload kind by the compose provider. A weights-pull step, running the same image at the same tag, resolves the model pointer into a named model-cache volume before the server starts, and a failed pull holds the server back; the compiler never fetches weights. The server serves what the cache holds, so the model name published to consumers is the weights identifier itself. An attached model_pull_token is rendered as the OLLAMA_PULL_TOKEN environment variable on the pull step, which is also the key the credential must publish under; unattached, weights are pulled anonymously. A GPU count is honoured: the compose provider reserves that many NVIDIA GPUs for the serving service, which the Docker host must be able to hand out. Compose is single-host: the replica count and the GPU runtime class are authored all the same, and the build reports each as an unsupported-field warning naming the setting. provides.host is the compose service name, which a consumer on the same compose network resolves directly.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/ollama:1.0.0 - v1.0.0
Recipe
nvidia-nim
NVIDIA Inference Microservice as a self-hosted, GPU-native inference server on a single Docker host via Docker Compose, synthesized from the Workload kind. The NIM container resolves the model profile into a named model-cache volume at start and serves the OpenAI-compatible API on the compose network; the compiler never fetches the profile. The image is infra.image, which the Release supplies: the catalog names no NIM image, since NIM images are distributed under NVIDIA's own terms. Docker pulls it with the host's own registry login, so where the image needs credentials, log the host in to nvcr.io first; infra.image_pull_secret names a Kubernetes Secret, and a Release setting it here halts the build. An attached ngc_api_key arrives on the inject pipe and is rendered as the NGC_API_KEY variable, written into the file as delivered: attach a Compose variable reference such as ${NGC_API_KEY} to keep the key out of it. Unattached, no such variable is emitted. The service always reserves infra.gpu.count NVIDIA GPUs, at least one, which the host must hand to containers through the NVIDIA Container Toolkit. The replica count and the GPU runtime class are reported as unsupported-field warnings, and the service declares no healthcheck, because NVIDIA does not document a health client inside the image. provides.host is the compose service name; provides.metrics_endpoint is always set, since compose runs one instance.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/nvidia-nim:1.0.0 - v1.0.0
Recipe
nvidia-api
The managed NVIDIA API as a reference realization for local development: no service runs. The compiler carries the endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/nvidia-api:1.0.0 - v1.0.0
Recipe
nats
NATS for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, port, exposed service). Used as an in-memory messaging server with no persistence, so no data volume is declared. Authentication is off by default; when a password credential is attached its value arrives on the env pipe and the server is started with user/password authentication flags. With no credential attached the server runs unauthenticated and the render stays reproducible. provides.host is the bare service name, resolved on the compose network.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/nats:1.0.0 - v1.0.0
Recipe
nemo-guardrails
The NeMo Guardrails server as a guard in a model's path, for local development via Docker Compose: one service running the image in infra.image, synthesized from the Workload kind by the compose provider. Two compose configs carry the guard's configuration as the product's own two files — config.yaml, naming the Hugging Face classifier rail pointed at the detector wired on the detector socket, with values.threshold and values.blocked_labels; and rails.co, the flow that keeps a request body from switching that rail off. The guard starts in the product's single-configuration mode and refuses a request naming any other configuration. It writes no NemoGuard rail, and halts while the content_safety or jailbreak_detection socket holds a connection. Compose has no env-from-secret primitive, so an attached upstream_key credential is rendered as a plain environment variable, under the name config.yaml gives it, which is therefore also the key the attached credential must publish under; unattached, the guard presents none. The upstream address reaches the guard as MAIN_MODEL_BASE_URL, from infra.upstream_endpoint. The service publishes no port on the host: callers sit on the same compose network. provides.host is the compose service name, and the published contract is key for key the one the Kubernetes realization publishes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/nemo-guardrails:1.0.0 - v1.0.0
Recipe
mysql
MySQL for local development via Docker Compose. Synthesized-mode Workload — image, environment, ports, persistent volume, healthcheck are translated from the Workload kind by the compose provider. The database name and credentials are delivered through the MySQL image's environment contract (MYSQL_DATABASE, MYSQL_USER, MYSQL_ROOT_PASSWORD, MYSQL_PASSWORD). Character set and collation are chart- and managed-target knobs; the MySQL container image takes them as server command-line arguments rather than environment variables, so this local-development target leaves them at the image defaults. Secrets injected via the dotenv-secret FeatureRecipe surface as env vars in the rendered service. Bootstrap scripts (init_scripts) and a custom server configuration block (custom_config) are both realized: each script mounts under /docker-entrypoint-initdb.d and runs once when the database first initializes, and the configuration block mounts as a .cnf the server reads through its defaults-extra-file argument.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/mysql:1.0.0 - v1.0.0
Recipe
mistral
Mistral AI's managed chat API as a reference realization: no compute runs in-cluster. The compiler carries the endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/mistral:1.0.0 - v1.0.0
Recipe
milvus
Milvus for local development via Docker Compose. Synthesized-mode Workload: one service running the product's own image in its single-container mode, with the metadata store inside the server process and the data on a named volume, so no etcd or object-store service is deployed beside it. provides.host is the bare service name, resolved on the compose network; the port is published on no host interface, so only the stack's own services reach the server. Authentication is off unless a password is attached; attached, it is switched on for the API and for the management endpoints the metrics port serves, whose health check and metrics stay open, and the password is the value the credential delivers, written into the compose file as delivered — a Compose variable reference such as ${MILVUS_ROOT_PASSWORD} keeps the material out of the file. Reach for this realization for local development; on Kubernetes the component is realized by the Helm Recipe milvus-chart.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/milvus:1.0.0 - v1.0.0
Recipe
metagpt
A MetaGPT agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as METAGPT_ROLE and METAGPT_OBJECTIVE by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/metagpt:1.0.0 - v1.0.0
Recipe
memcached
Memcached for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload kind into a service entry (image, port, exposed service). In-memory cache with no persistence, so no data volume is declared. The cache runs without authentication, and a password attached to the password slot halts the build naming the slot: the memcached image enforces no password, and its two authentication modes, SASL over the binary protocol and an ASCII authentication file, are not spoken by the catalog's Memcached clients, so the cache would accept every client whatever the attachment. provides.host is the bare service name, resolved on the compose network. Encrypted client connections are not served here: infra.client_tls set to true, or a certificate attached to server_certificate, halts the build by name, because the server reads its certificate from files and no credential Feature realized on Compose names a secret a service can mount. The helm realization memcached serves TLS from that slot.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/memcached:1.0.0 - v1.0.0
Recipe
mastra
A server built by Mastra, for local development via Docker Compose. Synthesized-mode Workload: the image the Release supplies and its environment are translated from the Workload kind by the compose provider. The server receives infra.port as PORT beside MASTRA_HOST set to 0.0.0.0, and provides is the service's name on the compose network with that port, and, where values.agent_id is set, the address of that agent's Agent2Agent card. The model wired into the `model` socket reaches the server as on helm/mastra: its endpoint as OPENAI_BASE_URL, and the connection's credential as OPENAI_API_KEY. Each tool wired into the `tools` socket reaches it under the variables helm/mastra names: its address as AGGREGA_TOOL_<ALIAS>_URL, or the whole address the connection holds as a secret in its place, and the connection's credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header the tool publishes, AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block, or the component's own, delivers the names it dictates, as on helm/mastra, a file as a compose config mounted at the path it names. A set values.agent_id reaches the server as MASTRA_AGENT_ID unless an entry of the component's own block reads it alone. Compose has no secret reference, so a secret held on a connection is written into the file as it was delivered; a Compose variable reference such as ${MODEL_API_KEY} is written as that reference, and Docker Compose substitutes it when the stack starts. Two secrets the Release alone attaches, a model credential beside a tool credential, both arrive under the key value and halt the build naming both; an Aggregate naming values.key on one of them avoids it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/mastra:1.0.0 - v1.0.0
Recipe
loki
Grafana Loki log-aggregation server for local development via Docker Compose, synthesized from the Workload kind by the compose provider: one service running the grafana/loki image at the release the Kubernetes realization pins, in single-binary mode, from one configuration file mounted from a compose config, with the chunk store on a named volume. The configuration carries what the Kubernetes realization configures — authentication off, one replica, a TSDB schema at v13 over a filesystem store, and the developer's retention_period enforced by the in-process compactor — with structured metadata accepted, so the server takes OTLP logs over HTTP at /otlp on provides.port beside its push and query API, and serves its own metrics in the Prometheus format at /metrics on the same port, published as provides.metrics_endpoint. A memcached wired to the chunk_cache socket becomes the server's chunk-cache client, driven by that edge's batch size, parallelism, timeout and entry lifetime, over TLS to one publishing memcached-tls, checked against the image's trust store; a server_ca on that connection halts the build. Unwired, there is none. Nothing is published to the host: a consumer on the compose network reaches the server at the service name, provides.host. Docker Compose sets no size on a volume, so infra.storage.size reaches nothing here. The image carries no shell or client and Loki ships no health command, so the service declares no healthcheck. Reach for it under Docker Compose; on Kubernetes, reach for `loki`.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/loki:1.0.0 - v1.0.0
Recipe
mariadb
MariaDB for local development via Docker Compose. Synthesized-mode Workload — image, environment, ports, persistent volume, healthcheck are translated from the Workload kind by the compose provider. Secrets injected via the dotenv-secret FeatureRecipe surface as env vars in the rendered service. Bootstrap scripts (init_scripts) and a custom server configuration block (custom_config) are both realized: each script mounts under /docker-entrypoint-initdb.d and runs once when the database first initializes, and the configuration block mounts as a .cnf the server reads through its defaults-extra-file argument. It runs one server, so infra.ha.enabled and a non-zero infra.ha.read_replicas halt the build naming the key; helm/mariadb-cluster honors both.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/mariadb:1.0.0 - v1.0.0
Recipe
llamaindex
A LlamaIndex agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as LLAMAINDEX_AGENT_MODE, LLAMAINDEX_SYSTEM_PROMPT (while set), LLAMAINDEX_MAX_FUNCTION_CALLS and LLAMAINDEX_LOG_LEVEL by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/llamaindex:1.0.0 - v1.0.0
Recipe
litellm
LiteLLM as a self-hosted, OpenAI-compatible model proxy for local development via Docker Compose, synthesized from the Workload kind by the compose provider. A compose config carries the proxy's configuration, which publishes the alias in values.model and points it at values.upstream_model. The configuration contract is authored against LiteLLM v1.103.1. Compose has no env-from-secret primitive, so an attached credential is rendered as a plain environment variable, and a credential written as a Compose variable reference lands as that reference rather than its value. On every Release two further configs carry a request guard, which forwards only the request fields the proxy declares for the route, plus any the Release admits in infra.admitted_request_fields, and refuses the rest naming the field; and a path allowlist, a service of its own running the official nginx image (BSD-2-Clause) in the proxy's network namespace on infra.port, which answers 403 to every path but the model routes, the model listing, the health routes and key management. Both come from the LiteLLMGuard bundle. PROXY_BASE_URL is emitted only when infra.external_base_url is set, the build halting on one carrying a path. With `values.metrics_enabled` true the proxy serves its Prometheus metrics on its loopback address, and the path allowlist forwards a GET of /metrics alone there from port 9090, which every peer on the compose network reaches, published as provides.metrics_endpoint; false leaves that key empty. The port checks no caller, and the metrics name the hashed keys, key aliases and user emails the proxy attributes spend to.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/litellm:1.0.0 - v1.0.0
Recipe
litellm-gateway
The AI gateway for local development via Docker Compose, running the official upstream proxy image. Its configuration, the routing table from values.models and the models socket, is a compose config the proxy reads, from the bundle the Kubernetes realizations import. It is proven at LiteLLM v1.103.1 and halts on any other engine_version. As on Kubernetes, a request guard forwards only the request fields the proxy declares for the route, plus any in infra.admitted_request_fields, and a path allowlist, the official nginx image (BSD-2-Clause) in the proxy's network namespace on infra.port, answers 403 to every path but the gateway's routes. values.log_masking is honoured, entity types through the Presidio services on the pii_analyzer and pii_anonymizer sockets. core/prompt-deny-patterns and core/fallback-chain are written into the configuration from their attachments. Compose has no env-from-secret primitive, so an attached credential is rendered as a plain variable; a Compose variable reference stays one. The cache_store socket is refused: the semantic cache it serves has no Compose realization.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/litellm-gateway:1.0.0 - v1.0.0
Recipe
lindy
A Lindy assistant on Lindy's managed platform as a reference realization for local development, so no service runs. The compiler carries the operator-supplied platform endpoint and the assistant identifier to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/lindy:1.0.0 - v1.0.0
Recipe
letta
The Letta stateful-agent platform as a reference realization for local development: no service is defined for it and nothing is provisioned. The compiler carries the platform's API base address and the agent identifier to consumers through the injection contract, so an application developed locally reaches the same platform it will reach in a cluster. This recipe pins no artifact. It names no image and no chart — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here, and none can be: with no service defined, a credential attached to the component's slot would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the platform, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/letta:1.0.0 - v1.0.0
Recipe
kong
Kong, a database-backed API gateway, for local development via Docker Compose, synthesized from the Workload kind by the compose provider. The PostgreSQL backend on the `backend` edge is the system of record, so it renders stateless with no volume. The connection's host, port, user and database arrive on the inject pipe under the four KONG_PG_* variables the edge's inject block names; an edge leaving any missing halts the build naming the edge and the variables. A schema-provisioning step runs the same image and connection before the gateway starts. With no credential on the backend connection, the steps and the gateway share the inline development default the PostgreSQL Recipe for this target also falls back to; an attached one is rendered as KONG_PG_PASSWORD, the key it must publish under. Each model wired to the models socket gets a chat and a completion route through Kong's AI proxy, in Kong's declarative configuration that a second step imports into the database before the gateway starts. A connection's credential is the value it publishes as AI_PROXY_<ALIAS>_API_KEY, written into the bearer token the routes name through Kong's environment vault: a Compose variable reference such as ${OPENAI_API_KEY} lands as that reference, a literal as itself. With `values.metrics_enabled` true the Status API moves off loopback to port 8007 of every interface and provides.metrics_endpoint publishes /metrics on it; false, that key is empty. While it is on, every peer on the Compose network reaches that port, but only a GET of /status, /status/ready or /metrics: the KongStatusGuard bundle refuses the rest, POST /status/ready among them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/kong:1.0.0 - v1.0.0
Recipe
haystack
A Haystack agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. Nothing is delivered from the `document_store` connection by default. A connection writing an `inject:` block receives the names it dictates instead of the defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as HAYSTACK_PIPELINE_YAML_PATH and HAYSTACK_MAX_LOOPS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/haystack:1.0.0 - v1.0.0
Recipe
huggingface
Hugging Face's managed Inference API as a reference realization: no compute runs in-cluster. The compiler carries the endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/huggingface:1.0.0 - v1.0.0
Recipe
gumloop
A Gumloop automation pipeline on Gumloop's managed platform as a reference realization for local development, so no service runs. The compiler carries the managed endpoint and the pipeline identifier to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/gumloop:1.0.0 - v1.0.0
Recipe
griptape
A Griptape agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL; the store wired into the `task_memory` socket contributes its host and port; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. A task-memory store served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as GRIPTAPE_OFF_PROMPT and GRIPTAPE_STREAM by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/griptape:1.0.0 - v1.0.0
Recipe
greeter-green
The ALTERNATE physical implementation of example/compute/greeter — the "green" build. No greeter instance reaches it through the type tier; it is selected only when the Manifest name tier overrides a specific instance to it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/greeter-green:1.0.0 - v1.0.0
Recipe
greeter-blue
The DEFAULT physical implementation of example/compute/greeter — the "blue" build. Every greeter instance resolves here through the Manifest type tier unless a per-instance name override redirects it elsewhere.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/greeter-blue:1.0.0 - v1.0.0
Recipe
grafana
Grafana dashboards server for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload into a service entry (official grafana/grafana image, web port, admin environment, named volume for state). The admin username is a developer value; the admin password is wired through the optional `admin_password` credential slot: when attached its value arrives on .context.inject.env and becomes GF_SECURITY_ADMIN_PASSWORD; when none is attached the Recipe falls back to a deterministic development default. Grafana's local state is persisted to a named volume. With `infra.external_base_url` set, it reaches the server as its root URL (GF_SERVER_ROOT_URL); unset, that variable is not emitted and the server keeps its default. The server listens on `infra.port`, which also carries its own metrics in the Prometheus format at /metrics, published as provides.metrics_endpoint. provides.host is the bare service name, resolved on the compose network. The packaged Grafana server is AGPL-3.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/grafana:1.0.0 - v1.0.0
Recipe
gemini
Google's managed Gemini API as a reference realization for local development: no service runs. The compiler carries the endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/gemini:1.0.0 - v1.0.0
Recipe
gemini-agent
A Gemini assistant on Google's generative-language API as a reference realization for local development, so no service runs. The compiler carries the endpoint and the model string to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/gemini-agent:1.0.0 - v1.0.0
Recipe
fms-guardrails-orchestrator
The FMS Guardrails Orchestrator for local development via Docker Compose: one service running the image in infra.image, synthesized from the Workload kind by the compose provider. A compose config carries its configuration file, naming the upstream server in infra.upstream_host and infra.upstream_port as the one chat requests are forwarded to and registering the detector wired on the detector socket under values.detector_id with values.default_threshold; it is mounted as one file, at a path the image does not use, which the ORCHESTRATOR_CONFIG variable names. The orchestrator's one healthcheck asks its health route with the curl the image carries. The service publishes no port on the host: callers sit on the same compose network, and an orchestrator that authenticates no caller is kept off the host's interfaces. provides.host is the compose service name, and the published contract is key for key the one the Kubernetes realization publishes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/fms-guardrails-orchestrator:1.0.0 - v1.0.0
Recipe
fluentd
Fluentd log aggregator for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload into a service entry. Fluentd is the receiving end: a minimal fluent.conf with a `forward` input source bound to the infra port is written at start-up, and records whose tag matches the developer's output match are routed to standard output. The buffer is in-memory by default; when infra.storage.size is set a named volume backs a file buffer. provides.host is the bare service name, resolved on the compose network, and provides.port is the forward input port upstream forwarders ship records to.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/fluentd:1.0.0 - v1.0.0
Recipe
fluent-bit
Fluent Bit log forwarder for local development via Docker Compose. Synthesized-mode Workload — the compose provider translates the Workload into a service entry. The sink host and port resolved from the `forward_target` uses-edge arrive on .context.inject.env and are written into a minimal fluent-bit.conf `forward` output at start-up, so the forwarder ships records to that sink. Fluent Bit holds no persistent state. Its monitoring endpoint on the infra port also serves its metrics in the Prometheus format at /api/v2/metrics/prometheus, published as provides.metrics_endpoint. provides.host is the bare service name, resolved on the compose network. The monitoring port, which checks no caller, is also published on the host: on its loopback address, 127.0.0.1, only, or on every host interface while `infra.publish_on_all_interfaces` is true. Containers on the compose network reach it either way.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/fluent-bit:1.0.0 - v1.0.0
Recipe
etcd
Single-node etcd for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry (image, environment, client and peer ports, a named volume for the data directory). etcd holds its data on the named volume and serves the v3 client API over HTTP without authentication. Its own metrics, in the Prometheus format at /metrics on the client port, are published as provides.metrics_endpoint. provides.host is the bare service name, resolved on the compose network.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/etcd:1.0.0 - v1.0.0
Recipe
ferretdb
A MongoDB-wire-compatible document-database proxy for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry (image, environment, wire port). The proxy holds no local state, so no volume is rendered. The PostgreSQL backend identity resolved from the `backend` uses-edge (host, port, user, database) arrives on .context.inject.env and is composed into the proxy's PostgreSQL connection URL; the URL is rendered password-free and the backend password is supplied separately on the standard libpq PGPASSWORD environment variable, wired through the optional credential on the `backend` connection with an inline development default when none is attached. provides.host is the bare service name, resolved on the compose network. FERRETDB_DEBUG_ADDR binds FerretDB's debug listener, port 8088, to the container's loopback address while `values.metrics_enabled` is false, and provides.metrics_endpoint is empty. True, it binds it on every interface, and provides.metrics_endpoint publishes its /debug/metrics. While it is on, every peer on the compose network reaches that listener with no caller check, and beside the metrics it serves /debug/pprof, which takes CPU profiles and execution traces on request, /debug/vars, /debug/archive and /debug/graphs.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/ferretdb:1.0.0 - v1.0.0
Recipe
elasticsearch
Elasticsearch for local development via Docker Compose. Synthesized-mode Workload: one service running Elastic's own image as one node, with its indexes on a named volume. provides.host is the bare service name, resolved on the compose network; the port is published on no host interface, so only the stack's own services reach the server. Authentication stays on with the built-in superuser, elastic, whose password is the value the attached credential delivers, written into the compose file as delivered — a Compose variable reference such as ${ELASTIC_PASSWORD} keeps the material out of the file; with none attached it is a development default. The image's first-start setup, which generates certificates and prints a password, is switched off. The REST API is served over plain HTTP, not encrypted, and is reachable only inside the Release; the transport layer listens inside the container only. Reach for this realization for local development; on Kubernetes the component is realized by the Helm Recipe elasticsearch.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/elasticsearch:1.0.0 - v1.0.0
Recipe
documentdb
PostgreSQL with the DocumentDB extension for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry (image, environment, port, a named volume for the data directory). POSTGRES_USER / POSTGRES_DB are developer values; POSTGRES_PASSWORD is wired through the optional `password` credential slot, whose value arrives on the inject pipe when a credential is attached and falls back to an inline development default when none is. provides.host is the bare service name, resolved on the compose network.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/documentdb:1.0.0 - v1.0.0
Recipe
dynamiq
A Dynamiq agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The image carries both the framework and the workflow it executes. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_URL — the variable Dynamiq's own OpenAI connection reads, which is not the one the bare client SDK would consult. The datastore wired into the `state_store` socket contributes its host and port; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. A connection writing an `inject:` block receives the names it dictates instead of these defaults, as variables or as files mounted at the paths it names, while OPENAI_URL and the API key stay. A state store served over TLS alone or over HTTPS halts the build on a connection writing no `inject:` block, naming the connection, since such a connection hands the application only the store's host and port. The agent's own values reach the container as DYNAMIQ_WORKFLOW_ENTRYPOINT and DYNAMIQ_STREAM_EVENTS by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/dynamiq:1.0.0 - v1.0.0
Recipe
docker-model-runner
A model served by the host's own Docker Model Runner, for local development via Docker Compose. The Recipe writes the model into the compose file's top-level `models` element under the component's own name, with its context size and runtime flags, and Compose pulls and configures it there when the stack starts. Beside it the Recipe adds one service under the same name, granted the model. Compose leaves out a model no service is granted, so without it a plain `docker compose up` would pull nothing. The service runs a minimal image that exits at once; Compose has already pulled and configured the model before it starts any container. provides.endpoint is the address containers reach the host's runner at, infra.endpoint, which defaults to Docker Desktop's http://model-runner.docker.internal/engines/v1; on Docker Engine a Release sets it to the address the host's runner answers containers at. Host, port and protocol are read from it. The settings that size or place an in-cluster server — port, engine_version, replicas, compute, gpu and model_cache — are refused here by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/docker-model-runner:1.0.0 - v1.0.0
Recipe
databricks
A Databricks model serving endpoint as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/databricks:1.0.0 - v1.0.0
Recipe
cloudflare-ai-gateway
References a Cloudflare AI Gateway that already exists, as a reference realization for local development: no service runs and nothing in the account is created or changed. The compiler carries the gateway's address to consumers through the injection contract. Reach for it where the gateway is created and configured outside the Release; reach for the terraform Recipe where the Release is to create it. It carries no Feature, because every control the gateway keeps is a setting of the gateway itself, which this realization does not write, so the build halts on any attachment. No credential is delivered: a consumer that needs one declares it as `credential` on the socket that reaches the gateway and delivers it from its own container.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/cloudflare-ai-gateway:1.0.0 - v1.0.0
Recipe
cohere
Cohere's managed Command API as a reference realization: no compute runs in-cluster. The compiler carries the endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/cohere:1.0.0 - v1.0.0
Recipe
claude-agent-sdk
An application built on the Claude Agent SDK, for local development via Docker Compose. Synthesized-mode Workload: the image the Release supplies and its environment are translated from the Workload kind by the compose provider. The model wired into the `model` socket reaches the command line each session runs in as the variables it reads: its endpoint as ANTHROPIC_BASE_URL, its model name as ANTHROPIC_MODEL, and the connection's credential as ANTHROPIC_API_KEY. Each tool wired into the `tools` socket reaches the application under the variables helm/claude-agent-sdk names: its address as AGGREGA_TOOL_<ALIAS>_URL, or the whole address the connection holds as a secret in its place, and the connection's credential as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header the tool publishes, AGGREGA_TOOL_<ALIAS>_CREDENTIAL_HEADER. A connection writing an `inject:` block also receives the names it dictates, as on helm/claude-agent-sdk, a file as a compose config mounted at the path it names. Compose has no secret reference, so a secret held on a connection is written into the file as it was delivered; a Compose variable reference such as ${ANTHROPIC_KEY} is written as that reference, and Docker Compose substitutes it when the stack starts. Two secrets the Release alone attaches, a model credential beside a tool credential, both arrive under the key value and halt the build naming both; an Aggregate naming values.key on one of them avoids it. The component's own `inject:` block delivers its entries as a connection's are.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/claude-agent-sdk:1.0.0 - v1.0.0
Recipe
bland
The Bland voice-agent platform as a reference realization for local development: no service is defined for it and nothing is provisioned. The compiler carries the platform's API base address and the pathway coordinates to consumers through the injection contract, so an application developed locally reaches the same platform it will reach in a cluster. This recipe pins no artifact. It names no image and no chart — the deployment calls a platform its own operator contracts for directly, and the wiring is the whole of what is emitted. No credential is delivered here, and none can be: with no service defined, a credential attached to either of the component's slots would reach no emitted file, so the build halts naming the slot. On compose the credential belongs to the component that calls the platform, attached there under the environment-variable name that component reads.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/bland:1.0.0 - v1.0.0
Recipe
chroma
Chroma for local development via Docker Compose. Synthesized-mode Workload: the compose provider translates the Workload kind into a service entry running the official chromadb/chroma image, with a named volume at the server's data path. provides.host is the bare service name, resolved on the compose network. The server checks no credential, so the Recipe publishes no port on the host: the service is reachable from the other services of the stack and from nothing outside it. Reach for this realization for local development; on Kubernetes the component is realized by the Helm Recipe of the same name, which synthesizes a StatefulSet and needs nothing installed in the target cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/chroma:1.0.0 - v1.0.0
Recipe
bedrock
Amazon Bedrock's managed model API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied regional endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/bedrock:1.0.0 - v1.0.0
Recipe
bedrock-embedding
Amazon Bedrock's managed embedding API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied regional endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: consumers authenticate with ambient AWS IAM (for example EKS IRSA) or inject their own credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/bedrock-embedding:1.0.0 - v1.0.0
Recipe
bedrock-agent
An Amazon Bedrock agent as a reference realization for local development, so no service runs. The compiler carries the operator-supplied regional endpoint and the agent name to consumers through the injection contract, and delivers no credential — consumers authenticate with their own AWS credentials.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/bedrock-agent:1.0.0 - v1.0.0
Recipe
azure-openai
Azure OpenAI's managed chat API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer. Reach for it where the Azure OpenAI resource already exists outside the Release; reach for terraform/azure-openai where the Release is to create the resource and its model deployment.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/azure-openai:1.0.0 - v1.0.0
Recipe
azure-openai-embedding
Azure OpenAI's managed embedding API as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer. Reach for it where the Azure OpenAI resource already exists outside the Release; reach for terraform/azure-openai-embedding where the Release is to create the resource and its model deployment.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/azure-openai-embedding:1.0.0 - v1.0.0
Recipe
autogpt
An AutoGPT agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as AUTOGPT_AI_NAME, AUTOGPT_AI_ROLE, AUTOGPT_AI_GOALS (as JSON text), AUTOGPT_CONTINUOUS_MODE and AUTOGPT_CONTINUOUS_LIMIT (while set) by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/autogpt:1.0.0 - v1.0.0
Recipe
azure-api-management
References an API already served by an Azure API Management instance, as a reference realization for local development: no service runs and nothing in Azure is created or changed. The compiler carries the API's address to consumers through the injection contract. Reach for it where the API is configured outside the Release; reach for the terraform Recipe where the instance exists and the Release is to configure the API. It carries no Feature, because every control the gateway keeps is a policy in the API's own configuration, which this realization does not write, so the build halts on any attachment. No credential is delivered: a consumer that needs one declares it as `credential` on the socket that reaches the gateway and delivers it from its own container.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/azure-api-management:1.0.0 - v1.0.0
Recipe
azure-managed
An Azure managed online endpoint as a reference realization: no compute runs in-cluster. The compiler carries the operator-supplied endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/azure-managed:1.0.0 - v1.0.0
Recipe
autogen
An AutoGen agent application for local development via Docker Compose. Synthesized-mode Workload: image and environment are translated from the Workload kind by the compose provider. The model endpoint published by the component wired into the `model` socket is rendered as OPENAI_BASE_URL, the environment variable the OpenAI-compatible client SDK reads; a model API key attached on the `model` connection arrives on the inject pipe as OPENAI_API_KEY. The model name is delivered as OPENAI_MODEL by default; a connection writing an `inject:` block receives the names it dictates in its place, as variables or as files mounted at the paths it names, while OPENAI_BASE_URL and the API key stay. The agent's own values reach the container as AUTOGEN_AGENT_NAME and AUTOGEN_SYSTEM_MESSAGE by default. The component's own `inject:` block delivers its entries as a connection's are, and one reading a value alone replaces that value's default name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/autogen:1.0.0 - v1.0.0
Recipe
anthropic
Anthropic's managed Claude API as a reference realization for local development: no service runs. The compiler carries the endpoint to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/anthropic:1.0.0 - v1.0.0
Recipe
anthropic-agent
A Claude assistant on Anthropic's managed API as a reference realization for local development, so no service runs. The compiler carries the endpoint and the model string to consumers through the injection contract. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/anthropic-agent:1.0.0 - v1.0.0
Recipe
ai21
AI21 Labs' managed Jamba API as a reference realization: no compute runs in-cluster. The compiler carries the endpoint to consumers through the injection contract; no service runs. No credential is delivered by this recipe: a consumer that needs one declares it as `credential` on the socket that reaches this endpoint and delivers it from its own container, because the environment-variable name belongs to the consumer.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/ai21:1.0.0 - v1.0.0
Recipe
agentgateway
agentgateway for local development via Docker Compose, synthesized from the Workload kind by the compose provider: one service running the project's image, pinned by digest, from one configuration file mounted from a compose config. Nothing is published to the host; provides.host is the compose service name. Each model wired to the models socket is one entry of that file, named by the connection's alias and sent to the endpoint and model name the model publishes. A connection's credential is the value it publishes as AI_PROXY_<ALIAS>_API_KEY on the inject pipe, written into the service's environment as it was delivered — a Compose variable reference such as ${OPENAI_API_KEY} lands as that reference, a literal as itself — and the entry names the variable, never the value. The file attaches no UI to the data-plane port. It realizes none of the AI Features the Component's labels admit, so attaching one halts the build naming its type, and it halts on infra.callers: this target holds no Secret to read the callers' keys from. Reach for it under Docker Compose; reach for `agentgateway-chart` on Kubernetes. The image carries no shell or client to run a check from, so the service declares no healthcheck. The gateway's metrics are always on, on port 15020, which checks no caller and which any peer on the compose network reaches. With `values.metrics_enabled` true their address is published as provides.metrics_endpoint; false, that key is empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/agentgateway:1.0.0 - v1.0.0
Recipe
a2a-remote
An agent another party runs over the Agent2Agent (A2A) protocol, reached at the address infra.endpoint names, as a reference realization for local development via Docker Compose: no service is defined for it and nothing runs. The address reaches a consuming service through the injection context rather than through a service definition, because compose has no counterpart to an in-cluster alias. provides.endpoint is the address as written, protocol is its scheme, agent_card_url is where the agent's card is read, and credential_header is the request header the agent reads a caller's credential from, Authorization unless infra.credential_header names another. token_url, oidc_discovery_url and scopes are where and for what a caller obtains a token under the card's OAuth 2.0 or OpenID Connect scheme, as infra names them, and empty where it names none. No credential is held here: a consumer that needs one holds it on its own connection to the agent and presents it under that header, because the credential belongs to the consumer rather than to the address. infra.endpoint is the address without any part that carries a key, at least a scheme and a host. An address carrying a user name or password is refused, for infra.endpoint and infra.agent_card_url alike. infra.token_url and infra.oidc_discovery_url, where set, are each refused unless they are an https URL naming a host with no user name or password, since OAuth 2.0 sends a client secret only over TLS. Reach for this realization under Docker Compose; helm/a2a-remote is the same reference on Kubernetes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/a2a-remote:1.0.0 - v1.0.0
Recipe
adminer
Adminer for local development via Docker Compose. Synthesized-mode Workload: image, environment, port, and the exposed service are translated from the Workload kind by the compose provider. The database host resolved from the `databases` uses-edge arrives on .context.inject.env and is rendered verbatim into the service env.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/compose/adminer:1.0.0 - v1.0.0
Recipe
worker-ecs
The generic background-worker primitive as one Amazon ECS service on Fargate, running infra.replicas tasks of the Release's image, rendered as a CloudFormation fragment the provider assembles into the boundary's stack. Reach for it where the target is an AWS account with an ECS cluster and a VPC; reach for helm/worker on a Kubernetes cluster. The cluster, subnets, security groups and execution role come from the operator, under infra.overrides, and each is required; a task role, a public IP and a log group are optional, and nothing is emitted for one left unset. Only a task definition and a service are emitted — no cluster, role, security group, log group or load balancer, each of which the operator already runs — and the container maps no port, so nothing connects to the worker. No container health check is declared, because ECS runs one inside an image this Recipe does not name. Each env entry an edge's inject block, or the component's own, dictates becomes an entry of the container's environment under that name, its value written whole, so a neighbour's address arrives as an Fn::GetAtt in the same stack and as an Fn::ImportValue from another; concurrency and queue_depth_hint arrive by default as WORKER_CONCURRENCY and WORKER_QUEUE_DEPTH_HINT, which no edge may dictate while they are the defaults. File, arg and other entries are refused by name: Fargate writes no file, and no command line is rendered. The task runs at the smallest Fargate size meeting both infra.compute.cpu and infra.compute.memory, so the Component's defaults run at 0.25 vCPU and 512 MiB. It is written as a fragment, as the catalog writes its CloudFormation Recipes, rather than sourced from a published template.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/worker-ecs:1.0.0 - v1.0.0
Recipe
valkey-elasticache
Valkey on Amazon ElastiCache, rendered as a CloudFormation fragment: one AWS::ElastiCache::ReplicationGroup with cluster mode disabled and one node, encrypted at rest, and a parameter group beside it where maxmemory_policy or databases is set. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. Reach for this realization where the cache is an ElastiCache service in a CloudFormation stack; reach for terraform/valkey-elasticache on terraform, and valkey or valkey-chart on helm, or valkey on compose, where it runs as a container. host publishes a CloudFormation reference to the primary endpoint's address, which resolves inside the stack and through its export, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the cache instead. infra.overrides.CacheSubnetGroupName is required. A value ElastiCache cannot apply — maxmemory, appendonly, persistence — halts the build naming the key; compute and storage are not read. With infra.client_tls true the service encrypts in transit and protocol is valkey-tls; an attached password becomes the AUTH token through a Secrets Manager dynamic reference, and needs client_tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/valkey-elasticache:1.0.0 - v1.0.0
Recipe
s3-object-store-amazon-s3
An S3-compatible object store on Amazon S3, rendered as a CloudFormation fragment: one AWS::S3::Bucket, a general purpose bucket in the account and Region the stack is deployed to, and an AWS::S3::BucketPolicy beside it where infra.overrides.policy supplies one. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. Reach for it where a consumer speaks the S3 API to a bucket in a CloudFormation stack; for terraform/s3-object-store-amazon-s3 through Terraform; for terraform/s3-object-store-gcs on Google Cloud; and for helm/s3-object-store or compose/s3-object-store where the Release runs the store itself, on SeaweedFS, in a cluster or on a Compose host. Every public access is blocked: all four S3 Block Public Access settings are on and ACLs are disabled, and no key a Release sets changes either, so S3 rejects a public ACL or a public bucket policy. S3 takes no key from a bucket: each consumer brings the access key of an IAM principal, created outside the build, on its own connection, and an attachment on access_key_id or secret_access_key halts naming the slot. The fragment declares no IAM resource. infra.region is required, and a template rule stops the stack unless it is deployed to that Region. It publishes the bucket and the Region as references, the Regional endpoint https://s3.<region> under the partition's domain suffix, its host, port 443 and protocol s3. A bucket name S3 does not take halts naming values.bucket.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/s3-object-store-amazon-s3:1.0.0 - v1.0.0
Recipe
redis-elasticache
Redis OSS on Amazon ElastiCache, rendered as a CloudFormation fragment: one AWS::ElastiCache::ReplicationGroup with cluster mode disabled and one node, encrypted at rest, and a parameter group beside it where maxmemory_policy or databases is set. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. ElastiCache runs Redis OSS up to 7.1, so infra.engine_version takes 7.1, 7.0, 6.2, 6.0, 5.0.6 or 4.0.10, read off the start of the value; from 8 on it halts naming the key, the Component's default included, so the Release sets the version on this target. ElastiCache serves vector search on Valkey alone, so vector_search is empty and a gateway's semantic cache halts on this store. Reach for this realization where the cache is an ElastiCache service in a CloudFormation stack; reach for terraform/redis-elasticache on terraform, cloudformation/valkey-elasticache for a newer engine, or redis on helm or compose where it runs as a container. host is a reference to the primary endpoint's address, or infra.overrides.host while set; port is infra.port. infra.overrides.CacheSubnetGroupName is required. maxmemory, appendonly, persistence, infra.overrides.image, and databases below 6.0 halt naming the key; compute and storage are not read. With infra.client_tls the service encrypts in transit and protocol is redis-tls; an attached password becomes the AUTH token through a Secrets Manager dynamic reference, and needs client_tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/redis-elasticache:1.0.0 - v1.0.0
Recipe
postgres-rds
PostgreSQL via AWS RDS, rendered as a CloudFormation fragment. The artifact emits only the RDS DBInstance resource under a Resources block; the cloudformation provider stamps the stack envelope and projects this recipe's provides entries into the stack Outputs when it assembles the boundary's fragments into one stack template. The managed-service realization: the operator's ha sizing maps onto the instance's native MultiAZ setting, an attached core/backup onto its automated backups, and RDS owns the master password, in AWS Secrets Manager: a credential on the password slot other than core/generated-secret, or any on app_user_password, halts the build. host publishes a CloudFormation reference to the instance's address, which resolves inside the stack and through its export; port publishes the port infra.port configures, which a consumer on any target can use. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead. The Recipe emits nothing for it — no DNS record, no alias — so pointing the name at the instance is the operator's. A value RDS cannot apply — an encoding, a locale, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu and compute.memory are not read, because the instance class sizes the instance. On terraform the same component is also realized by terraform/postgres-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/postgres-rds:1.0.0 - v1.0.0
Recipe
opensearch-service
OpenSearch on Amazon OpenSearch Service, rendered as a CloudFormation fragment: one AWS::OpenSearchService::Domain of one data node in the operator's VPC. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. Reach for this realization where the search engine is an OpenSearch Service domain in a CloudFormation stack; reach for terraform/opensearch-service on terraform, and helm/opensearch-chart or compose/opensearch where it runs as a container. The domain serves HTTPS alone on 443, with fine-grained access control and one account, admin, whose password is the secret core/random-password creates in the stack, handed over as a Secrets Manager dynamic reference, so the rendered template carries none; with no password attached, or another credential, the build halts naming the slot. The attachment must set values.symbols true, so the password holds every character class. host publishes a reference to the domain's endpoint, port 443, protocol https, endpoint https://<host> and user admin. infra.overrides.host publishes the operator's DNS name instead and names the domain's custom endpoint, presenting the certificate infra.overrides.CustomEndpointCertificateArn names, so a client checking the certificate reaches the domain by it; either key alone halts, as does a name the endpoint cannot take, such as an IP literal. infra.overrides.SubnetIds is required. infra.engine_version must name a version the service offers and infra.storage.size 10Gi or more, so both Component defaults halt here; an infra.port other than 9200 halts, and so do an io1 volume and any instance or volume type the service cannot serve this domain on.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/opensearch-service:1.0.0 - v1.0.0
Recipe
rabbitmq-amazon-mq
RabbitMQ on Amazon MQ, rendered as a CloudFormation fragment: one AWS::AmazonMQ::Broker in the operator's VPC. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. Reach for it where the broker is an Amazon MQ broker in a CloudFormation stack; reach for terraform/rabbitmq-amazon-mq on terraform, and rabbitmq or rabbitmq-cluster on helm, or rabbitmq on compose, where it runs as a container. Amazon MQ serves AMQP over TLS alone, so port is 5671 and protocol amqps, which is what tells a consumer to connect over TLS. host publishes the host of the broker's AMQP endpoint, read out of the endpoint URL when the stack deploys, inside the stack and through its export; infra.overrides.host publishes the operator's DNS name instead, which a client checking the certificate cannot use, since the certificate names the broker's own endpoint. The password is required: with nothing attached the build halts naming the slot. It is the secret core/random-password creates in the stack, handed over as a Secrets Manager dynamic reference, so the rendered template carries none, and the attachment must leave values.symbols off and keep values.length at 12 or more, as the service's password rules need. Any other credential halts. infra.overrides.SubnetIds is required. infra.engine_version must name a version the service offers, so the Component's default halts here; the container-sizing, port and controller keys, and values.metrics_enabled, halt away from the Component's defaults.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/rabbitmq-amazon-mq:1.0.0 - v1.0.0
Recipe
mysql-rds
MySQL via AWS RDS, rendered as a CloudFormation fragment. The artifact emits only the RDS DBInstance resource under a Resources block; the cloudformation provider stamps the stack envelope and projects this recipe's provides entries into the stack Outputs when it assembles the boundary's fragments into one stack template. host publishes a CloudFormation reference to the instance's address, which resolves inside the stack and through its export; port publishes the port infra.port configures, which a consumer on any target can use. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead. The Recipe emits nothing for it — no DNS record, no alias — so pointing the name at the instance is the operator's. A consumer in the same stack then receives the name too, and CloudFormation no longer creates it after the instance: order the two with the boundaries' syncWave. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu, compute.memory, charset and collation are not read. On terraform the same component is also realized by terraform/mysql-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/mysql-rds:1.0.0 - v1.0.0
Recipe
mariadb-rds
MariaDB via AWS RDS, rendered as a CloudFormation fragment. The artifact emits only the RDS DBInstance resource under a Resources block; the cloudformation provider stamps the stack envelope and projects this recipe's provides entries into the stack Outputs when it assembles the boundary's fragments into one stack template. host publishes a CloudFormation reference to the instance's address, which resolves inside the stack and through its export; port publishes the port infra.port configures, which a consumer on any target can use. Set infra.overrides.host in the Release to the DNS name or IP literal the instance is reached at, and host publishes that name instead. The Recipe emits nothing for it — no DNS record, no alias — so pointing the name at the instance is the operator's. A consumer in the same stack then receives the name too, and CloudFormation no longer creates it after the instance: order the two with the boundaries' syncWave. RDS owns the master password, in AWS Secrets Manager: root_password takes core/generated-secret, and any other credential there or on user_password halts the build. A value RDS cannot apply — an application user, init scripts, a raw configuration file, read replicas — halts the build naming the key; compute.cpu, compute.memory, charset and collation are not read. On terraform the same component is also realized by terraform/mariadb-rds-terraform-aws-modules, the terraform-aws-modules project's published RDS module. Point-in-time recovery comes with the automated backups, so infra.point_in_time_recovery: false on core/backup halts by name.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/mariadb-rds:1.0.0 - v1.0.0
Recipe
memcached-elasticache
Memcached on Amazon ElastiCache, rendered as a CloudFormation fragment: one AWS::ElastiCache::CacheCluster of infra.replicas nodes, and a parameter group beside it where max_item_size is set. The cloudformation provider stamps the stack envelope and projects this Recipe's provides into the stack Outputs. Reach for this realization where the cache is an ElastiCache service in a CloudFormation stack; reach for terraform/memcached-elasticache on terraform, and memcached on helm or on compose, where it runs as a container. host publishes a CloudFormation reference to the cluster's configuration endpoint address, which resolves inside the stack and through its export, and port the configured infra.port. Set infra.overrides.host to publish the DNS name the operator points at the cache instead. infra.overrides.CacheSubnetGroupName is required, and infra.engine_version must name a full version ElastiCache offers, such as 1.6.22. A value ElastiCache cannot apply — memory_limit, max_connections, threads, a password — halts the build naming the key; compute is not read. With infra.client_tls true the service encrypts in transit and protocol is memcached-tls.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/memcached-elasticache:1.0.0 - v1.0.0
Recipe
db-reader
A database-consuming application rendered as a CloudFormation fragment. The artifact emits a single SSM parameter whose value is the database host the `database` uses-edge injects, stored as a whole scalar. When the database assembles into the same stack the value is an intra-stack Fn::GetAtt; when it assembles into another stack the provider replaces the injected value with a cross-stack reference. The provider stamps the stack envelope when it assembles the boundary.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/db-reader:1.0.0 - v1.0.0
Recipe
function-lambda
The generic event-driven compute primitive as one AWS Lambda function run from a container image, rendered as a CloudFormation fragment the provider assembles into the boundary's stack. Reach for it where the target is an AWS account; reach for helm/function-knative or helm/function-keda on Kubernetes. Only the function is emitted, with a published version holding provisioned concurrency for a floor: no role, log group, trigger, URL or event source mapping: what invokes it, its role and its log group are the operator's, the role required under infra.overrides. The image is in Amazon ECR, in the function's Region, and implements Lambda's runtime API — an AWS base image for Lambda, or a linux/amd64 image with a runtime interface client; one serving HTTP carries an adapter forwarding Lambda's invocations. An image outside ECR is refused by name. compute.memory is the function's memory, max_replicas its reserved concurrency, and a min_replicas above zero provisioned concurrency on a version published at each change, serving invocations of that version; as a replaced version holds its share until deleted, the floor is half the ceiling at most. port, compute.cpu, control_plane_present and scaler are not read; the comment above the assertions gives each reason. Each env entry an edge or the component's own inject block dictates becomes a variable of the function's environment, its value written whole; runtime arrives by default as FUNCTION_RUNTIME. A name Lambda reserves or cannot take, or FUNCTION_RUNTIME while it is that default, is refused by name, and so are file, arg and other entries. It is written as a fragment, as the catalog writes its CloudFormation Recipes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/recipe/cloudformation/function-lambda:1.0.0 - v1.0.0
MappingLibrary
GCSBucket
The module arguments, and the refusals, every Recipe provisioning one Google Cloud Storage bucket through the terraform-google-modules project's cloud-storage module writes alike, held once. A Recipe consumes it with `valuesMapping: { $import: GCSBucket#/mappings/ModuleArguments }`; the import replaces the whole `valuesMapping`, so the entry is the Recipe's complete argument list. The entry names the one bucket values.bucket names, exactly as written, places it in the operator's project and location, keeps public access prevention enforced and uniform bucket-level access on, and creates no IAM binding, service account or HMAC key; no key a Release sets reaches any of those. It refuses a bucket name Cloud Storage does not take, a missing project or location, and every infra.overrides key but project_id. What differs between importers is read from the locals each declares: the location, the name of the infra key it came from, and, where the Recipe writes them, the storage class and versioning. terraform/s3-object-store-gcs and terraform/gcs import it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/mappinglibrary/gcsbucket:1.0.0 - v1.0.0
MappingLibrary
SourcedChartValues
Reusable value mappings for a recipe that sources an upstream chart. A recipe consumes an entry with `valuesMapping: { $import: SourcedChartValues#/mappings/<Entry> }`; the version is supplied by the per-Release lockfile and no registry host enters the reference. The import replaces the whole `valuesMapping`, so each entry below is a complete value document rather than a fragment spliced into one. Each entry carries the two obligations a sourced recipe owes: the generated object names are pinned to the node's rendered name — the release name the helm provider gives the chart, which spells the component's name in the alphabet a Kubernetes name allows — so a value the component publishes through `provides` from the same `.context.rendered_name` names something the chart emits, and the image the chart runs is delivered explicitly rather than left to the chart's own default, so repeated renders are reproducible. The pin reads `.context.rendered_name` inside the imported body, which renders the release name on a `capability: helm` Recipe sourcing a chart, because that provider reports one for such a node. Under a provider that reports no such name, a reference inside an imported body renders empty rather than halting, so import these entries into helm Recipes only. The key names are the ones most charts use for these settings — `fullnameOverride`, `image`, `resources`, `nodeSelector`, `tolerations`, `persistence`. A chart that spells one of them differently, or that nests its workload under a named sub-component, needs its own mapping: adapt an entry here rather than importing it, since only the chart settles what belongs in its value document.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/mappinglibrary/sourcedchartvalues:1.0.0 - v1.0.0
Feature
example/storage-alarm
Raises an alarm when a datastore's free storage falls below the headroom the application needs. Its FeatureRecipe is component-scope: it emits one alarm per attached component, watching that component's own storage. The headroom is developer intent (values), because how much spare capacity an application needs before it is in trouble is a property of the application; where the alarm notifies and how long a dip has to persist before it fires are operator sizing (infra), because they describe the operator's on-call arrangements rather than the application. This is an example-namespace Feature: it exists to show how an attachment contributes resources to a target's output, and it is authored for the worked example that demonstrates it rather than as a portable standard-library capability.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/example/storage-alarm:1.0.0 - v1.0.0
Feature
example/db-binding
A database-binding feature: attach it to an application component to wire that component to a stateful-server database and deliver the connection. The feature carries the `databases` uses-edge to the target plus the delivery declaration that lands the database's published host on the consumer's environment as ADMINER_DEFAULT_SERVER. The binding is the vehicle for the edge — the application component consumes the database through it, so the connection wiring is declared once as a reusable attachment rather than inline on every consumer. It renders no resource of its own — its FeatureRecipe (db-binding) is a no-op. Like any feature instance it still resolves to a FeatureRecipe; its whole contribution is the bound edge and its injection.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/example/db-binding:1.0.0 - v1.0.0
Feature
core/upstream-key-pinning
Pins the public keys a gateway accepts from the services it forwards to. For each upstream the Release names, the gateway accepts the connection only when a certificate that upstream presents carries one of the keys listed for it, and refuses the connection otherwise. It attaches to a component standing in the request path that declares it can pin an upstream by its public key (dataplane.aggrega.ai/upstream-pinning: public-key). The gateway is the side that checks, and each entry names the upstream it checks by the host the gateway dials. Each key is a public key in PEM form, the key itself rather than a certificate carrying it, so a certificate renewed over the same key still matches and one issued over a new key is refused until that key is pinned. A pin is a control, not an assurance: it narrows which key an upstream may present and says nothing else about the connection. What a gateway does for an upstream no entry names is the product's own, and each realization states it; on some gateways, setting any pin stops certificate chains being verified for every upstream.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/upstream-key-pinning:1.0.0 - v1.0.0
Feature
core/target-selection
Declares which way an AI gateway chooses among several targets that can serve the same request. Its strategy is latency, which picks the target observed to answer quickest, or cost, which picks the target that is cheapest to call. It is the choice between targets rather than the routing itself: which targets are candidates comes from the gateway's own routing table, so a table that publishes each model alias once leaves nothing to choose between. The developer attaches it in the Aggregate, since the strategy is a value, and a Release binds its realization in the Manifest. On core/gateway/litellm, target-selection-litellm writes latency as the proxy's routing strategy and refuses cost by name: the proxy's price table does not name OpenAI's and Anthropic's models in the provider/model form the gateway's targets take, so it would choose by order. On core/gateway/kong-enterprise and core/gateway/traefik-hub, target-selection-kong-enterprise and target-selection-traefik-hub halt every attachment, since each of their model routes carries one target. On core/gateway/higress, target-selection-higress halts every attachment, since the gateway can measure a service's latency but cannot make its choice route the request, and chooses no target by cost. On core/gateway/agentgateway, target-selection-agentgateway halts every attachment, since the gateway takes the quickest target only when a random draw picks it, and chooses no target by cost. Each halts on any other gateway.
ОфіційнийЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/target-selection:1.0.0 - v1.0.0
Feature
core/traffic-shadowing
Sends a copy of live requests to a second service and discards that service's answers, so a candidate — a new model version, a rewritten service — can be watched under real traffic while every caller is still answered by the destination it was sent to. It acts at the gateway hop: it attaches to a component standing in the request path that declares it can mirror a request (dataplane.aggrega.ai/request-mirroring, whichever value), and that component copies the requests it forwards to the Service the Release names. The component needs to be able to copy a matching request to a second destination and drop that destination's response; that is all the whole stream asks for, so the default attaches wherever the capability is declared. Copying only a share of the requests asks for more: a component declaring the sampled fidelity, which offers a settable share. A realization for a component that copies every request refuses a share below 100, naming the fidelity that component declares, rather than copying more traffic than was asked for. Every input is the operator's, so the operator attaches it in the Release, on the mirroring component's entry under that component's features, and writes in infra the Service the copies go to and the share copied.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/traffic-shadowing:1.0.0 - v1.0.0
Feature
core/shared-config
A boundary-level shared-config feature. Its FeatureRecipe runs once per boundary (executionScope: boundary), emitting a ConfigMap of settings shared by every component in the boundary. The per-boundary tier is operator-set infra, so each boundary's ConfigMap carries its own resolved value.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/shared-config:1.0.0 - v1.0.0
Feature
core/semantic-caching
Lets an AI gateway answer a prompt from the replies it stored earlier when the prompt is close enough in meaning to one it answered before, rather than calling the model again. The gateway turns each prompt into a vector with embedding_model, serves a stored reply whose prompt is at least similarity_threshold alike, and keeps a reply for ttl_seconds. scope decides who shares a cache. With key, every caller presenting the same key to the gateway is served from one cache, so an end user behind that key can be served a reply generated for another, tool calls included. The default, end_user, keeps each end user the gateway identifies apart. A match on meaning suits single prompts: on multi-turn or agentic traffic consecutive turns read alike, so a reply to an earlier turn can be served for a later one. It does not meet core/data-masking on one component. The developer attaches it in the Aggregate, with its inputs in values; it takes no infra. core/gateway/litellm, azure-api-management, apisix and traefik-hub realize it, the last three keeping scope key alone, apisix and traefik-hub for the callers each Release declares; litellm, apisix and traefik-hub keep the replies in a core/datastore/redis wired to the gateway, litellm requiring its administrative key. On core/gateway/kong-enterprise and higress every attachment halts, naming the caller separation each cache lacks, and on agentgateway, naming the reply cache it lacks; on any component but these seven an attachment halts naming the type, rendering nothing.
ОфіційнийЗалежні: 8Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/semantic-caching:1.0.0 - v1.0.0
Feature
core/secret-provided
A secret value the developer delivers explicitly. Unlike core/secret-value (which the FeatureRecipe injects directly into the component pipe), this credential exposes its value as a `provides` key, so the developer dictates delivery with the attachment's own `inject:` block — the third inject placement, allowed only on a transparent featuresSchema slot. The feature.aggrega.ai/kind: secret label names what it produces; it is attached on the features: pillar like any other Feature. origin is written by the developer in values or by the operator in infra, exactly one of the two on each attachment; the build halts naming the attachment when both or neither carry it. Only an attachment the Aggregate declares can carry the inject: block that delivers the value, so on a connection, or on an attachment the Release alone makes, no inject block delivers it: use core/referenced-secret or core/secret-value there.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/secret-provided:1.0.0 - v1.0.0
Feature
core/retry-policy
Sets how many times an AI gateway may send a request before giving up. It is meant for the transient failure — a refused connection, a rate limit, a timeout — that succeeds on being sent again; which of those a gateway sends again is the gateway's own, and not every gateway sends all three again. The retries are the gateway's rather than the caller's, so a caller sees one answer whichever send produced it. The value is the total number of sends including the first, which is one more than the number of retries it buys. It brings no resource of its own: the gateway's Recipe writes the setting where the gateway reads one, into the routing configuration it already renders, a traffic policy the gateway's controller reconciles, or a hosted gateway's own configuration.
ОфіційнийЗалежні: 8Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/retry-policy:1.0.0 - v1.0.0
Feature
core/secret-value
A single secret value, delivered into the consuming component's inject pipe under a named key. Provider-agnostic: the Manifest maps it to a FeatureRecipe per target, which delivers it as an environment or chart value on helm (where k8s-secret-from-value also holds it in a Kubernetes Secret a Recipe can read by reference), an environment value in the Compose file, or a value a Terraform Recipe reads. origin is written by the developer in values or by the operator in infra, exactly one of the two on each attachment; the build halts naming the attachment when both or neither carry it.
ОфіційнийЗалежні: 4Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/secret-value:1.0.0 - v1.0.0
Feature
core/response-validation
Checks a model's reply passing through an AI gateway against a JSON Schema before the reply reaches the caller, so an application expecting structured output is not left parsing whatever came back. schema carries the schema itself, written as a mapping, and a reply that does not match it is refused rather than passed on. invert turns the check around and refuses a reply that does match. The schema is the attachment's, the same for every request the gateway serves. Checking a reply against a schema the caller sends with its own request is a different control, and not this one. The developer attaches it in the Aggregate, writing schema and invert in values; it takes no infra. No gateway in the standard library realizes it: an attachment halts the build, naming why on core/gateway/agentgateway, core/gateway/higress and core/gateway/traefik-hub and naming the type on any other gateway, and renders nothing.
ОфіційнийЗалежні: 3Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/response-validation:1.0.0 - v1.0.0
Feature
core/random-password
A credential the target generates when it deploys, so no one authors it and no default ships for every reader of a catalog to know. The build cannot make the password: its output is identical for identical input and never carries a secret's value. So the FeatureRecipe renders a request instead — a generator resource naming the secret resource it fills — and the target draws the value at deploy time: a controller in a Kubernetes cluster, the random provider during a Terraform apply, Secrets Manager when a CloudFormation stack is created. The rendered output names the generator and the secret resource, and carries no value. It publishes the same locator core/referenced-secret publishes, the secret resource's name and the data key within it, so a consuming component references a generated credential exactly as it references one that already exists, and a slot selecting the credential family admits it. Where several components share one credential, attach core/random-password on one of them and core/referenced-secret, bound to k8s-existing-secret, on the others, with its origin set to the name this attachment publishes; pin that name with the name value so both attachments can be written from one parameter. Attaching core/random-password on each of them, or once at a scope that reaches them all, instead renders one request per component, and each request draws a different password. Realized on helm (through the External Secrets Operator, or through the kubernetes-secret-generator controller), on terraform and on cloudformation. The compose target has no generator, so a Release binding this Feature there halts for want of a realization; supply the credential with core/secret-value instead.
ОфіційнийЗалежні: 4Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/random-password:1.0.0 - v1.0.0
Feature
core/referenced-secret
A credential held in a referenceable secret resource, whose identity it publishes for consumers to point at. The referenceable-secret member of the credential family: where core/secret-value delivers a value into the inject pipe, core/referenced-secret names a resource (a Kubernetes Secret, an AWS Secrets Manager secret, a Vault entry) — one the FeatureRecipe's output creates, or one that already exists — and exposes its locator via provides, so a consumer references it (e.g. a Kubernetes secretKeyRef) instead of re-deriving the name. Provider-neutral: the Manifest maps it per target. origin is written by the developer in values or by the operator in infra, exactly one of the two on each attachment; the build halts naming the attachment when both or neither carry it. So a Release can attach it to a connection it wires, or to a component, and name the secret itself.
ОфіційнийЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/referenced-secret:1.0.0 - v1.0.0
Feature
core/referenced-secret-set
A credential that materializes a single multi-key secret resource and publishes that resource's identity for consumers to point at. The multi-key member of the referenceable-secret family: where core/referenced-secret produces a resource around one caller-named data key and publishes both the resource locator and that key, this Feature carries a caller-supplied set of data keys in one resource and publishes the locator only. The consumer already knows the fixed key names its target reads (a Helm chart that expects a Secret with several named keys), so it references the resource by name and lets its target read the keys it already knows. Provider-neutral: the Manifest maps it to the appropriate FeatureRecipe per target. entries is written by the developer in values or by the operator in infra, exactly one of the two on each attachment; the build halts naming the attachment when both or neither carry it.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/referenced-secret-set:1.0.0 - v1.0.0
Feature
core/public-ingress
Publishes a server component at a hostname outside the deployment. Its FeatureRecipe (component-scope) emits one ingress resource per attached component, routing the host to the Service the component's own Recipe emits. The developer attaches it in the Aggregate and writes the request path in values, because which requests an application serves is part of what the application is. The operator sizes the same attachment on the Release's entry for the component, writing in infra the hostname, which is the environment's name for the component, and the cluster knobs, which describe the cluster the Release lands in rather than the application: which Service and named port to route to, the ingress class that reconciles the resource, and the certificate the host is served with over HTTPS (tls_secret). The Service defaults to the primary Service a synthesized workload emits, so a Release whose Recipe sources an upstream chart naming its Service otherwise sets service_name to that name. The certificate is referenced, never issued: the Secret already holds it — created by the operator, or written by a certificate controller such as the one core/security/cert-manager installs, once a Certificate resource in the cluster asks it to — and this Feature emits neither the Secret nor that request.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/public-ingress:1.0.0 - v1.0.0
Feature
core/prometheus-exporter
Marks a component as a Prometheus scrape target. Its FeatureRecipe (component-scope) emits one ServiceMonitor per attached component, a custom resource the Prometheus Operator reads, so a cluster not running that operator scrapes nothing through it. The scrape knobs — interval, named port, and metrics path — are all operator sizing (infra): monitoring is operational configuration, set by the platform, not application intent. Designed to be attached in bulk via a Manifest features.annotations binding — one annotation on a boundary expands to every contained component, routing the scrape knobs into each component's feature infra — as well as by a per-component features: block.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/prometheus-exporter:1.0.0 - v1.0.0
Feature
core/prompt-deny-patterns
Refuses a model request whose text matches one of the regular expressions the attachment lists, before the request reaches a model. It acts at the gateway hop: it attaches to a component standing in the request path that declares it can match the model payload it routes against a list of patterns (dataplane.aggrega.ai/prompt-guard: pattern-list), and that component turns a matching request away instead of forwarding it. The control is a literal match and is only as good as its list: it refuses text an expression matches, and it does not catch a paraphrase, the same request in another language, text disguised by an encoding, text split across messages, tool-call arguments, images or anything in the model's reply. A request that does not pass through the gateway is not seen at all. Attaching it therefore does not establish that a payload can be trusted and does not stop prompt injection; judging whether a request tries to override the instructions a model was given is a different control, declared as core/prompt-injection-shield. Which messages a gateway reads, whether letter case counts, and what a refused caller receives are the gateway's, and each realization states them for its own.
ОфіційнийЗалежні: 7Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/prompt-deny-patterns:1.0.0 - v1.0.0
Feature
core/prompt-injection-shield
Declares that the model requests a gateway forwards are checked for attempts to override the instructions a model was given, before the model is called, at one of two sensitivities: strict refuses more such attempts, and more legitimate requests with them. The check is a probabilistic classifier's verdict: it misses some attacks, reworded, encoded or respelled ones among them, refuses some legitimate requests, and is known to be evadable, so it is one control beside others. It does not judge the model's reply, tool-call arguments or images. A realization either sends the requests through a guard wired to the gateway's guard socket, which checks each one and forwards what it admits to the model, or turns on a check the gateway runs itself. core/gateway/litellm and core/gateway/nginx realize it with core/security/nemo-guardrails as the guard, each halting the build with no guard wired or unless the guard's threshold is the one the sensitivity names. core/gateway/azure-api-management realizes it as Azure's prompt shield, which is on or off, so it keeps standard alone. What each check reads, and what the gateway still serves, is stated on each realization. On a gateway with no realization, attaching it halts the build naming the type.
ОфіційнийЗалежні: 7Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/prompt-injection-shield:1.0.0 - v1.0.0
Feature
core/period-budget
Bounds what the traffic through an AI gateway may consume over a period. Once what the gateway has counted over period exceeds limit, it refuses further model requests until the count falls back under the limit. unit names what is counted, a cost in US dollars or tokens, and scope names whose traffic one limit covers: everything the gateway serves, each caller apart, or each model apart. The developer attaches it in the Aggregate, writing unit and scope in values; the operator sizes that attachment in the Release, writing limit and period in infra under the same attachment name on the Release's entry for the component. It attaches to a component in the request path that declares it can keep such a budget (dataplane.aggrega.ai/period-budget, whichever value). Each realization says how its gateway counts a period, and refuses a unit, a scope or a period its gateway cannot keep, naming the value asked for and what the gateway offers, rather than accepting it and counting something else. Each realization writes the budget where its gateway reads one: into the gateway's own configuration, or as a resource the gateway's controller reconciles.
ОфіційнийЗалежні: 9Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/period-budget:1.0.0 - v1.0.0
Feature
core/k8s-dns-resolver
Caller-side cross-namespace host qualification for Kubernetes. When a consumer reaches a target deployed in a different namespace, the in-cluster DNS name it must dial is the target's bare service name qualified with the target's namespace (service.namespace). This feature, attached to the consumer, rewrites the consumer's own injected host to that qualified form — leaving the target's published host untouched, so the same target stays reachable by same-namespace callers under its bare name. The qualification is derived from the target's resolved placement, never hardcoded.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/k8s-dns-resolver:1.0.0 - v1.0.0
Feature
core/manual-trigger
Lets a core/compute/task with no schedule be deployed without starting, for someone to start each run by hand, as a backup on request or a one-off data fix is. Attached to a task with no schedule, no run starts when the Release deploys; each run is one someone starts on the target. Attached to a task with a schedule, the task runs on its schedule exactly as it would without the Feature, and can also be started by hand. Running on demand is part of what the task is, so the developer attaches it in the Aggregate, under the task's features. It takes no values and no infra, so the operator writes nothing for it in the Release. It applies to components of type task only. No script is written into the build output: a run is started with the target's own command, which is the operator's to run. On Kubernetes that is kubectl create job <run-name> --from=cronjob/<cronjob>, the CronJob named as the task is rendered. On Docker Compose it is docker compose run --rm <service>, the service named after the task, which docker compose up leaves stopped. On Amazon ECS it is aws ecs run-task, given the cluster, the task definition and the network configuration the task's Recipe reads from infra.overrides. Each target realizes the Feature through a FeatureRecipe the Manifest binds to it, and a Manifest binding none halts the build, naming the attachment.
ОфіційнийЗалежні: 4Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/manual-trigger:1.0.0 - v1.0.0
Feature
core/http-route
Publishes a server component on a gateway programmed through the Gateway API. Its FeatureRecipe (component-scope) emits one HTTP route per attached component, naming the gateway the route attaches to as its parent and the Service the component's own Recipe emits as its backend. The developer attaches it in the Aggregate and writes the request path in values, because which requests an application serves is part of what the application is. The operator sizes the same attachment on the Release's entry for the component, writing in infra the hostname, which is the environment's name for the component, and the cluster knobs, which describe the cluster the Release lands in rather than the application: which gateway and listener to attach to, and which Service and port to forward to. Left without a hostname, the route carries none and serves every hostname the gateway's listener accepts, with the path match separating it from the other routes on that listener. The Service defaults to the primary Service a synthesized workload emits, so a Release whose Recipe sources an upstream chart naming its Service otherwise sets service_name to that name. A route names its backend port by number rather than by name, and an attachment cannot read the port the component it is attached to was given, so service_port carries no default and the build halts until the Release supplies it.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/http-route:1.0.0 - v1.0.0
Feature
core/generated-secret
A credential whose material the target engine generates and manages. The Release declares that a component's credential slot is delegated to the engine: no value is authored on the attachment, none travels the inject pipe, and no secret resource is rendered. It is the engine-managed peer of the developer-supplied credential Features — it carries the same feature.aggrega.ai/kind: secret label and binds the same slot — so a credential slot is always bound and what varies between Releases is what binds it. Attach it where the target provisions the credential itself: a managed cloud database, for example, whose Recipe sets the provider's own managed-password directive so the provider generates and rotates it, or a controller-driven Recipe that hands the credential to its controller, which generates it into a Secret it creates and owns. That choice lives in the component's Recipe, never here, which is why this Feature declares no values and its FeatureRecipes render nothing. Bind it only on a target the catalog ships a FeatureRecipe for — the standard library ships the terraform, cloudformation and helm capabilities, matching the Recipes that leave the credential to the engine.
ОфіційнийЗалежні: 3Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/generated-secret:1.0.0 - v1.0.0
Feature
core/fallback-chain
Sends a request whose model fails on to other models, in an order the attachment names. Attached to an AI gateway that routes on the model named in a request, it takes chains, whose entries each name a model the gateway publishes and, in order, the models a failed request for it falls back to, tried in turn until one answers; and max_hops, which bounds how far along the chain a request is passed. Every name in chains is a name the gateway publishes. The build halts on a model listed among its own fallbacks and on a model given two entries, and a realization halts on a name its gateway does not publish. A request falls back once the gateway has given up on the model it asked for, after the gateway's own retries of that model, whose count core/retry-policy sets. It brings no resource of its own: the chain and its bound are written into the routing configuration the gateway's Recipe already renders.
ОфіційнийЗалежні: 6Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/fallback-chain:1.0.0 - v1.0.0
Feature
core/gcp-workload-identity
Binds a workload to a Google Cloud service account through GKE Workload Identity. Its FeatureRecipe emits one Kubernetes ServiceAccount annotated with the Google service-account email; the pod runs under that ServiceAccount and impersonates the Google service account at runtime, so no static key material is stored. It fills a workload's identity slot: the workload's Component declares the need, and the operator attaches this Feature in the Release, on the Release's entry for the component. The Google service-account email and the ServiceAccount name are operator sizing (infra) — which identity a workload runs under is operational configuration. Emitting the ServiceAccount and binding the pod are two explicit steps; attaching this Feature performs the first. The consumer's Recipe performs the second when it reads the published ServiceAccount name; otherwise set the aggrega.ai/service-account annotation on the consumer to that name. The annotation is honoured on the synthesized render path; a consumer rendered from an external chart via `source:` does not honour it, and the binding has to be expressed through that chart's own inputs instead.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/gcp-workload-identity:1.0.0 - v1.0.0
Feature
core/egress-allowlist
Lets the component it is attached to connect only to the components it is wired to: the targets of its uses edges, each on the ports its pods declare, plus its own pods, the cluster's DNS and the address ranges the attachment names. It keeps, say, an agent platform reaching only the models and tools its Aggregate wires it to. It restricts where a connection may go, not what it carries. The cluster's network enforces the rule, and only where its network plugin implements network policy, so the Release affirms it. A target the cluster does not run, such as a hosted model API, is reached only through the address ranges the attachment lists for it, and a range admits those addresses on any port for the whole component, not for that connection alone. Attached by the operator in the Release, on the component it restricts, under components.<id>.features. Its destinations are the uses edges the Aggregate already declares, so a connection the Aggregate does not declare as an edge can be refused by the cluster wherever a Release attaches it. Realized on helm: one network policy per workload the component renders, in that workload's namespace. The build halts, naming the connection, where a target renders no pods and no range is listed for it, or where a target declares no port; and, as core/caller-allowlist does, where a workload is emitted into no namespace or carries neither a controller selector nor the build's identity labels. On compose the build halts, naming this Feature: compose has no network policy.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/egress-allowlist:1.0.0 - v1.0.0
Feature
core/data-masking
Replaces sensitive text in the chat and completion traffic an AI gateway routes with a placeholder before it travels past the gateway: in requests before they reach the model provider, in replies before they reach the caller, or both, as scope says. It attaches to a gateway declaring dataplane.aggrega.ai/content-masking and masks what the attachment lists, regular expressions under patterns and, under entities, entity types a recognizer service detects, in the text of the conversation's messages; which roles, turns and fields count as that text each realization states. An attachment listing nothing to mask halts the build: no list is assumed. action redact replaces a match for good; restore puts the original back in the reply, where a realization can do so reliably; refuse lets no text carrying a match travel on. An unreachable or erring recognizer fails the request. By default, text a realization does not mask is refused rather than forwarded, and what a masked reply carries unread is withheld; allow_unmasked_text and allow_unmasked_reply_parts relax that. It shares no component with core/semantic-caching nor a gateway with values.log_masking. Its values refuse display_name, description, severity and as. Realized on core/gateway/litellm, kong-enterprise, agentgateway, higress and traefik-hub, each refusing restore. Masking alone falls short: text no pattern matches and no recognizer detects passes; the gateway and its recognizers see the original; tool definitions, response formats, stop sequences, reasoning a client sends back and the names of JSON keys pass as sent; the gateway's log and a logging integration can hold unmasked text; and traffic not crossing the gateway is untouched.
ОфіційнийЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/data-masking:1.0.0 - v1.0.0
Feature
core/context-propagation
Carries the original caller's identity across the gateway hop, so the service behind a gateway can decide on who called rather than on the gateway that relayed the call. It attaches to a component standing in the request path that declares it can convey a caller's identity (dataplane.aggrega.ai/caller-identity, whichever value), and that component passes the caller's identity along on each request it forwards. propagation_mode names which identity travels, and the two modes are two different identities rather than two strengths of one: client_certificate forwards details of the TLS certificate the client presented, which the component first verifies against a certificate authority the Release names; injected_header forwards headers the component composes after authenticating the caller against its own key store. A component implements one mechanism, declared as the value of its caller-identity label, and a realization refuses a mode its component does not implement, naming the mode asked for and the mechanism the component declares. destinations names which of the places the component forwards to receive the identity, in the component's own vocabulary — model aliases on an AI gateway, routes on an HTTP gateway — and a realization whose component cannot limit the identity that way refuses the attachment when it is set, rather than sending the identity everywhere. The Feature covers the hop from the gateway to what the gateway forwards to, and no further: identity between two workloads that call each other directly, with no gateway between them, is checked at the receiving workload, which is a service mesh's hop rather than a gateway's, and attaching this Feature does not protect it.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/context-propagation:1.0.0 - v1.0.0
Feature
core/caller-allowlist
Lets only the components declared to call the component it is attached to connect to it: those whose uses edges point at it, plus the component's own pods and the sources the attachment names, on the ports its pods declare. It keeps, say, a model server reachable only through the gateway in front of it. It restricts who may connect, not what a connection carries. The cluster's network enforces the rule, and only where its network plugin implements network policy. The build writes the policy and cannot know whether it will be enforced, so the Release affirms it. Traffic from a source outside the Aggregate, such as an ingress controller or a monitoring agent, is admitted only when the attachment names it. A chart that runs several workloads is protected as one, and pods a controller creates at runtime are outside the rule. Attached by the operator in the Release, on the component it protects, under components.<id>.features. Its callers are the uses edges the Aggregate already declares, so a call the Aggregate does not declare as an edge can be refused by the cluster wherever a Release attaches it. Realized on helm: one network policy per workload the component renders, in that workload's namespace, selecting pods by a selector carrying no version. The build halts, naming the component, where it or a caller renders no pods, emits a workload into no namespace or with neither a controller selector nor the build's identity labels (a chart's Job, say), or where it declares no port. Compose cannot express the rule, so binding this Feature there halts for want of a realization.
ОфіційнийЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/caller-allowlist:1.0.0 - v1.0.0
Feature
core/compose-bridge
Lets components sitting on different networks of one local compose file reach one another. On the local target a component joins the network of its compose file, or the one its boundary names with the aggrega.ai/namespace annotation: boundaries sharing one compose file without that annotation share one network, and the annotation is what separates them. An edge crossing between two networks has no route; attached at boundary scope, this renders one small proxy into the boundary's own compose file, joined to the networks of the crossing edges' ends, and rewrites each consumer's address so it reaches its target through that proxy. Only an edge whose two ends are assembled into that same file is relayed; a connection between two files is left as written. Every input is the operator's, so the operator attaches it in the Release, on the boundary holding the seam, under that boundary's features. Both the listening ports and the rewritten addresses are derived from the edges that actually cross, so no address is hand-authored: what a Release may set, in infra, is the base port a collision falls back to, and which published field is taken as the upstream address. It is a network-architecture concern rather than a per-component one, which is why it attaches to a boundary and gates on no component type.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/compose-bridge:1.0.0 - v1.0.0
Feature
core/concurrency-limit
Caps how many requests a gateway carries at once — the control that keeps a burst of callers from becoming a queue at the upstream. scope names what one cap counts: gateway, every request the gateway forwards, counted together; or model, each target the gateway sends requests to, counted apart. A target is each upstream entry the gateway routes to, so where one model name has several entries each entry is counted apart. The count is per running proxy, so where several replicas run each of them admits max_in_flight: size it against one replica, not against the whole. Every input is the operator's, so the operator attaches it in the Release, on the gateway's entry under that component's features, and sets max_in_flight in infra, with scope where the default, gateway, is not the count wanted. It attaches to a component in the request path that declares it can keep such a cap (dataplane.aggrega.ai/concurrency-cap, whichever value), and each realization refuses the scope its gateway does not count, naming the cap the gateway keeps. On core/gateway/nginx scope gateway is realized and a request past the cap is answered 503; scope model is refused, because NGINX keeps one total. On core/gateway/envoy-gateway scope model is realized, one cap for each wired model, and a request past it is answered 503; scope gateway is refused, because Envoy Gateway keeps a cap for each backend and none on their total. On core/gateway/litellm scope model is realized, one cap for each entry of its model map, and a request past it is answered 429 unless a fallback chain sends it on; scope gateway is refused, because the proxy keeps no total that holds.
ОфіційнийЗалежні: 4Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/concurrency-limit:1.0.0 - v1.0.0
Feature
core/aws-irsa
Binds a workload to an AWS IAM role through IAM Roles for Service Accounts (IRSA). Its FeatureRecipe emits one Kubernetes ServiceAccount annotated with the IAM role ARN; the pod runs under that ServiceAccount and the EKS OIDC provider mints role credentials at runtime, so no static credential material is stored. It fills a workload's identity slot: the workload's Component declares the need, and the operator attaches this Feature in the Release, on the Release's entry for the component. The role ARN and the ServiceAccount name are operator sizing (infra) — which identity a workload runs under is operational configuration. Emitting the ServiceAccount and binding the pod are two explicit steps; attaching this Feature performs the first. The consumer's Recipe performs the second when it reads the published ServiceAccount name; otherwise set the aggrega.ai/service-account annotation on the consumer to that name. The annotation is honoured on the synthesized render path; a consumer rendered from an external chart via `source:` does not honour it, and the binding has to be expressed through that chart's own inputs instead.
ОфіційнийЗалежні: 1Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/aws-irsa:1.0.0 - v1.0.0
Feature
core/backup
Has the datastore it is attached to backed up on a schedule, each backup kept for a number of days and then deleted. Every input is the operator's, who attaches it in the Release, under the component's features; the developer writes nothing for it. It applies to components of type stateful-server. On a managed database service, attaching it configures the service's own automated backup: when it starts and how many days each backup is kept. On AWS, the Manifest can also select AWS Backup, which backs up an Amazon RDS instance into a vault of its own under the IAM role infra.role_arn names, the instance keeping its own automated backups beside unless infra.keep_automated_backups is false. On Kubernetes, a backup product the Manifest selects writes to the bucket the store connection points at or, where it can, to a volume infra.volume sizes; an attachment naming both, or neither for such a product, halts. A backup on a volume sits in the same cluster as the data it protects, so it guards against a bad change or deletion but not against losing the cluster. A target refuses by name a value its product cannot keep, rather than running a backup other than the one written: a schedule it cannot express, a retention outside the days it keeps, a key, a role, a user, a store or a volume it has no use for. infra.point_in_time_recovery asks for restores to any moment of the retention window, with no single default: left out, each product keeps its own. It is on, refusing false, on Amazon RDS, Azure Database for PostgreSQL flexible server and CloudNativePG; on in Cloud SQL; off on AWS Backup until true; off, refusing true, on mariadb-operator, Velero, K8up, KubeStash and Kasten.
ОфіційнийЗалежні: 10Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/backup:1.0.0 - v1.0.0
Feature
core/audit-logging
Keeps a record of what an AI gateway served. While attached to a gateway that routes on the model named in a request, the gateway keeps a record of each model request it served, naming the model that answered and the tokens the reply reports. Each realization states where the record lands. By default the record also says what each request cost and names the caller the gateway verified, and a gateway that cannot keep either refuses the attachment at build. require_cost and require_verified_caller relax that. With require_verified_caller off, the record names whatever caller a request claims to be, or none, so anyone sending a request can have it recorded against someone else. retention_days has the gateway delete a record once it is that old. With none, the gateway itself deletes no record, and how long the place a record lands holds it is that place's own. Every input is the operator's, so the operator attaches it in the Release, on the gateway's entry under that component's features, and sets any of the three there in infra.
ОфіційнийЗалежні: 6Опубліковано 2026-11-10oci://hub.aggrega.ai/feature/core/audit-logging:1.0.0 - v1.0.0
FeatureRecipe
variable-secret
Terraform realization of core/secret-value, for the author-supplied credential model. The material originates in the Aggregate or the Release — the developer authors it in the attachment's values, an operator fulfils the parameter that origin reads, or an operator writes it in the attachment's infra — and this recipe delivers that value into the consuming node's inject pipe under the chosen key. Nothing is emitted from here; the delivery is the whole contribution, so the value lands only where a Terraform Recipe reads that key and renders it into its own configuration — as the value of an input variable, or directly as a resource argument. A Recipe that does not read the key receives nothing. The managed-database Recipes the standard library ships take no such value and halt the build naming the slot: the RDS ones set the provider's managed-password directive and take core/generated-secret on that slot instead, and the Cloud SQL and flexible-server ones take core/random-password. This is the peer of terraform/generated-secret, not its replacement, and the two serve opposite models. generated-secret is pure wiring for the model where the target engine provisions and rotates the credential itself: it binds the slot the component declares and carries no injectMapping, because there is no authored value to hand over. Here the material is authored, so there is a value to deliver and injectMapping is how it travels. Satisfying a credential slot and injecting a value are different acts — which is exactly why a pure-wiring recipe can bind a required slot while rendering nothing at all.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/variable-secret:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-azure-api-management
Realizes core/semantic-caching on core/gateway/azure-api-management's terraform Recipe; a Release binds it in its Manifest. It writes a cache lookup and a cache store into the API's policy and emits no file. embedding_model names an API Management backend for an embeddings model that the instance already holds and reaches with its managed identity; the instance also needs a Redis cache with RediSearch configured as its external cache. The build cannot see either, so the gateway's Recipe halts until the operator affirms that cache by setting infra.cache_serves_vector_search. A reply is kept for ttl_seconds, apart per API and per API Management subscription, so it honours scope key and refuses end_user by name: the API validates no end user's token, so the gateway has no verified end user to keep apart. Key needs the subscription key the API requires by default. similarity_threshold is written as score-threshold, 1 minus the threshold, reading Azure's score as a cosine distance: Microsoft gives the score's range and direction, lower requiring closer prompts, but not its measure. Microsoft warns a similar prompt can be served a reply that is wrong for it. That the policy applies as described is not observed by any build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/semantic-caching-azure-api-management:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-cloudflare-ai-gateway
Realizes core/retry-policy on core/gateway/cloudflare-ai-gateway's terraform Recipe; a Release binds it in its Manifest. It writes the gateway's own retry count into the gateway's resource and emits no file. max_attempts counts every send, the first included, while the gateway's retry count names the sends after the first, from 1 to 5, so the number written is one fewer than the attempts asked for and max_attempts runs from 2 to 6; any other value is refused by name. The retries are the gateway's, and the delay between them and its backoff are left as the gateway's own settings have them. It halts when the attachment is carried by any component other than core/gateway/cloudflare-ai-gateway, and when the gateway carries a second retry policy. That the gateway retries as described is Cloudflare's documented behaviour and no build observes it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/retry-policy-cloudflare-ai-gateway:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-azure-api-management
Realizes core/retry-policy on core/gateway/azure-api-management's terraform Recipe; a Release binds it in its Manifest. It writes Azure's retry around the forwarding, in the backend section of the API's policy, and emits no file. A request answered with HTTP 429 or a 5xx status is sent again, its body held so it can be resent, one second after the previous send, up to max_attempts sends in all. A refused or dropped connection, and a model that answers nothing within the forwarding's timeout, are not sent again: Azure ends the retry on such an error and hands it to the policy's on-error section. Azure retries 1 to 50 times after the first send, so max_attempts 1 writes no retry and above 51 is refused by name. The retry takes the place of the forwarding the API would inherit, so while it is attached a forwarding setting made at a broader scope of the instance, such as a timeout, does not apply to this API, which forwards with Azure's defaults. It halts on a second retry policy and on any component other than core/gateway/azure-api-management. That the retry applies as described is Microsoft's documented behaviour, and no build observes it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/retry-policy-azure-api-management:1.0.0 - v1.0.0
FeatureRecipe
random-password
Terraform realization of core/random-password through the random provider's random_password resource. The attachment renders a module of its own holding one random_password resource and one output, named values.key and marked sensitive, that carries the password out of the module; Terraform draws the password when the configuration is applied, so the rendered configuration carries no value. The value is kept in the Terraform plan and state, so the state is where it must be protected. The published provides are the attachment's module call and the output holding the password, which a consumer joins with a dot and wraps in an interpolation — for example ${module.db--password.value}. The reference resolves in the boundary's root configuration, so the consuming Recipe is a source: Recipe, whose valuesMapping becomes its module's arguments there; a Recipe that authors its own body cannot reach it, and for a host composed in from another Aggregate Terraform refuses the reference at plan time. The password holds at least one upper-case letter, one lower-case letter and one digit, and one punctuation character while values.symbols is on, as Secrets Manager draws for the cloudformation realization by default, so with values.symbols on a service requiring every character class accepts it. It is drawn once and kept in state; a change to values.length or values.symbols draws a new one. The name random-password is shared with the cloudformation realization, so a Manifest references this one as terraform/random-password:1.0.0; a bare random-password:1.0.0 matches more than one and resolves to none, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/random-password:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-azure-api-management
Realizes core/prompt-injection-shield on core/gateway/azure-api-management's terraform Recipe; a Release binds it in its Manifest. The Recipe writes Azure's content safety check into the API's policy with its prompt shield on, and it emits no file. Azure AI Content Safety checks each request for a user attack before the model is called, and the gateway answers one it detects with HTTP 403. The shield is on or off, so it keeps sensitivity standard and refuses strict by name. It runs at Microsoft's own sensitivity, which is not measured against the threshold standard means on the gateways that send requests through a guard. The check calls the instance's Azure AI Content Safety resource through the backend infra.content_safety_backend_id names, which the instance reaches with its managed identity; the build halts while that key is unset, and cannot see the backend. It halts on any component other than core/gateway/azure-api-management. That the check applies as described is Microsoft's documented behaviour, and no build observes it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/prompt-injection-shield-azure-api-management:1.0.0 - v1.0.0
FeatureRecipe
period-budget-cloudflare-ai-gateway
Realizes core/period-budget on core/gateway/cloudflare-ai-gateway's terraform Recipe; a Release binds it in its Manifest. It writes one spend-limit rule into the gateway's own resource and emits no file. The gateway counts what its traffic cost in US dollars, priced from the pricing it keeps per model, over a window that moves forward with time: the last period, written as a number of seconds. Once the count reaches limit it answers further model requests with HTTP 429 until the window has moved on. It records a request's cost once the request completes, so concurrent requests can take it past limit before the first refusal, and it counts only models whose pricing it knows. So it honours unit usd, scope gateway and a period of seconds, minutes, hours or days, and refuses tokens, a budget per caller or per model, and calendar months by name. It halts when the attachment is carried by any component other than core/gateway/cloudflare-ai-gateway, and when a gateway carries more rules than the 20 Cloudflare keeps. That the gateway reads the window in seconds, and refuses as described, is Cloudflare's documented behaviour and no build observes it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/period-budget-cloudflare-ai-gateway:1.0.0 - v1.0.0
FeatureRecipe
period-budget-azure-api-management
Realizes core/period-budget on core/gateway/azure-api-management's terraform Recipe; a Release binds it in its Manifest. It writes one token limit into the API's policy and emits no file. The gateway counts the prompt and completion tokens each reply reports, together, and prices nothing: past a quota over an hour, a day, a week, a month or a year, each a fixed window on the UTC clock, it answers HTTP 403 until the window ends, and past a rate over one minute HTTP 429. So it honours unit total_tokens, a period of 1h, 1d, 7d, 1mo, 12mo or 1m, and a whole-number limit, for everything the API serves or for each API Management subscription, and refuses the rest by name. It counts after a reply arrives, so concurrent requests can pass the limit, and keeps a separate count at each gateway of the instance. It halts on any component other than core/gateway/azure-api-management. That the policy applies and refuses as described is Microsoft's documented behaviour on every tier but Consumption, and no build observes it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/period-budget-azure-api-management:1.0.0 - v1.0.0
FeatureRecipe
manual-trigger
Realizes core/manual-trigger on core/compute/task's terraform Recipe, terraform/task-ecs; a Release binds it in its Manifest. It emits nothing: the host's Recipe carries the change, deploying a task with no schedule as its task definition alone, without starting it. A scheduled task runs on its schedule as it does without the attachment. A run is started by hand by the operator with aws ecs run-task, giving the cluster, the task definition, named after the component, and the subnets, security groups and public IP setting the Release sets under the task's infra.overrides.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/manual-trigger:1.0.0 - v1.0.0
FeatureRecipe
backup-native
Component-scope FeatureRecipe for core/backup on terraform through the managed database service's own automated backup. It renders nothing: the Recipe realizing the attached database configures the backup from the attachment, because the backup is a setting of the database that Recipe writes, and a FeatureRecipe writes files of its own and cannot set one on its host. It binds on a database realized by terraform/postgres-rds, terraform/mysql-rds, terraform/mariadb-rds, terraform/mysql-rds-module, terraform/mariadb-rds-module, the -terraform-aws-modules Recipe of each engine, terraform/postgres-cloud-sql, terraform/mysql-cloud-sql or terraform/postgres-flexible-server, each of which keeps infra.retention_days, and a daily infra.schedule where its service takes a time, as the service's automated backups, and refuses what its service cannot keep. infra.point_in_time_recovery is the host's Recipe's to keep, each with its service's own default. Amazon RDS and the flexible server keep it with their backups, so left out or true is what they give and false halts by name. Cloud SQL keeps it on in this catalog, as high availability needs: false turns it off, and halts beside high availability, or beside read replicas on MySQL. The build halts, naming the attachment, on a host whose Recipe configures no backup of its own from core/backup, on infra.role_arn, which AWS Backup reads, on a store connection, infra.volume or infra.credentials_secret, since the service keeps its backups in its own storage, and on infra.run_as_user or infra.run_as_group, since it runs no job in a cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/backup-native:1.0.0 - v1.0.0
FeatureRecipe
generated-secret
Terraform realization of core/generated-secret. Pure wiring: it declares nothing but the capability it binds to and the Feature it implements, and renders no resource of its own. The credential is provisioned and rotated by the target the Terraform Recipe manages — a managed database Recipe sets the provider's own managed-password directive — so there is no value to inject and no secret to emit. Its whole contribution is that the slot the component declares is bound, and bound by the form that says the engine owns the credential. The name generated-secret is shared with the cloudformation and helm realizations of the same Feature, so a Manifest references this one provider-qualified, as terraform/generated-secret:1.0.0. A bare generated-secret:1.0.0 matches more than one and resolves to none, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/generated-secret:1.0.0 - v1.0.0
FeatureRecipe
backup-aws-backup
Component-scope FeatureRecipe for core/backup on terraform through AWS Backup: writes a backup vault, a backup plan and a selection naming the attached database's Amazon RDS DB instance, in a module of the attachment's own. It backs up an instance realized by terraform/postgres-rds, terraform/mysql-rds or terraform/mariadb-rds. Each backup is a snapshot of the instance that AWS Backup takes on infra.schedule, read in UTC and started within an hour of that time, or daily at 05:00 UTC where it is unset, and keeps in the vault for infra.retention_days days. AWS Backup keeps a continuous backup, its point-in-time recovery, only where a rule asks for one, so left out or false it keeps snapshots alone; infra.point_in_time_recovery true asks, for at most 35 days. AWS Backup runs each backup under the existing IAM role infra.role_arn names, which the attachment must set. The build halts, naming the attachment, without infra.role_arn, on a host this realization cannot name — one realized by another Recipe, or a datastore AWS Backup does not protect, such as an OpenSearch Service domain, an MSK cluster or an Amazon MQ broker — on a schedule AWS Backup cannot keep, such as one starting more than once an hour, on a retention over 36500 days, on a store connection, infra.volume or infra.credentials_secret, since AWS Backup keeps its backups in its own vault, on infra.run_as_user or infra.run_as_group, since it runs no job in a cluster, and on a continuous backup kept over 35 days or beside infra.keep_automated_backups false. The instance keeps its automated backups beside these, for 7 days or as long as the continuous backup, unless the attachment sets infra.keep_automated_backups to false.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/backup-aws-backup:1.0.0 - v1.0.0
FeatureRecipe
upstream-key-pinning-tyk
Component-scope FeatureRecipe for core/upstream-key-pinning on the Tyk gateway, core/gateway/tyk, on Kubernetes. The pins are a setting the gateway reads rather than a resource of their own, so this recipe emits no file: the gateway's Recipe reads the attachment and hands every entry to the gateway through the chart's environment extension point, as the one setting the gateway takes its upstream pins from. An upstream is the host the gateway dials, `*.example.com` for every host one label under a suffix, or `*` for every upstream; the most specific entry applies, and it replaces the `*` entry rather than adding to it. Setting any pin changes how the gateway reaches every upstream: from then on it checks an upstream's key rather than the chain the upstream presents, so an upstream no entry covers is reached with its certificate not verified at all. `*` covers every upstream with no entry of its own and holds each of them to its key. It halts on an upstream given more than one key, on an upstream holding a comma, a colon or a misplaced `*`, and when the attachment is carried by any component other than core/gateway/tyk.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/upstream-key-pinning-tyk:1.0.0 - v1.0.0
FeatureRecipe
audit-logging-azure-api-management
Binds core/audit-logging on core/gateway/azure-api-management's terraform Recipe to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file and changes no log. Azure API Management logs each model request — its token counts, the model and, where turned on, the prompt and completion — to Azure Monitor, through a diagnostic setting on the instance into a Log Analytics workspace. That setting belongs to the instance, which the API's Recipe reads and never configures, so the build halts on every attachment, naming it. The halt also names what the attachment asks for that the log cannot keep: a cost while require_cost is on, since the log keeps none, and a retention where retention_days is given, since how long a row lasts is the workspace's retention rather than a setting of the gateway. It halts on any component other than core/gateway/azure-api-management.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/terraform/audit-logging-azure-api-management:1.0.0 - v1.0.0
FeatureRecipe
target-selection-traefik-hub
Binds core/target-selection on the gateway core/gateway/traefik-hub to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no balancer. Traefik picks by lowest response time only among the servers of one Kubernetes Service, which one Chat Completion middleware covers with one model name and one credential; between Services it balances by weight or by a hash of the client's address, or fails over in a fixed order. On this gateway each model wired to the models socket is served on a route of its own carrying that one model, so there is nothing to choose between, and no Traefik strategy chooses a target by cost. So the build halts on every attachment, naming the strategy asked for and why the gateway cannot honour it; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/traefik-hub.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/target-selection-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
traffic-shadowing-nginx
Component-scope FeatureRecipe for core/traffic-shadowing on the NGINX gateway, core/gateway/nginx, on Kubernetes. The copy is a setting inside the server configuration the gateway's own Recipe writes rather than a resource of its own, so this recipe emits no file: the NGINX Recipe reads the attachment and adds the mirror to the location it forwards, sending a copy of every forwarded request, with its original URI, to the Service the attachment's infra names, and discarding that Service's response. NGINX copies every request it forwards and has no setting that copies a share of them — the gateway declares dataplane.aggrega.ai/request-mirroring: full-rate — so this recipe halts on a shadow_percent below 100 rather than copying more traffic than was asked for. It also halts when the attachment is carried by any component other than core/gateway/nginx, because no other component's Recipe reads it, and the attachment would render no copy at all. The Service the copies go to has to resolve when NGINX starts, as the upstream does.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/traffic-shadowing-nginx:1.0.0 - v1.0.0
FeatureRecipe
target-selection-kong-enterprise
Binds core/target-selection on the gateway core/gateway/kong-enterprise to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no balancer. Kong Gateway Enterprise chooses among a route's targets with AI Proxy Advanced, whose balancer can pick the lowest-latency target or the lowest-usage one. On this gateway each model wired to the models socket is served on routes of its own, which carry that one model, so a balancer there has one target and nothing to choose between. Choosing by cost would also need a price per target, which nothing the catalog renders supplies. So the build halts on every attachment, naming the strategy asked for and why the gateway cannot honour it; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/kong-enterprise. It writes no plugin. Kong sells AI Proxy Advanced under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips the plugin silently.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/target-selection-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
target-selection-litellm
Realizes core/target-selection on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes the proxy's routing strategy into the configuration the gateway's own Recipe renders, and emits no file. The choice applies only among entries of the gateway's model map that publish the same alias. A model wired to the models socket is served under its connection's own alias, which no other entry may publish, so the entries that share an alias are the ones typed into values.models. With no attachment the proxy picks among them at random. Strategy latency is written as latency-based-routing: the proxy sends a request to the entry whose last ten replies took the least time on average, per output token where a reply reports its tokens, or to its first token for a streamed request, and an entry not yet tried counts as quickest. Strategy cost is refused by name. The proxy looks an entry's target up, as written, in a price table that names OpenAI's and Anthropic's models without the provider prefix a target carries, and prices an entry it does not find at a fixed fallback, so the entries sharing an alias cost alike and the first is taken. The build also halts when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/target-selection-litellm:1.0.0 - v1.0.0
FeatureRecipe
target-selection-agentgateway
Binds core/target-selection on the gateway core/gateway/agentgateway to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no virtual model. agentgateway 1.6.0 chooses among the targets of one model name by drawing two of them at random and taking the one whose health, divided by the time it has taken to answer, scores higher. The target observed to answer quickest is therefore taken only when it is drawn, and not always then, so strategy latency is refused by name; nothing in its configuration chooses a target by cost, and its model cost catalog prices none of the models wired to it, so strategy cost is refused by name too. Each model wired to the models socket is also served under its own alias with one target, which leaves nothing to choose between. With nothing wired to that socket the build halts saying the gateway serves no model route. It halts on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/target-selection-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
target-selection-higress
Binds core/target-selection on the gateway core/gateway/higress to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no balancer. Higress has a plugin that measures latency — ai-load-balancer with LeastTotalLatency picks the service of lowest recent response time — but its choice does not take effect on the routes this gateway renders. That plugin runs after the AI proxy, which by then has set the request's model name, upstream address and credential from the candidate the route already resolved, and the balancer only writes its choice into a request header that nothing routes by unless an Envoy filter the vendor requires is added, which this catalog does not render. So the chosen service would be reached, if at all, carrying another candidate's proxy rule. strategy cost is refused for a second reason: Higress has no strategy that chooses a target by cost. The build halts on every attachment, naming why. An alias with one candidate has nothing to choose between, and with nothing wired to the models socket the gateway serves no model route; each halts saying so. It halts on any component other than core/gateway/higress. Without this Feature the gateway spreads an alias's requests across its candidates in equal shares.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/target-selection-higress:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-traefik-hub
Realizes core/semantic-caching on the gateway core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, reads the attachment and writes Traefik Hub's semantic cache onto the route of each model wired to the gateway's models socket, keeping the replies in the core/datastore/redis wired to its cache_store socket. A reply is stored per wired model and per caller: under the caller the gateway's key check verified, one of the callers infra.callers declares, each by its own key, so no caller is served a reply stored for another. The cache embeds a request's text with embedding_model, the alias of a models connection, through that model's own endpoint and credential, and serves a stored reply at least similarity_threshold alike, written as the cache's distance bound, one minus the threshold, which is the cosine distance: Traefik's reference gives the bound's direction, not its measure. A reply is kept ttl_seconds. A request too large to embed is answered uncached. It needs a Traefik Hub release from v3.19.11, or from v3.20.6, which keep callers apart by default; an older image halts the build. So it keeps scope key alone, and refuses end_user, the default, by name: the gateway cannot tell end users behind one key apart. It halts with no caller declared, with nothing wired to cache_store or the models socket, on an embedding_model no models connection carries, on a second such attachment, and on any component other than core/gateway/traefik-hub. That the gateway caches is not observed by any build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
shared-config
Boundary-scope FeatureRecipe (executionScope: boundary): renders ONCE per boundary, emitting a single <boundary>-shared-config ConfigMap from its local chart source. The boundary-derived values flow through valuesMapping, so each boundary's ConfigMap carries that boundary's own resolved name and tier. The artifact is unmistakably boundary-level — one per boundary, named for the boundary — distinct from any per-instance ConfigMap a component's own chart emits.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/shared-config:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-litellm
Realizes core/semantic-caching on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes the gateway's own cache setting into the configuration the gateway's Recipe renders, and emits no file. The gateway keeps its replies in the core/datastore/redis wired to its cache_store socket, reached over TLS while it publishes redis-tls, turns each prompt into a vector with embedding_model, an alias of its model map, and serves a stored reply whose prompt is at least similarity_threshold alike; the store removes a reply ttl_seconds after it was stored. Chat requests, streamed or not, and responses and messages requests are cached; completions, embeddings and every other kind of request are not. Replies are kept apart by the key each request presents, which the gateway verifies and a request cannot name for itself. Under scope end_user it also keeps apart the end users requests name, in an x-litellm-end-user-id or x-litellm-customer-id header or in the body's user, metadata.user_id, litellm_metadata.user or safety_identifier field. The gateway verifies none of them, so a caller naming another end user is served that end user's replies, and a request naming none shares one cache with every other request naming none under the same key. Under scope key the end user is left out, though the body's user field, a request parameter, still keeps apart requests naming different ones. The build halts unless the gateway's api_key slot is attached and its cache_store socket wired, on an embedding_model the map does not publish, on a second such attachment on one gateway, and when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-litellm:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-kong-enterprise
Binds core/semantic-caching on the gateway core/gateway/kong-enterprise to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file, writes no cache and binds no store. Kong Gateway Enterprise carries a semantic cache, AI Semantic Cache, but has no setting that keeps one caller's cached replies apart from another's. Kong tells callers apart only as the consumers an authentication plugin names, a consumer being the holder of a key, not the end user behind it. The gateway declares no callers: the import of its routes file adds and updates what the file names and removes nothing, so a consumer and its key, once written, would stay able to authenticate after the caller is removed from the Release, and a rotated key would leave the old one working. Nor can a key-auth key be held by reference, so a caller's key could reach Kong only as a literal in a rendered file. Under end_user and under key alike every request would share one cache, and a reply generated for one caller could be served to another. So the build halts on every attachment, naming the scope asked for and the caller separation the cache lacks. With nothing wired to the models socket the gateway serves no model route to cache, and it halts saying so. It halts on any component other than core/gateway/kong-enterprise. It writes no plugin. Kong sells the cache under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips the plugin silently.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-higress
Binds core/semantic-caching on the gateway core/gateway/higress to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file, writes no cache and binds no store. Higress 2.2.5 carries a semantic cache, its ai-cache plugin, but it keeps no caller's replies apart: every stored reply sits under one key prefix, set once for the plugin, followed by the prompt text taken from the request body, and nothing in that key names a caller or an end user. The gateway also declares no callers, so it identifies none. Under end_user and under key alike every request would share one cache, and a reply generated for one caller could be served to another sending a similar prompt. So the build halts on every attachment, naming the scope asked for and the caller separation the cache lacks. With nothing wired to the models socket the gateway serves no model route to cache, and it halts saying so. It halts on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-higress:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-agentgateway
Binds core/semantic-caching on the gateway core/gateway/agentgateway to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file, writes no cache and binds no store. agentgateway 1.6.0 keeps no cache of model replies: it answers every model request by calling the model, and nothing in its configuration stores a reply or compares one prompt with another. Its prompt caching is a different thing, marking in a request where a provider that caches prompts itself may do so; the reply still comes from the model. So the build halts on every attachment, naming the scope asked for and the cache the gateway lacks. With nothing wired to the models socket the gateway serves no model route to cache, and it halts saying so. It halts on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
semantic-caching-apisix
Realizes core/semantic-caching on the gateway core/gateway/apisix; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, apisix-chart, reads the attachment and writes APISIX 3.18.0's ai-cache onto the chat route of each model wired to the gateway's models socket, keeping the replies in the core/datastore/redis wired to its cache_store socket. A reply is stored per route and per caller: under the consumer the request authenticated as, one of the callers infra.callers declares, each by its own key, so no caller is served a reply stored for another. The gateway embeds a request's last user message with embedding_model, the alias of a models connection, through that model's own endpoint and credential, and serves a stored reply whose message is at least similarity_threshold alike and whose earlier messages are the same; a request identical to a stored one is served from the exact layer. A reply is kept ttl_seconds, and one served by meaning is stored again under that request for ttl_seconds more. Completion requests are not cached. A store or embedding failure is a miss: the request is answered. So it keeps scope key alone, and refuses end_user, the default, by name: the gateway cannot tell end users behind one key apart. It halts with no caller declared, with nothing wired to cache_store or the models socket, on an embedding_model no models connection carries, on a second such attachment, and on any component other than core/gateway/apisix.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/semantic-caching-apisix:1.0.0 - v1.0.0
FeatureRecipe
retry-policy
Component-scope FeatureRecipe for core/retry-policy on an AI gateway whose realization renders its routing table as a configuration body. The control it turns on is a setting the running proxy reads out of that body rather than a resource of its own, so this recipe emits no file: it writes the setting into the configuration the attached gateway already renders, and every other byte of that body is left as the gateway's own Recipe authored it. max_attempts is the total number of times a request may be sent, while the gateway counts only the sends that follow the first one, so the number written is one fewer than the attempts asked for — three attempts is two retries. The retries apply to the request as a whole and are the gateway's, not the caller's: a caller sees one answer whichever send produced it. The body it writes into is the one core/gateway/litellm renders, so the build halts when the attachment is carried by any other component.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-higress
Realizes core/retry-policy on the gateway core/gateway/higress; a Release binds it in its Manifest. It emits no file: the gateway's Recipe, higress-chart, reads the attachment and writes the route's own retry as Ingress annotations onto every alias's model route. max_attempts counts every send, the first included, while Higress's proxy-next-upstream-tries counts the sends after the first, so the number written is one fewer than the attempts asked for: three attempts is two retries. max_attempts 1 writes the retry off, so each request is sent once; this is written out rather than left unset, because a Higress route carrying no retry annotation takes the retry the gateway's Istio control plane applies by default. The retries are the gateway's, so a caller sees one answer whichever send produced it. A request is sent again, at once, after a refused or reset connection and after a reply of 500 to 599 or 429; a POST is sent again, which the gateway's own default retry excludes. The route sets no per-try deadline, so a model that accepts the connection and then answers nothing or slowly is waited on and not sent again, and a reply the gateway forms itself is not sent again. A streamed request is sent again only before its reply has begun, and a request whose body is too large to buffer is sent once. It refuses max_attempts above 2147483648, the most sends Higress's retry count holds, and halts on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy-higress:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-kong-enterprise
Binds core/retry-policy on the gateway core/gateway/kong-enterprise to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no retry count. Kong Gateway Enterprise's AI Proxy Advanced holds a count of retries on each route's balancer, spent across that route's targets. On this gateway each model wired to the models socket is served on routes of its own, which carry that one model, so a retry has to send the request to that same target again, and Kong's reference reads both ways on whether the balancer does: the reference has a failed request move to the next target, and Kong's changelog fixed the balancer retrying the first target more than once, while its load-balancing page says a failure retries the same or another target. The code that decides it ships compiled, and no measurement on a licensed gateway has settled it. The retry count of the service a route belongs to resends no POST once sent and acts on no status code, so it does not honour the Feature either. So the build halts on every attachment, naming why; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/kong-enterprise. It writes no plugin. Kong sells AI Proxy Advanced under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips the plugin silently.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-traefik-hub
Realizes core/retry-policy on the gateway core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, reads the attachment and writes Traefik's retry middleware as the last Middleware of every model route, or of each model a fallback chain tries, so a model is sent a request again before the request falls back. max_attempts counts every send, the first included, and so does the middleware's attempts, so it is written as given; 1 writes nothing, and each request is sent once. A request is sent again at once after a connection that carried nothing to the model and after a reply of 429 or of 500 to 504, a POST included; any other reply ends the retries. A reply the gateway forms itself — a key, guard, masking or quota refusal — is never sent again, and neither is a streamed reply once it has begun. The retries are the gateway's, so a caller sees one answer. It halts on a second such attachment, with nothing wired to the models socket, on a gateway image older than Traefik Hub v3.20 while max_attempts is above 1, and on any component other than core/gateway/traefik-hub. No build proves that a running gateway sends a request again.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-envoy-gateway
Realizes core/retry-policy on the gateway core/gateway/envoy-gateway; a Release binds it in its Manifest. It emits no file: the gateway's Recipe, envoy-gateway, reads the attachment and writes it as the retry of the one traffic policy on the route serving the models wired to its models socket, the policy that also carries a period budget's rate limit. max_attempts counts every send, the first included, while Envoy Gateway counts the sends after the first, so the number written is one fewer than the attempts asked for: three attempts is two retries, and one attempt is a single send. The retries are the gateway's, so a caller sees one answer whichever send produced it. A send is repeated on a refused connection or stream, on a reset — an upstream that does not answer at all — and on a reply of 429, 503 or 504. The route's request timeout spans every send, so a model that answers nothing within it is not sent again. Each wired model's rule holds one backend, so a request for any model the gateway serves is sent at most max_attempts times. It refuses max_attempts above 2147483648, the most sends Envoy Gateway's retry count holds, and halts on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/envoy-gateway.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy-envoy-gateway:1.0.0 - v1.0.0
FeatureRecipe
retry-policy-agentgateway
Realizes core/retry-policy on the gateway core/gateway/agentgateway; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, agentgateway-chart, writes a retry policy into the configuration, attached to the listener every model request arrives on. max_attempts counts every send, the first included. agentgateway 1.6.0 sends a request once more than its retry policy's attempts, so the policy's attempts is max_attempts less one, and max_attempts 1 writes no policy: each request is sent once. A request is sent again, at once and to the same model, after a reply of 429, 500, 502, 503 or 504, a refused or reset connection, a connection taking longer than 11 seconds to open or a failed name lookup. The configuration sets no timeout, so a model that accepts the connection and answers nothing is waited on and not sent again. A request whose body is larger than 64 KiB is sent once, and a refusal the gateway answers itself — a guard's, a budget's, a missing key's — is never sent again. It refuses max_attempts above 256, the most sends that policy holds, and halts on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/retry-policy-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
response-validation-higress
Binds core/response-validation on the gateway core/gateway/higress to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file and writes no check. Higress 2.2.5 checks a model's reply against a JSON Schema with its ai-json-resp plugin, but that check is weaker than the Feature's name. A reply that passes is answered with the bare JSON the plugin finds in the reply's message, not with the chat completion the route's OpenAI API returns. Any client skips the check by sending the request header X-HIGRESS-AI-JSON-RESP: true. And a schema nested more than six levels deep is not checked at all: the plugin then checks only that the reply holds JSON. So the build halts on every attachment, naming all three. With nothing wired to the models socket the gateway serves no model route whose replies could be checked, and it halts saying so. It halts on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/response-validation-higress:1.0.0 - v1.0.0
FeatureRecipe
response-validation-traefik-hub
Binds core/response-validation on the gateway core/gateway/traefik-hub to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file and writes no check. No middleware in Traefik Hub's reference checks a model's reply against a JSON Schema. Its Chat Completion middleware checks a request against the chat completion schema, not a reply; its API management checks request bodies against an OpenAPI document, not a model's reply; and its LLM Guard can send a reply to an outside service and judge that service's answer, which this realization does not wire. So the build halts on every attachment, naming why; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/traefik-hub.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/response-validation-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
response-validation-agentgateway
Binds core/response-validation on the gateway core/gateway/agentgateway to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no check. agentgateway 1.6.0 checks no model reply against a JSON Schema. What it runs on a reply is its guards: a regular-expression guard, a content-safety service it calls, or a webhook of the operator's own, none of them a schema check this realization could write. So the build halts on every attachment, naming the check the gateway lacks. With nothing wired to the models socket the gateway serves no model route whose replies could be checked, and it halts saying so. It halts on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/response-validation-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
random-password-kubernetes-secret-generator
Helm realization of core/random-password through the kubernetes-secret-generator controller. It renders, from a co-located Helm chart, a StringSecret, which the controller reconciles into a Secret of the same name holding a generated password; the rendered output names the StringSecret and carries no value. The name is values.name where one is supplied, and otherwise {rendered_name}-{attachment_name} — the host component's rendered name, or its identity on a host reporting none, then the attachment name with underscores folded to hyphens — so the derived name is a valid Kubernetes object name wherever its host part is one. The password is written under values.key, and the published provides point a consumer at that Secret and that key. The password is drawn once: the controller generates a field only while it is empty and the StringSecret does not ask for regeneration, so a password shared by a running database and its clients does not change under them. The controller produces exactly values.length characters by encoding random bytes: with values.symbols off the encoding is hexadecimal, so the password holds digits and the letters a to f only; with it on the encoding is the controller's default, base64, which adds upper-case letters and the characters + and /. The kubernetes-secret-generator controller must be running in the cluster; the catalog does not install it. The build halts unless the Release affirms it with infra.control_plane_present. Reach for this realization where that controller runs in the cluster, and for random-password-external-secrets where the External Secrets Operator runs instead — it is installed by core/security/external-secrets.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/random-password-kubernetes-secret-generator:1.0.0 - v1.0.0
FeatureRecipe
provided-secret
Helm realization of core/secret-provided. origin — from values or infra, whichever carries it — is published as the `value` provides key, and the attachment dictates where it is delivered with its own `inject:` block on a transparent slot — this realization carries no injectMapping of its own, so an attachment without that block delivers the value nowhere. The value reaches a host Recipe that reads the slot's inject pipe, and it is written into the rendered output as it stands, so the material sits in the artifact. A host Recipe that renders the slot as a workload secret reference, a secretKeyRef, cannot take it: a value carries no locator for the reference to point at, and the build halts naming the slot. Fill such a slot with a referenceable credential instead. The name provided-secret is shared with the compose realization of the same Feature, so a Manifest references this one provider-qualified, as helm/provided-secret:1.0.0. A bare provided-secret:1.0.0 matches both and resolves to neither, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/provided-secret:1.0.0 - v1.0.0
FeatureRecipe
public-ingress
Component-scope FeatureRecipe for core/public-ingress: emits one Kubernetes Ingress per attached component from a co-located chart, routing the host the operator names (infra.host) and the path the developer writes (values.path) to the named port of the Service the component's own Recipe emits. Resource Generation via source + valuesMapping; the per-component .context.* root is available at component scope. The backend Service name is the operator-set service_name where one is supplied, and otherwise the primary Service a synthesized workload emits for the attached component — the component's rendered name followed by the primary service key, which is how workload synthesis names it. The Ingress itself is named from the same rendered name, the name the helm provider spells in the characters a Kubernetes name allows. An ingress controller is a runtime prerequisite: the build halts unless the Operator affirms one is running in the target cluster (infra.control_plane_present), so an Ingress is never emitted into a cluster with nothing to reconcile it into routing. Where the Operator names a certificate Secret (infra.tls_secret), the Ingress gains a TLS section listing the host and that Secret, which is how an Ingress asks its controller to serve the host over HTTPS; the Secret is referenced and never emitted, and with the key unset the Ingress carries no TLS section.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/public-ingress:1.0.0 - v1.0.0
FeatureRecipe
random-password-external-secrets
Helm realization of core/random-password through the External Secrets Operator. It renders, from a co-located Helm chart, a Password generator and an ExternalSecret that fills a Secret from it, so the operator's controller draws the password inside the cluster and writes it into the Secret; the rendered output names the generator and the Secret and carries no value. The Secret, the generator and the ExternalSecret share one name: values.name where one is supplied, and otherwise {rendered_name}-{attachment_name} — the host component's rendered name, or its identity on a host reporting none, then the attachment name with underscores folded to hyphens — so the derived name is a valid Kubernetes object name wherever its host part is one. The password is written under values.key, and the published provides point a consumer at that Secret and that key. The password is drawn once. The generator produces a new value on every invocation, so the ExternalSecret is set to create the Secret once and never refresh it, and a password shared by a running database and its clients does not change under them. Deleting the Secret draws a new password. The External Secrets Operator must be running in the cluster, with its generator resources installed — core/security/external-secrets installs both — and the build halts unless the Release affirms it with infra.control_plane_present. Reach for this realization where the External Secrets Operator runs in the cluster, and for random-password-kubernetes-secret-generator where the kubernetes-secret-generator controller runs instead.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/random-password-external-secrets:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-nginx
Component-scope FeatureRecipe for core/prompt-injection-shield on the NGINX gateway, core/gateway/nginx, on Kubernetes. It emits no file: it points the proxy_pass target the gateway's Recipe writes at the guard wired to the gateway's guard socket, in place of the component on its upstream socket, so every request the gateway forwards goes to the guard. The guard checks each chat request and forwards what it admits to the model its own infra.upstream_endpoint names, which the Release sets to the model the upstream socket names; it answers no completions, embeddings, responses or messages route, and its model listing passes the caller's Authorization header to the model's. The guard must be core/security/nemo-guardrails running the threshold the sensitivity names, 0.5 for standard and 0.001 for strict, the values measured for its detector's default classifier. The build halts otherwise, with nothing on the guard socket, naming it, with more than one guard, on a second attachment, beside core/traffic-shadowing, whose copies would bypass the guard, and on any component other than core/gateway/nginx. The guard reads only the latest user message, and judges no reply, tool-call argument or image. A request it refuses is answered with an ordinary completion, streamed when asked; when the guard or its detector cannot be reached the request is refused and nothing reaches the model.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-injection-shield-nginx:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-traefik-hub
Binds core/prompt-injection-shield on the gateway core/gateway/traefik-hub to a halt naming why the gateway does not keep it; a Release binds it in its Manifest. It emits no file, writes no guard and routes nothing through one. Traefik Hub checks a request for an attempt to override a model's instructions only through a service it calls. Its LLM Guard and Parallel LLM Guard middlewares send a request to a guard service or a model in a shape the realization would write and judge the answer by conditions it would write too; how they work with a prompt-attack detector is not measured, so neither sensitivity maps to a check on this gateway. Its Azure Prompt Shield, Amazon Bedrock Guardrails and GCP Model Armor middlewares call hosted services this catalog does not wire. The build halts on every attachment, naming the sensitivity asked for, and on any component other than core/gateway/traefik-hub.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-injection-shield-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-higress
Binds core/prompt-injection-shield on the gateway core/gateway/higress to a halt naming why the gateway does not keep it; a Release binds it in its Manifest. It emits no file, writes no guard and routes nothing through one. Higress 2.2.5 judges a request for an attempt to override a model's instructions only through a service it calls: its ai-security-guard plugin calls Alibaba Cloud's content-security service, and the qwen3guard plugin its source carries, which its 2.2.5 plugin release does not publish, a Qwen3Guard model service the operator runs. Each lets the request through to the model when that call fails or its answer cannot be read, so a guard that is down or unreachable stops nothing. This realization wires neither and sends no request through a guard, so nothing would check the requests. The build halts on every attachment, naming the sensitivity asked for. It halts on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-injection-shield-higress:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-litellm
Component-scope FeatureRecipe for core/prompt-injection-shield on the AI gateway core/gateway/litellm. It emits no file. The gateway's own Recipe, reading the attachment, sends every entry of the model map to the guard wired to the gateway's guard socket, which checks each chat request and forwards what it admits to the model; this recipe writes a request policy under which the gateway refuses every other kind of request with HTTP 400, forwarding nothing, since the guard checks chat completions only. The guard must be core/security/nemo-guardrails running the threshold the sensitivity names, 0.5 for standard and 0.001 for strict, the values measured for its detector's default classifier. The build halts otherwise, with nothing on the guard socket, naming it, with more than one guard, on a second attachment, on a map target not written openai/<model>, while the provider_key slot or core/semantic-caching is attached or a model is wired to the models socket, and on any component other than core/gateway/litellm. The guard reads only the latest user message, and judges no reply, tool-call argument or image. A request it refuses is answered with an ordinary completion, streamed when asked; when the guard or its detector cannot be reached the request is refused and nothing reaches the model.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-injection-shield-litellm:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-agentgateway
Binds core/prompt-injection-shield on the gateway core/gateway/agentgateway to a halt naming why the gateway does not keep it; a Release binds it in its Manifest. It emits no file, writes no guard and routes nothing through one. agentgateway 1.6.0 judges a request for an attempt to override a model's instructions only through a service it calls: a webhook the operator runs, or a hosted service — Azure AI Content Safety, whose jailbreak detection it can turn on, Google Model Armor or AWS Bedrock Guardrails, each judging by a policy defined in that service. This realization wires none of them, and it sends no model request through a guard, so nothing would check the requests. Its regular-expression guard, which core/prompt-deny-patterns turns on, matches patterns rather than judging with a classifier. So the build halts on every attachment, naming the sensitivity asked for. It halts on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-injection-shield-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-kong
Realizes core/prompt-deny-patterns on the gateway core/gateway/kong; a Release binds it in its Manifest. It emits no file: the KongModelRoutes bundle each of the gateway's Recipes imports reads the attachment and writes Kong 3.9.3's AI Prompt Guard onto the chat and completion routes of each model wired to the gateway's models socket. The guard reads the text of every message of every role across the whole conversation, joined into one string with a space after each message, so an expression can match across two of them, and a completion's prompt. A message whose content is not a single string, such as content in parts or a turn carrying only tool calls, refuses the whole request. Letter case counts unless an expression says otherwise, as (?i) does. A request any of deny_patterns matches is refused whatever allow_patterns says; with allow_patterns set, a request none of them matches is refused. A refused caller receives HTTP 400 with the message "bad request", which names no expression, and so does a request the guard cannot read, such as a body that outgrows Kong's 8 KiB request buffer and is set aside on disk, which the gateway refuses without the guard as well. The build does not check an expression: one Kong cannot compile answers the requests reaching it with HTTP 500. So it honours deny_patterns and allow_patterns, and halts on a list of more than 10 expressions or an expression over 500 bytes, on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/kong. The gateway imports its routes into its database, which keeps the guard after the attachment is removed, until it is deleted through Kong's Admin API.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-deny-patterns-kong:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-litellm
Component-scope FeatureRecipe for core/prompt-deny-patterns on the AI gateway core/gateway/litellm. It emits no file: it adds one entry to the guardrails list of the configuration the gateway's own Recipe renders, turning on the gateway's built-in content filter before a request is forwarded, with one blocking rule per entry of deny_patterns, on every request, so a caller cannot turn it off from the request. A request any expression matches, ignoring letter case, in the text of any message, is refused with HTTP 400 naming the rule by its position in the list — deny_pattern_1 for the first — and not the expression. Images are not examined; an administrator can exempt a key or a team from the gateway's always-on guardrails; and a request of a kind the gateway maps to no guardrail is forwarded unchecked — chat and text completions, the messages API and the responses API are among the kinds it maps. It halts on allow_patterns, because this gateway's filter only refuses a request that matches, and when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-deny-patterns-litellm:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-traefik-hub
Realizes core/prompt-deny-patterns on the gateway core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, reads the attachment and writes a Content Guard of Traefik Hub's AI gateway onto the route of each model wired to the gateway's models socket. The guard's regular-expression engine reads each request as a chat completion and refuses one that any of deny_patterns matches, with HTTP 403 and a plain-text body, before the request is forwarded. Expressions use Go's RE2 syntax, which has no lookahead and no backreference; letter case counts unless an expression says otherwise, as (?i) does, and an expression matches anywhere in the text. Which fields of a chat request the guard reads is not stated in Traefik's reference. The build does not check an expression: one the gateway cannot compile keeps the guard from starting. The guard reads requests only, so a streamed reply passes as it streams. So it honours deny_patterns, and several attachments, each its own guard. It refuses allow_patterns by name, since the guard refuses on a match and admits nothing by one, and halts with nothing wired to the models socket and on any component other than core/gateway/traefik-hub. That the gateway enforces the guard is not observed by any build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-deny-patterns-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
period-budget-traefik-hub
Realizes core/period-budget on core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, writes it as Traefik Hub's quota middleware on the route of each model wired to the gateway, counting in the Redis-protocol store on its cache socket. The quota counts the tokens each reply reports, read, written or both as unit names, or, under usd, their cost in US dollars, priced from a list of models the middleware ships with; a model it has no price for is refused with HTTP 422. It counts over a window that slides with time: what was counted at one moment frees again one period later. Once the count reaches limit, model requests are refused with HTTP 429; a request's count is added once its reply completes, so the request crossing the limit, and any served beside it, is served, and a reply the semantic cache serves counts nothing. Traefik's reference says the quota serves requests uncounted while the store cannot be reached. So it honours every unit, scope model, scope gateway while one model is wired, and a period of hours or days, or minutes or seconds making whole hours. It refuses scope caller, scope gateway with several models wired, a calendar month and any other period by name, and halts on a limit that is not a whole number of tokens, with nothing wired to the models or cache socket, and on any component other than core/gateway/traefik-hub. That the gateway enforces the budget is not observed by any build.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
prometheus-exporter
Component-scope FeatureRecipe for core/prometheus-exporter: emits one Prometheus ServiceMonitor per attached component from a co-located chart, templating the scrape interval, the named port and the metrics path from the operator-set feature infra. Resource Generation via source + valuesMapping; the per-component .context.* root is available at component scope. The ServiceMonitor is named from the attached component's rendered name, the name the helm provider spells in the characters a Kubernetes name allows, and selects the component's Service by its app.kubernetes.io/name label, which on a Service the helm provider synthesizes is the component's identity.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prometheus-exporter:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-agentgateway
Realizes core/prompt-deny-patterns on the gateway core/gateway/agentgateway; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, agentgateway-chart, writes agentgateway 1.6.0's regular-expression guard into the configuration, applied to every model the gateway serves, refusing a model request one of deny_patterns matches. The guard reads the system and developer prompts, every other message, the arguments of tool calls and the results of tool calls, each text on its own, so an expression does not match across two of them; images are not read. Letter case counts unless an expression says otherwise, as (?i) does. A refused caller receives HTTP 403 with the body "The request was rejected due to inappropriate content", which names no expression. The build does not check an expression: one agentgateway cannot compile stops a gateway reading the configuration from starting, and a running one keeps its previous configuration. So it honours deny_patterns, and refuses allow_patterns by name, since agentgateway's guard keeps no list a request must match. It halts with nothing wired to the models socket, on a second such attachment, and on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-deny-patterns-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-apisix
Realizes core/prompt-deny-patterns on the gateway core/gateway/apisix; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, apisix-chart, reads the attachment and writes APISIX 3.18.0's ai-prompt-guard onto the chat route of each model wired to the gateway's models socket, and writes no completion route, since the guard reads no completion request. The guard reads the text of every message, whatever its role, across the whole conversation, joined into one string with a space between messages, so an expression can match across two of them; images and tool-call arguments are not read. Letter case counts unless an expression says otherwise, as (?i) does. With allow_patterns set, a request none of them matches is refused with HTTP 400, "Request doesn't match allow patterns"; then a request any of deny_patterns matches is refused with HTTP 400, "Request contains prohibited content", which names no expression. A request the guard cannot read as a model request is refused with HTTP 400 rather than forwarded unchecked, and one carrying no message text at all is answered with an empty HTTP 200 and not forwarded. The build does not check an expression: one APISIX cannot compile drops every route it is written on when the gateway reads the file, and those answer HTTP 404. So it honours deny_patterns and allow_patterns, and halts with nothing wired to the models socket, on a second such attachment, and on any component other than core/gateway/apisix.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/prompt-deny-patterns-apisix:1.0.0 - v1.0.0
FeatureRecipe
period-budget-higress
Realizes core/period-budget on the gateway core/gateway/higress; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, higress-chart, writes Higress's ai-token-ratelimit plugin onto the model routes, counting in the Redis-protocol store on the cache socket. The gateway adds the prompt and completion tokens each reply reports once the reply completes, over a fixed window of period that opens when the first reply in it is counted, and answers model requests with HTTP 429 while the count exceeds limit, until the window ends; the request crossing the limit is served. Scope gateway keeps one count for all it serves, scope model one for each alias. A reply reporting no usage counts nothing, and a streamed request turning the usage report off is refused with HTTP 403, as are a request with a compressed body or an upgrade and one whose content type contains octet-stream or grpc, or neither contains application/json nor begins with multipart/form-data; a client crafting a body the gateway's validation cannot decode gets past them. Without masking, a body over 32 KiB may be refused with HTTP 413, depending on when the store answers. While the store cannot be reached, model requests are served and not counted. So it honours unit total_tokens, scope gateway or model, and a period of one second, minute, hour or day, however spelt. It refuses usd, prompt_tokens, completion_tokens, scope caller, since the gateway verifies no caller, any other period and a limit that is not a whole number by name, and halts on a second such attachment, with nothing wired to the models or cache socket, and on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-higress:1.0.0 - v1.0.0
FeatureRecipe
period-budget-kong-enterprise
Realizes core/period-budget on the gateway core/gateway/kong-enterprise; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, kong-enterprise-chart, reads the attachment and writes Kong's AI Rate Limiting Advanced into the routes file it imports, one plugin for the whole gateway under scope gateway, or one on each wired model's service under scope model. The gateway counts the tokens each model's reply reports, read, written or both as unit names, in fixed windows of period, each request counted in the one window its time falls in. Once the count reaches limit, the next model request is answered with HTTP 429 until the window ends, so the request that crosses the limit is served. The count is kept in the gateway's PostgreSQL backend, shared by every replica and kept across a restart. Under scope model a request is counted against the alias it named, whichever model answered it. So it honours unit total_tokens, prompt_tokens or completion_tokens, scope gateway or model, and a period of seconds, minutes, hours or days. It refuses usd, scope caller and calendar months by name, and halts on a second such attachment, with nothing wired to the models socket, on a gateway image older than 3.15, and on any component other than core/gateway/kong-enterprise. Kong sells the plugin under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips the plugin silently, so the gateway's Recipe halts until infra.ai_gateway_entitled affirms it. No build proves that a running gateway enforces the budget.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
period-budget-litellm
Realizes core/period-budget on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes the gateway's own budget setting into the configuration the gateway's Recipe renders, and emits no file. The gateway counts what its traffic cost in US dollars, priced from the cost table it keeps per model, and starts the count again at the start of every period: a number of days from midnight UTC, 7d from a Monday, 1mo from the first of the month. Once the count exceeds limit it refuses every further model request, its administrator's own among them, until the next period starts, which it notices within about ten minutes. It reads that count again at most once a minute, so the traffic of that minute can take it past limit before the first refusal. So it honours unit usd, scope gateway and a period of whole days other than 30d, or 1mo, and refuses any other unit, scope or period by name. The gateway keeps the count in its bookkeeping database and checks a budget only for callers it authenticates, so the build halts unless its database_url and api_key credential slots are both attached. It halts when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-litellm:1.0.0 - v1.0.0
FeatureRecipe
period-budget-envoy-gateway
Realizes core/period-budget on core/gateway/envoy-gateway; a Release binds it in its Manifest. It emits no file: the gateway's Recipe writes it as Envoy Gateway's usage-based rate limit on the route serving the wired models, counted by the control plane's rate-limit service in the store on the cache socket. The gateway counts the tokens each reply reports, read, written or both as unit names, over fixed windows on the UTC clock — 1s, 1m, 1h, 1d, 7d from a Thursday, 30d or 365d from 1 January 1970, or 1mo, a calendar month under a setting the Recipe turns on for every monthly rate limit that install keeps. Once the count exceeds limit it answers model requests with HTTP 429 until the window ends. While the store or the rate-limit service cannot be reached, or a check outlasts Envoy's rate-limit timeout, it answers them with HTTP 500, under a setting the Recipe turns on for every rate limit that install keeps, so a hand-applied one fails closed too. Tokens are added once a reply completes, so the request crossing the limit, any answered beside it, and one admitted just before an outage are served, the last uncounted. A store keeping the count in memory, as core/datastore/redis does by default, starts every window again on a restart. So it honours unit total_tokens, prompt_tokens or completion_tokens, scope gateway or model, and those periods, or the same length spelt otherwise. It refuses usd, scope caller, since nothing here verifies a caller, and any other period by name, and halts on a limit that is not a whole number of tokens up to 4294967295, on a second such attachment, with nothing wired to the models or cache socket, and on any component other than core/gateway/envoy-gateway.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-envoy-gateway:1.0.0 - v1.0.0
FeatureRecipe
manual-trigger
Realizes core/manual-trigger on Kubernetes, on core/compute/task's helm Recipe, helm/task; a Release binds it in its Manifest. It emits nothing: the host's Recipe carries the change, rendering a task with no schedule as a suspended CronJob in place of the Job that would run at deploy, because a FeatureRecipe sees neither the host's values nor can it withdraw a resource the host emits. A scheduled task runs on its schedule as it does without the attachment. A run is started by hand by the operator with kubectl create job <run-name> --from=cronjob/<cronjob>, the CronJob named as the component is rendered. The build halts, naming the attachment, on a component whose render emits a Job, which runs as soon as it is deployed, or no CronJob to start a run from, as helm/agenta-migrations renders its run.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/manual-trigger:1.0.0 - v1.0.0
FeatureRecipe
period-budget-agentgateway
Realizes core/period-budget on the gateway core/gateway/agentgateway; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, agentgateway-chart, writes the budget into the configuration, counted in the total tokens each model reply reports. Scope caller is a budget on the API key of each caller infra.callers declares, each counted apart over fixed windows of period aligned to the Unix epoch, a day starting at midnight UTC, and kept across a restart in the gateway's store, which the chart Recipe renders, with one replica. Scope gateway is a token bucket for every model request, and scope model one for each model name a request asks for, filled to limit every period from when the gateway read its configuration and kept in its memory, so a restart, or a changed configuration, fills it again. Once what was counted reaches limit, model requests are answered with HTTP 429 until the window ends or the bucket is filled; the request that reaches it is served, and a reply reporting no tokens counts nothing. It refuses usd, prompt_tokens, completion_tokens, a calendar month, scope caller with no caller declared and a limit that is not a whole number by name, and halts on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
period-budget-apisix
Realizes core/period-budget on the gateway core/gateway/apisix; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, apisix-chart, reads the attachment and writes APISIX's token rate limiting onto the chat routes of the models wired to the gateway's models socket, and writes no completion route, since APISIX counts no tokens on one. The gateway counts the tokens each model's reply reports, read, written or both as unit names, over a window of period that opens with the first model request after the previous window closed rather than on the clock. Once the count reaches limit it answers further model requests with HTTP 503 until the window closes. It adds a request's tokens once the reply is complete, so the request that crosses the limit, and any answered beside it, is served, and a reply the gateway's semantic cache serves counts nothing and is served past the limit. The count is kept in the memory of the gateway process, so the chart runs one replica while the budget is attached, and a restart starts a new window. So it honours unit total_tokens, prompt_tokens or completion_tokens, scope gateway, model or caller, and a period of seconds, minutes, hours or days. Scope caller gives each caller infra.callers declares a count of its own across every route, the gateway telling them apart by the key each presents; with no caller declared it is refused by name. It refuses usd and calendar months by name, and halts on a limit that is not a whole number of tokens, on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/apisix.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/period-budget-apisix:1.0.0 - v1.0.0
FeatureRecipe
k8s-secret-set
Generates a single Kubernetes Secret carrying several caller-supplied data keys, rendered from a co-located Helm chart. The generated Secret uses the deterministic name pattern {rendered_name}-{attachment_name}: the host component's rendered name, which the helm provider spells in the characters a Kubernetes name allows, then the slot name with underscores normalized to hyphens, so the name is a valid RFC-1123 label wherever the host reports a rendered name. On a host that reports none, the host part is the component's identity, and the name is a valid label only where that identity is one. The caller supplies the fixed key names its consuming target reads, so the consumer references the Secret by name and its target reads the keys it already knows. Each entry's value is written into the emitted Secret's stringData, so the material sits in the rendered output: this suits evaluation and local work, not a production credential.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/k8s-secret-set:1.0.0 - v1.0.0
FeatureRecipe
k8s-external-secret
Generates a Kubernetes ExternalSecret resource from a co-located Helm chart, so a secret-store operator running in the cluster fetches the credential and reconciles it into a Secret. origin — from values or infra, whichever carries it — is read as the key to fetch from the store, so the rendered output carries the locator rather than the credential value. The ExternalSecret reads from a ClusterSecretStore named platform-secret-store and refreshes hourly. The reconciled Secret uses the deterministic name pattern {rendered_name}-{attachment_name}: the host component's rendered name, which the helm provider spells in the characters a Kubernetes name allows, then the slot name with underscores normalized to hyphens, so the name is a valid RFC-1123 label wherever the host reports a rendered name. On a host that reports none, the host part is the component's identity, and the name is a valid label only where that identity is one. The published provides point a consumer at that Secret and the data key within it. Reach for this realization where the External Secrets Operator runs in the cluster with a ClusterSecretStore named platform-secret-store, and the credential should be fetched from that store rather than written into the rendered output. Reach for k8s-existing-secret where the Secret already exists in the target namespace, and for k8s-secret-from-parameter where neither holds and the material may sit in the rendered output, as it may for evaluation.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/k8s-external-secret:1.0.0 - v1.0.0
FeatureRecipe
k8s-secret-from-parameter
Generates a Kubernetes Secret resource from an authored literal value, rendered from a co-located Helm chart. The generated Secret uses the deterministic name pattern {rendered_name}-{attachment_name}: the host component's rendered name, which the helm provider spells in the characters a Kubernetes name allows, then the slot name with underscores normalized to hyphens, so the name is a valid RFC-1123 label wherever the host reports a rendered name. On a host that reports none, the host part is the component's identity, and the name is a valid label only where that identity is one. The value is written into the emitted Secret, so the credential material sits in every rendered artifact the build produces. Reach for this realization where no Secret exists yet and no secret-store operator runs in the cluster, and the material may sit in the rendered output — for evaluation and local work, not for a production credential. Reach for k8s-existing-secret where the Secret already exists in the target namespace, and for k8s-external-secret where a secret-store operator runs in the cluster and should fetch the credential from its store.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/k8s-secret-from-parameter:1.0.0 - v1.0.0
FeatureRecipe
k8s-secret-from-value
Helm realization of core/secret-value that a consuming Recipe can read by reference. It renders one Kubernetes Secret holding the value and publishes the Secret's name and data key, so a Recipe reading the slot as a secretKeyRef points at it, and it also delivers the value into the component's inject pipe, as env-secret does, so a Recipe reading the pipe receives it as a variable. The value is written into the emitted Secret, and a Recipe reading the pipe writes it into its own manifest as well, so the material sits in the rendered output: for evaluation and local work, not for a production credential. The Secret is emitted for every attachment, whether or not its host reads it by reference. An origin holding a Docker Compose variable reference of the form ${…} halts the build, since Kubernetes substitutes none; set the real value through the Release instead. Reach for this realization where a host reads the credential by reference, or where one Manifest maps core/secret-value for hosts of both kinds. Reach for env-secret where every host reads the inject pipe and no Secret is wanted, and for core/referenced-secret where the Secret already exists or a secret-store operator fetches the credential.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/k8s-secret-from-value:1.0.0 - v1.0.0
FeatureRecipe
generated-secret
Helm realization of core/generated-secret. Pure wiring: it declares nothing but the capability it binds to and the Feature it implements, and renders no resource of its own. It is for a controller-driven Recipe that hands the credential to its controller: helm/mariadb-cluster with infra.credentials set to generated emits a MariaDB custom resource naming no root-password Secret, and the mariadb-operator controller generates the password into a Secret it creates and owns. Binding the component's required credential slot here says the controller owns the material, so no value is authored, none travels the inject pipe, and no Secret is emitted. Bind it only where the Recipe leaves the credential to its controller. The synthesis helm/mariadb delivers its credential slots itself, so it cannot take this binding and the build halts. helm/postgres-cloudnative-pg, whose controller also owns its credentials, realizes a Component whose credential slots are optional and refuses any attachment on them, so there they are left unbound instead. The name generated-secret is shared with the terraform and cloudformation realizations of the same Feature, so a Manifest references this one provider-qualified, as helm/generated-secret:1.0.0. A bare generated-secret:1.0.0 matches more than one and resolves to none, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/generated-secret:1.0.0 - v1.0.0
FeatureRecipe
http-route
Component-scope FeatureRecipe for core/http-route: emits one Gateway API HTTPRoute per attached component from a co-located chart, attaching the route to the gateway the Operator names and forwarding the host the operator names (infra.host) and the path the developer writes (values.path) to a port on the Service the component's own Recipe emits. Resource Generation via source + valuesMapping; the per-component .context.* root is available at component scope. The route is emitted at gateway.networking.k8s.io/v1, which is the version the routing API both serves and stores it at. The backend Service name is the operator-set service_name where one is supplied, and otherwise the primary Service a synthesized workload emits for the attached component — the component's rendered name followed by the primary service key, which is how workload synthesis names it. The emitted resource is named for the component that carries the attachment, by the same rendered name — the name the helm provider spells in the characters a Kubernetes name allows — and for the attachment itself, so one component carrying two of these attachments emits two routes rather than one overwriting the other. A gateway control plane is a runtime prerequisite: the build halts unless the Operator affirms one is running in the target cluster (infra.control_plane_present), so a route is never emitted into a cluster with nothing to reconcile it into routing. The build also halts where the attachment names no gateway to attach to or no Service port to forward to, because neither has a fallback that could be right.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/http-route:1.0.0 - v1.0.0
FeatureRecipe
k8s-dns-resolver
Helm implementation of core/k8s-dns-resolver. Inject-only: before the base Recipe renders, the injectMapping qualifies the consumer's injected host with the target's Kubernetes namespace, producing the cross-namespace in-cluster DNS name (service.namespace). It reads the target's namespace from the consumer's resolved connection to that target, and falls back to the "default" namespace when none is set — matching where the helm provider deploys an unannotated workload, so the qualified host is always reachable and never malformed. Only the consumer's inject pipe is rewritten; the target's published host is untouched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/k8s-dns-resolver:1.0.0 - v1.0.0
FeatureRecipe
gcp-workload-identity
Component-scope FeatureRecipe for core/gcp-workload-identity: emits one Kubernetes ServiceAccount per attached component from a co-located chart, annotated with the operator-set Google service-account email for GKE Workload Identity. An operator-set service_account_name is used exactly as written, so a name a Workload Identity binding already trusts is never rewritten; left unset, the name composes the host component's rendered name — the name the helm provider spells in the characters a Kubernetes name allows — with the name the attachment is carried under, so attaching this Feature to several components in one namespace emits a distinct ServiceAccount for each. The name is published as provides so the operator can mirror it onto the consumer's aggrega.ai/service-account annotation. Resource Generation via source + valuesMapping; the per-component .context.* root is available at component scope.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/gcp-workload-identity:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain
Realizes core/fallback-chain on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes the chain and its bound into the configuration the gateway's own Recipe renders, and emits no file. Once the gateway gives up on the alias a request asked for, after its retries within that alias, whose count core/retry-policy sets, it sends the request to the aliases the alias's entry lists, in order, until one answers, and a fallback that fails follows its own entry in turn. Each alias tried after the first counts one fallback; a fallback's own entry is followed only while fewer than max_hops have been counted, and every alias one entry lists is tried whatever the count. The build halts where chains names an alias the gateway's model map does not publish, naming it and the aliases the map publishes, and when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/fallback-chain:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain-kong-enterprise
Realizes core/fallback-chain on the gateway core/gateway/kong-enterprise; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, kong-enterprise-chart, reads the attachment and writes each chained alias's chat and completion routes through Kong's AI Proxy Advanced instead of its AI proxy, with the alias's model and the models its walk falls back to as targets in priority order. The walk is the alias, then its entry's fallbacks in order, each counting one fallback; a fallback's own entry is followed in turn only while fewer than max_hops have been counted, and every alias one entry lists is placed whatever the count. The balancer sends a request to each target of the walk once, in order, until one answers, moving on after a connection error or timeout, an invalid reply header or a reply of 429, 500, 502, 503 or 504; every other reply is returned as it came. An alias with no entry keeps its AI proxy. The build halts where chains names an alias the models socket does not carry; beside core/retry-policy, because Kong keeps one retry count for the whole walk and cannot send to a model again before falling back; on a walk reaching an alias twice; on a second such attachment; with nothing wired; on a gateway image older than 3.12; and on any component other than core/gateway/kong-enterprise. Kong sells AI Proxy Advanced under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips the plugin silently, so the gateway's Recipe halts until infra.ai_gateway_entitled affirms it. No build proves that a running gateway falls back.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/fallback-chain-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain-traefik-hub
Realizes core/fallback-chain on the gateway core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, reads the attachment and sends the route of each alias an entry names through nested failovers of Traefik's own, trying the models of the alias's walk in order, each with its own model name, credential, quotas and retry. The walk is the alias, then its entry's fallbacks in order, each counting one fallback; a fallback's own entry is followed in turn only while fewer than max_hops have been counted, and every alias one entry lists is placed whatever the count. A model's reply of 429 or of 500 to 504, after any retry of that model, moves the request on to the next; any other reply is returned as it came, and the last model's whatever it is. A request body over 1 MiB is refused with HTTP 413. The key check, the guards, the masking and the cache stay on the route, once per request. The build halts where chains names an alias the models socket does not carry, on a walk reaching an alias twice, on a second such attachment, with nothing wired, on a gateway image older than Traefik Hub v3.20, and on any component other than core/gateway/traefik-hub. No build proves that a running gateway falls back.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/fallback-chain-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain-higress
Binds core/fallback-chain on the gateway core/gateway/higress to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no fallback. Higress 2.2.5 sends a failed request on at most once, to one other destination: a route's default backend, a Kubernetes Service, on the reply statuses the route lists, or the one fallback route its console writes for an AI route, whose target is picked once — at random among several, or the single one an ordered fallback allows. Its AI proxy's failover moves between the keys of one connection, not between models. So no request is sent on along a chain of models until one answers, and a fallback's own entry is never followed; the build halts on every attachment, naming why. With nothing wired to the models socket the gateway serves no model route, and it halts saying so. It halts on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/fallback-chain-higress:1.0.0 - v1.0.0
FeatureRecipe
egress-allowlist
Component-scope FeatureRecipe for core/egress-allowlist on Kubernetes: writes one NetworkPolicy for each workload the attached component renders, in the namespace that workload is emitted into, admitting egress to four kinds of destination only — the attached component's own pods, the pods of every component its uses edges point at, the cluster's DNS pods on port 53, and the address ranges the attachment lists. Each component's pods are matched by their labels and by the namespace they are emitted into together, on the ports their own containers declare. Ingress is left untouched. A workload's pods are selected by its controller's own selector or, where it writes none, by the identity labels the build stamps on its pods, never by a label carrying a version. Selectors, namespaces and ports are read from what the renders emitted, never restated by hand; the attachment renders after every component has rendered, for that reason. The build halts unless the Operator affirms the cluster's network plugin enforces network policy, and wherever a policy could not admit exactly what the Aggregate and the attachment declare: naming every connection whose target renders no pods and has no address range listed; a target declaring no port; a workload emitted into no namespace, or selected neither by its controller nor by the build's identity labels; a component rendering no pods; an address-range entry naming no connection, or no range; and DNS pods named by nothing.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/egress-allowlist:1.0.0 - v1.0.0
FeatureRecipe
env-secret
Helm implementation of core/secret-value. Inject-only: the injectMapping merges the value into the attached component's env pipe, so the component Recipe renders it as a container environment variable on the synthesized workload (or feeds a managed chart's auth values). No Kubernetes Secret is generated; a slot read as a secretKeyRef takes core/referenced-secret, or core/secret-value mapped to k8s-secret-from-value. An origin holding a Docker Compose variable reference of the form ${…} halts the build, since Kubernetes substitutes none; set the real value through the Release instead.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/env-secret:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain-agentgateway
Binds core/fallback-chain on the gateway core/gateway/agentgateway to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no virtual model. agentgateway 1.6.0 fails over between models through a virtual model, a name whose targets sit in priority groups. A request is sent to the first group holding a target that has not been evicted, and a failed send evicts that target for the requests that come after it while the request it failed is answered with that failure; a retry beside it chooses again from whichever group is active at that moment, which can be the target that just failed. So no request is sent on to the next model of a chain once its own has failed, which is what the Feature promises, and a fallback's own entry could not be followed in turn either, since a virtual model's targets are model entries rather than other virtual models. The build halts on every attachment, naming why; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/fallback-chain-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
db-binding
Helm implementation of example/db-binding. This FeatureRecipe renders no resources. A component-attached feature is an instance that must resolve to a FeatureRecipe for resolution-completeness, the same way a component resolves to a Recipe. This binding's effect — injecting the database server address into the consumer (db's host provides → ADMINER_DEFAULT_SERVER) — flows through the uses-edge, not through this recipe; so the recipe is intentionally empty.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/db-binding:1.0.0 - v1.0.0
FeatureRecipe
data-masking-kong-enterprise
Realizes core/data-masking on the gateway core/gateway/kong-enterprise; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe writes Kong's AI PII Sanitizer onto each wired model's routes, and the sanitizer sends the text it masks to the Kong PII service wired to the pii_sanitizer socket, failing the request when that service cannot be reached or errs. action redact replaces each match with a numbered placeholder; refuse, from Kong 3.12, refuses a request carrying one. Replies are masked from 3.12, and a request for a streamed reply is then refused. Reply text not sent as the one string the OpenAI API returns passes unread; a wired model not labelled as speaking that API halts the build. Kong's sanitizer reads the text of the user's turns, a completion's prompt and a reply's message text, and refuses or withholds nothing it does not read, so each strict default is refused by name and an attachment builds once it relaxes them, letting through unmasked: under allow_unmasked_text, a request's other turns, tool calls and other fields; under allow_non_text_content, its images, files and audio; under allow_unmasked_reply_parts, a reply's tool calls, refusal, audio, token probabilities and a failed call's reply; under mask_reasoning false, its reasoning. Patterns need placeholder_shows_name false under redact. It refuses restore, refuse on replies, allow_streamed_replies and entity types outside Kong's categories. Kong sells the plugin under its AI Gateway Enterprise entitlement, which no build can verify and without which Kong skips it silently, so the gateway's Recipe halts until infra.ai_gateway_entitled affirms it. No build proves that a running gateway masks.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/data-masking-kong-enterprise:1.0.0 - v1.0.0
FeatureRecipe
data-masking-litellm
Realizes core/data-masking on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes guardrails, one request policy and a setting turning the gateway's logging of message contents off into the configuration the gateway's own Recipe renders, and emits no file. Listed patterns are matched inside the gateway; listed entity types are found and replaced by the Presidio pair wired to its pii_analyzer and pii_anonymizer sockets. Every entry runs on every request; a caller cannot turn it off. By default, while requests are masked, it masks the messages, the tool calls and results a conversation carries and the other text beside them; while replies are, a reply's text, tool calls, added parts and, by entity type, its reasoning, and it withholds what masking cannot read. What its masking does not reach is refused with HTTP 400 before forwarding. action refuse refuses what carries a match instead; allow_unmasked_text and allow_unmasked_reply_parts let through, unmasked, what its comments list. An unreachable or erring recognizer fails the request. While patterns mask replies, a streamed reply is sent whole at the end. It refuses action restore at build time: this gateway numbers its stand-ins afresh in every text, so restoring could hand a caller another person's name. Masking is not sufficient on its own: text no pattern matches and no recognizer detects passes; the gateway and both recognizers see the original; tool definitions, response formats, stop sequences, reasoning a client sends back and the names of JSON keys pass as sent; the gateway's own log and a logging integration can still hold unmasked text; and traffic that does not cross the gateway is not touched.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/data-masking-litellm:1.0.0 - v1.0.0
FeatureRecipe
data-masking-traefik-hub
Realizes core/data-masking on the gateway core/gateway/traefik-hub; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, traefik-hub-chart, writes a Content Guard of Traefik Hub's AI gateway onto the route of each wired model, reading requests, replies or both as scope names. Patterns are matched by its regular-expression engine; entity types are found by the Presidio analyzer on the gateway's pii_analyzer socket, in English. redact replaces a match with a run of * as long as the match; refuse answers HTTP 403 in its place. A reply that is masked or checked is held whole and sent at the end, streamed or not. Traefik's reference does not say which fields of a chat request or reply Content Guard reads, and names no setting that refuses or withholds what it does not read, so each strict default is refused by name and the attachment builds once the Release relaxes it: allow_unmasked_text and allow_non_text_content on requests; allow_unmasked_reply_parts and mask_reasoning false on replies; and allow_streamed_replies with entities on replies, since a streamed reply is held, not refused. What each lets through is what Content Guard does not read. It refuses restore, a placeholder showing a pattern's name, patterns and entities in one attachment, a second attachment, and an entity type the analyzer does not know; and halts with no model wired, with entities but not one analyzer wired, and on any other component. No build proves the gateway masks.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/data-masking-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
data-masking-higress
Realizes core/data-masking on core/gateway/higress; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, higress-chart, writes Higress's ai-data-masking plugin onto the model routes, replacing each listed pattern's matches with [<pattern name>] before the request is forwarded. Patterns are Rust fancy-regex, grok's %{NAME} expanded. The plugin reads each message's content and reasoning while all are plain text, and nothing else. So while allow_unmasked_text is off, every other request is refused with HTTP 403 before it is forwarded: one not sent as JSON or not a chat completion, a message whose content is a list or null, a tool call or tool turn, a message's name, and any parameter but a chat completion's non-text ones and the developer's tool, format and stop settings. Set on, those pass unmasked, a request with a message not holding plain text passing whole. With only allow_non_text_content off, images, files, audio and every request but a chat completion are refused. Whatever it relaxes, a compressed, binary or gRPC body, an upgrade and a body holding U+FFFD, which an invalid UTF-8 byte reads as, get HTTP 403, and a body over 32 KiB 413. A client deliberately sending valid JSON the gateway's validation cannot decode (a number past a double's range) gets past them: only plain message text and reasoning are then masked, escaped text is not, and a body also not valid UTF-8 passes unmasked. Nor is escaped text masked in JSON nested over 127 levels. It refuses entity types, restore, refuse, scope response and both, and allow_non_text_content on while allow_unmasked_text is off. It halts on a second attachment, on an empty models socket and on any other component.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/data-masking-higress:1.0.0 - v1.0.0
FeatureRecipe
context-propagation-envoy-gateway
Component-scope FeatureRecipe for core/context-propagation on the Envoy Gateway ingress gateway, core/gateway/envoy-gateway. It emits one client traffic policy attached to the gateway that component emits, on its `https` listener: the gateway then requests a certificate from every client, verifies it against the certificate authority held in the Secret infra.client_ca_secret names (under the key ca.crt), and turns away a client presenting none or one that does not chain to it. For every request it forwards it replaces any x-forwarded-client-cert header the client sent with one describing the certificate it verified — its subject and the URI and DNS names it carries, beside the certificate's hash, which the gateway always adds — so the service behind the route reads an identity the gateway checked rather than one a client asserted. The listener exists only where the gateway serves HTTPS, which is its tls_secret. The gateway declares dataplane.aggrega.ai/caller-identity: client-certificate, so this recipe halts on any other propagation_mode; it halts when infra.destinations is set, because the policy attaches to the listener and the identity then travels on every route behind it, which a list of destinations cannot narrow; it halts when no certificate authority is named in infra.client_ca_secret, because a certificate is worth forwarding only once it has been verified; and it halts when the attachment is carried by any component other than core/gateway/envoy-gateway, whose gateway is the one the policy names. The policy is named for the component and the attachment, so two attachments on one gateway emit two policies.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/context-propagation-envoy-gateway:1.0.0 - v1.0.0
FeatureRecipe
context-propagation-litellm
Component-scope FeatureRecipe for core/context-propagation on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. The identity it forwards is the gateway's own: the virtual key a caller presents, verified against the keys the gateway issued. So it halts unless the gateway's api_key and database_url credential slots are both attached. The work is done by a request hook the gateway's own Recipe delivers; this recipe carries the checks and emits nothing. The gateway refuses a request whose body carries headers for the upstream, and on every request the hook also removes each x-litellm- header from the parts of the request body the gateway turns into upstream headers, so a caller cannot supply an identity of its own. On a request for a model alias infra.destinations lists, it adds headers naming the user and the team of the verified key, and its organization where the key names one, and nothing else about the key; the gateway release the catalog pins creates organizations only under an enterprise licence, so its keys name none. A request made with the master key names nobody. The headers reach the upstream of a listed alias only, fallbacks included. It halts when infra.destinations is missing or names an alias the model map does not publish, on any propagation_mode but injected_header, and when the attachment is carried by any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/context-propagation-litellm:1.0.0 - v1.0.0
FeatureRecipe
data-masking-agentgateway
Realizes core/data-masking on the gateway core/gateway/agentgateway; a Release binds it in its Manifest. It emits no file: the gateway's chart Recipe, agentgateway-chart, writes agentgateway's regular-expression guards into the configuration, replacing each match of a listed pattern with <masked> — under refuse, refusing the request or the reply with HTTP 403 — in requests, in replies or in both as scope says. While replies are masked a request asking for a streamed reply is refused with HTTP 403 before it is forwarded, since agentgateway does not mask a streamed reply, and so is any request whose body is larger than the 2 MiB the gateway reads for that check, streamed or not. A request guard reads the system and developer prompts, every other message and the arguments and results of tool calls; masking an argument can leave its JSON invalid. A reply guard reads the reply's message text. Nothing else is read or refused, so each strict default is refused by name and an attachment builds once the Release relaxes it, letting through unmasked: under allow_unmasked_text, a message's name and fields beside its text and every request field beside its messages; under allow_non_text_content, images, files and audio; under allow_unmasked_reply_parts, a reply's tool calls, refusal, audio and token probabilities and a failed call's reply; under mask_reasoning false, its reasoning. Patterns need placeholder_shows_name false under redact. It refuses entity types and restore. It halts on a second attachment, with nothing wired to the models socket, and on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/data-masking-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
concurrency-limit-nginx
Component-scope FeatureRecipe for core/concurrency-limit on the NGINX gateway, core/gateway/nginx, on Kubernetes. The cap is a setting inside the server configuration the gateway's own Recipe writes rather than a resource of its own, so this recipe emits no file: the NGINX Recipe reads the attachment and writes a limit_conn keyed on one fixed word, so NGINX counts every request the server is processing as one total, kept in shared memory across its worker processes, and answers a request past max_in_flight with 503. NGINX keeps that one total and no count for each upstream target — the gateway declares dataplane.aggrega.ai/concurrency-cap: total — so this recipe halts on scope model rather than capping the total in its place. It halts on a max_in_flight above 65535, the most limit_conn takes, and when the attachment is carried by any component other than core/gateway/nginx, because no other component's Recipe reads it, and the attachment would render no cap at all.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/concurrency-limit-nginx:1.0.0 - v1.0.0
FeatureRecipe
caller-allowlist
Component-scope FeatureRecipe for core/caller-allowlist on Kubernetes: writes one NetworkPolicy for each workload the attached component renders, in the namespace that workload is emitted into, admitting ingress on the ports its containers declare from three kinds of peer only — the attached component's own pods, the pods of every component whose uses edge points at it, and the sources the attachment names. Each component's pods are matched by their labels and by the namespace they are emitted into together. Egress is left untouched. A workload's pods are selected by its controller's own selector or, where it writes none, by the identity labels the build stamps on its pods, never by a label carrying a version, so a policy keeps matching the pods it describes across upgrades. Selectors, namespaces and ports are read from what the renders emitted, never restated by hand; the attachment renders after every component has rendered, for that reason. The build halts, naming the component, unless the Operator affirms the cluster's network plugin enforces network policy, and wherever a policy could not admit exactly what the Aggregate and the attachment declare: a component or a caller rendering no pods; a workload emitted into no namespace, or selected neither by its controller nor by the build's identity labels; a component declaring no port; a source entry naming both an address range and pods, or neither.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/caller-allowlist:1.0.0 - v1.0.0
FeatureRecipe
concurrency-limit-envoy-gateway
Realizes core/concurrency-limit on the gateway core/gateway/envoy-gateway; a Release binds it in its Manifest. It emits no file: the gateway's Recipe, envoy-gateway, reads the attachment and writes it as the circuit breaker of the one traffic policy on the route serving the models wired to its models socket, the policy that also carries a period budget's rate limit and a retry policy's retry. Envoy Gateway applies that circuit breaker to the backend of each rule of the route, and each wired model has a rule of its own, so each wired model is capped on its own: at most max_in_flight requests are sent to it at once, and a request past the cap is answered 503 and is not sent again; a send the retry policy repeats counts against the cap like the first. The count is each running data-plane replica's own. Envoy's worker threads check the count and then add to it, so requests arriving together can take it briefly above the cap. Envoy Gateway keeps a cap for each backend and none on the total across everything the gateway serves — the gateway declares dataplane.aggrega.ai/concurrency-cap: per-target — so this recipe halts on scope gateway, the Feature's default, rather than capping each model in its place. It refuses a max_in_flight above 4294967295, and halts with nothing wired to the models socket, on a second such attachment, and on any component other than core/gateway/envoy-gateway.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/concurrency-limit-envoy-gateway:1.0.0 - v1.0.0
FeatureRecipe
concurrency-limit-litellm
Realizes core/concurrency-limit on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. Each Recipe of the gateway reads the attachment and writes max_in_flight as the max_parallel_requests of every entry of the model map, typed by hand and wired alike. The proxy keeps each entry's count in memory, where a record lapses ten minutes after it is made and a new count starts at zero while requests counted by the old one are in flight, so the Recipe also sets that lifetime to a century, keeping each count as long as the proxy runs. The proxy takes one of an entry's slots for each request it sends there, so at most max_in_flight requests are in flight to each entry at once. A request finding its entry full is answered 429: after the gateway's retries where its alias has one entry, and at once where it has several, even while another of them has room, since the proxy picks the entry before it counts; a fallback chain attached to the gateway sends it on to its alias's fallbacks instead. The count is each running proxy's own. The proxy keeps no total that holds — its one total is read only by a request limiter an environment variable switches on, which counts a request for its first second — so this recipe halts on scope gateway, the Feature's default. It halts on a model map of more than 100 entries, past which the store the counts are kept in, 200 records, can drop one, on a second such attachment, and on any component other than core/gateway/litellm; and the gateway refuses a request carrying priority, whose queue the century lifetime would keep.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/concurrency-limit-litellm:1.0.0 - v1.0.0
FeatureRecipe
caller-allowlist-higress
Binds core/caller-allowlist on the gateway core/gateway/higress to a halt naming why the gateway cannot honour it; a Release binds it in its Manifest. It emits no file and writes no network policy. The allowlist admits callers on the ports the attached component's containers declare, read from what its render emitted. Higress's gateway container declares only its two telemetry ports; it programs the port that serves model traffic through its own control plane rather than as a container port, and the chart declares that port on the container only in a local or kind layout, and then bound to a host port. So a policy built from the gateway's declared ports would admit no model traffic and refuse every request to the gateway, its declared callers included. The build halts on every attachment, naming why. It halts on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/caller-allowlist-higress:1.0.0 - v1.0.0
FeatureRecipe
backup-kubestash
Component-scope FeatureRecipe for core/backup on Kubernetes through KubeStash: writes a KubeStash BackupConfiguration for the attached component's workload, with the BackupStorage and the RetentionPolicy it reads, beside it. The backups run only where core/storage/kubestash is deployed, under a licence AppsCode issues; nothing in the build checks either. Each backup copies the files under the mount path of the workload's one persistent volume, with restic, into a repository in the bucket the store connection points at or, with infra.volume set instead, on a claim of that size written beside it. It is a copy of the volume's files taken while the datastore runs, not a database-consistent dump, so point-in-time recovery is off: a restore returns to a backup's moment, and infra.point_in_time_recovery: true halts. It runs on infra.schedule, or daily at 03:00 where that is unset, and each backup is kept for infra.retention_days days. The jobs run as infra.run_as_user and infra.run_as_group. The jobs read the repository's password, under RESTIC_PASSWORD, and for a bucket its key pair, under AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY, the names KubeStash fixes, from the Secret infra.credentials_secret names or, unset, one named after the BackupConfiguration, in the component's namespace. The build halts, naming the attachment, with neither a store connection nor infra.volume, with a store publishing no S3 API, on a second core/backup attachment, on a component rendering not exactly one Deployment, StatefulSet or DaemonSet mounting a persistent volume at one path, on a volume for a DaemonSet or of a size that is not a Kubernetes quantity, and on infra.role_arn, which KubeStash has no use for.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-kubestash:1.0.0 - v1.0.0
FeatureRecipe
backup-mariadb-operator
Component-scope FeatureRecipe for core/backup on a MariaDB database realized by helm/mariadb-cluster: writes a mariadb-operator Backup beside the database's MariaDB, which the mariadb-operator controller runs as a CronJob. The backups run only where that controller is deployed; nothing in the build checks it. Each backup is a logical dump of every database on the server, taken with mariadb-dump on infra.schedule, or daily at 03:00 where that is unset, in the time zone of the cluster's CronJob controller; each backup run deletes the backups older than infra.retention_days days. It is written to the bucket the store connection points at, signing with the key pair in the Secret infra.credentials_secret names or, unset, in one named after the Backup, which the cluster must hold in the database's namespace; or, with infra.volume set instead, to a ReadWriteOnce claim of that size which the controller provisions beside the database. Point-in-time recovery is off: a restore returns to a dump's moment, and infra.point_in_time_recovery: true halts. The build halts, naming the attachment, on a database realized by another Recipe, with neither a store connection nor infra.volume, with a store publishing no S3 API, with a store in the cluster whose namespace is not declared, on a database carrying a second core/backup attachment, on a Backup name longer than a CronJob's name may be, on a volume size that is not a Kubernetes quantity, on infra.credentials_secret beside a volume, on infra.role_arn, the role AWS Backup assumes, which the Backup has no use for, and on infra.run_as_user and infra.run_as_group, since it writes no security context on the Backup.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-mariadb-operator:1.0.0 - v1.0.0
FeatureRecipe
backup-velero
Component-scope FeatureRecipe for core/backup on Kubernetes through Velero: writes a Velero Schedule and the BackupStorageLocation it writes to, both in the namespace velero, where core/storage/velero runs. The schedule runs only where that component is deployed; nothing in the build checks it. Each backup holds the attached component's pods, selected by the selectors of the controllers its render emits, in the namespaces they are emitted into; the persistent volume claims and volumes those pods mount; and the data of the pods' volumes, which Velero's node agent copies into the bucket. It runs on infra.schedule, or daily at 03:00 in the time zone of Velero's server where that is unset, and each backup is kept for infra.retention_days days. Point-in-time recovery is off: a restore returns to a backup's moment, and infra.point_in_time_recovery: true halts. The location is the bucket the store connection points at, under a prefix of the schedule's own name, reached at the endpoint and region the store publishes. The build halts, naming the attachment, with no store connection, with a store publishing no S3 API, with a store in the cluster whose namespace is not declared, on a component carrying a second core/backup attachment, on a component whose pods are emitted into no namespace, and on a component whose render emits no workload Velero can select, naming the realization to bind instead. It refuses infra.credentials_secret, infra.role_arn, infra.volume, infra.run_as_user and infra.run_as_group by name: Velero signs with the key core/storage/velero holds, assumes no role an attachment names, keeps backups only in a bucket, and starts no job an attachment writes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-velero:1.0.0 - v1.0.0
FeatureRecipe
backup-kasten
Component-scope FeatureRecipe for core/backup on Kubernetes through Veeam Kasten: writes a Kasten policy and the location profile it exports to, both in the namespace kasten-io, where core/storage/kasten runs. The policy runs only where that component is deployed; nothing in the build checks it. Each run snapshots the attached component's controllers, the resources carrying their selector labels and the volume claims they mount, then exports the snapshot, volume data included, to the bucket the store connection points at or, with infra.volume set instead, to a ReadWriteMany claim of that size written beside the profile. Kasten takes volume snapshots through a CSI snapshot class marked for it. It runs on infra.schedule, read in UTC, or daily at 03:00 UTC; Kasten keeps backups by count, so a schedule must run once a day or once an hour at a fixed minute, and each backup is kept for infra.retention_days days. Point-in-time recovery is off: a restore returns to the moment of a snapshot, and infra.point_in_time_recovery: true halts by name. For a bucket, Kasten reads the key pair under aws_access_key_id and aws_secret_access_key from the Secret in kasten-io that infra.credentials_secret names or, unset, the one named after the policy. The build halts, naming the attachment, with neither a store connection nor infra.volume, with a store publishing no S3 API, on any other schedule, on a second core/backup attachment, on a component rendering no workload or none in a namespace, and on infra.role_arn, infra.run_as_user and infra.run_as_group, which a Kasten policy has no place for.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-kasten:1.0.0 - v1.0.0
FeatureRecipe
aws-irsa
Component-scope FeatureRecipe for core/aws-irsa: emits one Kubernetes ServiceAccount per attached component from a co-located chart, annotated with the operator-set IAM role ARN for IAM Roles for Service Accounts (IRSA). An operator-set service_account_name is used exactly as written, so a name an IAM role's trust policy already names is never rewritten; left unset, the name composes the host component's rendered name — the name the helm provider spells in the characters a Kubernetes name allows — with the name the attachment is carried under, so attaching this Feature to several components in one namespace emits a distinct ServiceAccount for each. The name is published as provides so the operator can mirror it onto the consumer's aggrega.ai/service-account annotation. Resource Generation via source + valuesMapping; the per-component .context.* root is available at component scope.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/aws-irsa:1.0.0 - v1.0.0
FeatureRecipe
backup-cloudnative-pg
Component-scope FeatureRecipe for core/backup on a PostgreSQL database realized by helm/postgres-cloudnative-pg, through the Barman Cloud plugin core/storage/barman-cloud-plugin installs: writes the plugin's ObjectStore and a CloudNativePG ScheduledBackup beside the database's Cluster, which names the plugin as its write-ahead-log archiver. The backups run only where the plugin is deployed; nothing in the build checks it. The Cluster archives its write-ahead log into the bucket the store connection points at, continuously, and the schedule takes a base backup there on infra.schedule, or daily at 03:00 where that is unset. Backups and log older than infra.retention_days days are deleted, except the newest base backup taken before that window, which a restore into it needs. That archive is CloudNativePG's point-in-time recovery, on here, so infra.point_in_time_recovery: false halts by name. The plugin signs its requests with the key pair in the Secret infra.credentials_secret names or, unset, in one named after the ObjectStore, which the cluster must hold in the database's namespace. The build halts, naming the attachment, on a database realized by another Recipe, with no store connection, with a store publishing no S3 API, with a store in the cluster whose namespace is not declared, on a database carrying a second core/backup attachment, on infra.role_arn, the role AWS Backup assumes, which the plugin has no use for, on infra.volume, since the plugin keeps backups only in a bucket, and on infra.run_as_user and infra.run_as_group, since it runs inside the database's own pods.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-cloudnative-pg:1.0.0 - v1.0.0
FeatureRecipe
backup-k8up
Component-scope FeatureRecipe for core/backup on Kubernetes through K8up: writes a K8up Schedule beside the attached component. The schedule runs only where core/storage/k8up is deployed; nothing in the build checks it. Each backup copies the files of the component's persistent volume claims, selected by its workloads' labels, into a restic repository in the bucket the store connection points at or, with infra.volume set instead, on a claim of that size written beside the Schedule. It is a copy of the volumes' files taken while the datastore runs, not a database-consistent dump, so point-in-time recovery is off: a restore returns to a backup's moment, and infra.point_in_time_recovery: true halts. It runs on infra.schedule, or daily at 03:00 where that is unset; K8up keeps backups by count, so a schedule must run once a day or once an hour at a fixed minute, and each backup is kept for infra.retention_days days. The jobs run as infra.run_as_user and infra.run_as_group. The backup jobs read the repository's password, under RESTIC_PASSWORD, and for a bucket its key pair, under ACCESS_KEY_ID and ACCESS_SECRET_KEY, from the Secret infra.credentials_secret names or, unset, from one named after the Schedule, in the component's namespace. The build halts, naming the attachment, with neither a store connection nor infra.volume, with a store publishing no S3 API, on any other schedule, on a second core/backup attachment, on a component rendering no workload whose claims a schedule can select, on a volume for a DaemonSet or of a size that is not a Kubernetes quantity, and on infra.role_arn, which K8up has no use for.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/backup-k8up:1.0.0 - v1.0.0
FeatureRecipe
audit-logging
Realizes core/audit-logging on the AI gateway core/gateway/litellm; a Release binds it in its Manifest. It writes the gateway's request-log settings into the configuration the gateway's Recipe renders, and emits no file. The gateway keeps a row for each model request it served in its bookkeeping database, naming the model that answered and the tokens the reply reports, and it honours every value. The row says what the request cost in US dollars, priced from the table the gateway keeps per model; a model the table does not price is recorded as costing nothing. It names the key the gateway verified, as the hash of a key the gateway issued or as a fixed name for its administrative key, with the user and team the key belongs to; the end user a row names is whatever the request writes, unchecked. Where retention_days is given, the gateway deletes a row once it is that many days old, and otherwise it deletes none. The gateway writes no row without its bookkeeping database, and without an administrative key it checks no caller and writes no row for a request that presents no key and names no end user, so the build halts unless its database_url and api_key credential slots are both attached. It halts when the attachment is carried by any component other than core/gateway/litellm. While the gateway's values.log_masking lists what to mask, its request log keeps answered requests only: failed and refused requests are counted in the gateway's daily request count but not recorded per key, model or caller.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/audit-logging:1.0.0 - v1.0.0
FeatureRecipe
audit-logging-higress
Realizes core/audit-logging on the gateway core/gateway/higress; a Release binds it in its Manifest. It emits no file of its own: while the attachment is present, the gateway's chart Recipe, higress-chart, adds Higress's ai-statistics plugin to the model routes, and the access log the gateway writes to its standard output then carries, in each model request's line, an ai_log field naming the model that answered and the input, output and total tokens its reply reports, beside the request's path, status, timing and upstream. It keeps no prompt or reply text. So it realizes an attachment that turns require_cost and require_verified_caller off and gives no retention_days. It refuses by name require_cost on, since the gateway prices no request, require_verified_caller on, since the gateway declares no callers and so verifies none, and retention_days, since the gateway keeps no record it could delete: how long a line is kept is the log collection's own. It halts with nothing wired to the models socket, and on any component other than core/gateway/higress.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/audit-logging-higress:1.0.0 - v1.0.0
FeatureRecipe
audit-logging-traefik-hub
Binds core/audit-logging on the gateway core/gateway/traefik-hub to a halt naming why the gateway cannot keep it; a Release binds it in its Manifest. It emits no file and writes no record. Traefik Hub keeps a record of a model request only as an OpenTelemetry span its AI middlewares export to a tracing backend, naming the model and, at the span's detailed level, the tokens and their estimated cost in US dollars. The span names no caller, so a record naming the caller the gateway's key check verified, the default, cannot be kept; and this realization wires no tracing backend, so even a record naming no caller would land nowhere. The gateway keeps no record it could delete. The build halts on every attachment, naming why; with nothing wired to the models socket it halts saying the gateway serves no model route. It halts on any component other than core/gateway/traefik-hub.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/audit-logging-traefik-hub:1.0.0 - v1.0.0
FeatureRecipe
audit-logging-envoy-gateway
Realizes core/audit-logging on the gateway core/gateway/envoy-gateway as the gateway's access log; a Release binds it in its Manifest. It emits no file of its own: while the attachment is present, the gateway's Recipe names the access log from the gateway and has its model route report the token counts. The record lands on the data plane's standard output, one JSON line per request, and the access log keeps nothing itself: how long a line lasts is up to whatever collects that output. A line carries the fields Envoy Gateway writes by default, the model the request named, the model that answered and the prompt, completion and total tokens the reply reports; a request the model did not answer carries no model that answered and no tokens. No field names a caller the gateway verified: the client's address and the headers it sends are written as they arrive, unchecked. So the build halts, naming each value it cannot keep: require_cost while on, since the per-request cost Envoy AI Gateway can compute has no unit and the gateway keeps no prices; require_verified_caller while on, since nothing on this path verifies who the caller is; and retention_days where given, since the access log has no store and no retention setting. An attachment at the defaults is told both strict values in one failure. It halts with nothing wired to the models socket, and on any component other than core/gateway/envoy-gateway.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/audit-logging-envoy-gateway:1.0.0 - v1.0.0
FeatureRecipe
audit-logging-agentgateway
Realizes core/audit-logging on the gateway core/gateway/agentgateway; a Release binds it in its Manifest. It emits no file of its own: while the attachment is present, the gateway's chart Recipe, agentgateway-chart, gives the gateway its store, a SQLite file on a volume it renders, with one replica, and the gateway writes there a record of each model request that passes its key, budget and authorization checks, a guard's refusal included. A record names the model the gateway asked the model server for — the name the wired model publishes, not the alias the caller sent — and the model that answered, the input, output and total tokens the reply reports, the response's status and timing, and, for a caller infra.callers declares, that caller's name, from the API key the gateway verified. It keeps no prompt or reply text. The records are read through the gateway's admin interface, which listens inside its pod. So it honours require_verified_caller while callers are declared, and refuses by name require_cost while on, since no model wired to the gateway is priced, require_verified_caller while on with no caller declared, and retention_days, since the gateway deletes no record. It halts with nothing wired to the models socket, and on any component other than core/gateway/agentgateway. The compose realization of the gateway has no realization of it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/helm/audit-logging-agentgateway:1.0.0 - v1.0.0
FeatureRecipe
upstream-key-pinning-tyk
Component-scope FeatureRecipe for core/upstream-key-pinning on the Tyk gateway, core/gateway/tyk, under Docker Compose. The pins are a setting the gateway reads rather than a service of their own, so this recipe emits nothing: the gateway's Recipe reads the attachment and hands every entry to the gateway's container as the one environment variable the gateway takes its upstream pins from. An upstream is the host the gateway dials, `*.example.com` for every host one label under a suffix, or `*` for every upstream; the most specific entry applies, and it replaces the `*` entry rather than adding to it. Setting any pin changes how the gateway reaches every upstream: from then on it checks an upstream's key rather than the chain the upstream presents, so an upstream no entry covers is reached with its certificate not verified at all. `*` covers every upstream with no entry of its own and holds each of them to its key. It halts on an upstream given more than one key, on an upstream holding a comma, a colon or a misplaced `*`, and when the attachment is carried by any component other than core/gateway/tyk.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/upstream-key-pinning-tyk:1.0.0 - v1.0.0
FeatureRecipe
traffic-shadowing-nginx
Component-scope FeatureRecipe for core/traffic-shadowing on the NGINX gateway, core/gateway/nginx, under Docker Compose. The copy is a setting inside the server configuration the gateway's own Recipe writes rather than a service of its own, so this recipe emits nothing: the NGINX Recipe reads the attachment and adds the mirror to the location it forwards, sending a copy of every forwarded request, with its original URI, to the service the attachment's infra names, and discarding that service's response. NGINX copies every request it forwards and has no setting that copies a share of them — the gateway declares dataplane.aggrega.ai/request-mirroring: full-rate — so this recipe halts on a shadow_percent below 100 rather than copying more traffic than was asked for. It also halts when the attachment is carried by any component other than core/gateway/nginx, because no other component's Recipe reads it, and the attachment would render no copy at all. The service the copies go to has to resolve on the compose network when NGINX starts, as the upstream does.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/traffic-shadowing-nginx:1.0.0 - v1.0.0
FeatureRecipe
prompt-injection-shield-nginx
Component-scope FeatureRecipe for core/prompt-injection-shield on the NGINX gateway, core/gateway/nginx, under Docker Compose. It emits no file: it points the proxy_pass target the gateway's Recipe writes at the guard wired to the gateway's guard socket, in place of the component on its upstream socket, so every request the gateway forwards goes to the guard. The guard checks each chat request and forwards what it admits to the model its own infra.upstream_endpoint names, which the Release sets to the model the upstream socket names; it answers no completions, embeddings, responses or messages route, and its model listing passes the caller's Authorization header to the model's. The guard must be core/security/nemo-guardrails running the threshold the sensitivity names, 0.5 for standard and 0.001 for strict, the values measured for its detector's default classifier. The build halts otherwise, with nothing on the guard socket, naming it, with more than one guard, on a second attachment, beside core/traffic-shadowing, whose copies would bypass the guard, and on any component other than core/gateway/nginx. The guard reads only the latest user message, and judges no reply, tool-call argument or image. A request it refuses is answered with an ordinary completion, streamed when asked; when the guard or its detector cannot be reached the request is refused and nothing reaches the model.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/prompt-injection-shield-nginx:1.0.0 - v1.0.0
FeatureRecipe
provided-secret
Compose realization of core/secret-provided. origin — from values or infra, whichever carries it — is published as the `value` provides key, and the attachment dictates where it is delivered with its own `inject:` block on a transparent slot — this realization carries no injectMapping of its own, so an attachment without that block delivers the value nowhere. On compose the delivered value is written into the emitted docker-compose file as a literal environment value, exactly as authored: a literal origin puts the material in the rendered artifact, and an origin written as a Compose variable reference, such as ${DB_PASSWORD}, lands as that reference, which Docker Compose substitutes when the stack starts. The name provided-secret is shared with the helm realization of the same Feature, so a Manifest references this one provider-qualified, as compose/provided-secret:1.0.0. A bare provided-secret:1.0.0 matches both and resolves to neither, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/provided-secret:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-litellm
Realizes core/prompt-deny-patterns on the AI gateway core/gateway/litellm under Docker Compose; a Release binds it in its Manifest, as compose/prompt-deny-patterns-litellm. It emits no file: the LiteLLMGatewayConfig bundle the gateway's compose Recipe imports reads the attachment and adds one entry to the gateway's guardrails, turning on its built-in content filter before a request is forwarded, with one blocking rule per entry of deny_patterns, on every request, so a caller cannot turn it off from the request. A request any expression matches, ignoring letter case, in the text of any message, is refused with HTTP 400 naming the rule by its position in the list — deny_pattern_1 for the first — and not the expression. Images are not examined; an administrator can exempt a key or a team from the gateway's always-on guardrails; and a request of a kind the gateway maps to no guardrail is forwarded unchecked. It halts on allow_patterns, because this gateway's filter only refuses a request that matches, and on any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/prompt-deny-patterns-litellm:1.0.0 - v1.0.0
FeatureRecipe
prompt-deny-patterns-kong
Realizes core/prompt-deny-patterns on the gateway core/gateway/kong under Docker Compose; a Release binds it in its Manifest. It emits no file: the KongModelRoutes bundle the gateway's compose Recipe imports reads the attachment and writes Kong 3.9.3's AI Prompt Guard onto the chat and completion routes of each model wired to the gateway's models socket. The guard reads the text of every message of every role across the whole conversation, joined into one string with a space after each message, so an expression can match across two of them, and a completion's prompt. A message whose content is not a single string, such as content in parts or a turn carrying only tool calls, refuses the whole request. Letter case counts unless an expression says otherwise, as (?i) does. A request any of deny_patterns matches is refused whatever allow_patterns says; with allow_patterns set, a request none of them matches is refused. A refused caller receives HTTP 400 with the message "bad request", which names no expression, and so does a request the guard cannot read, such as a body that outgrows Kong's 8 KiB request buffer and is set aside on disk, which the gateway refuses without the guard as well. The build does not check an expression: one Kong cannot compile answers the requests reaching it with HTTP 500. So it honours deny_patterns and allow_patterns, and halts on a list of more than 10 expressions or an expression over 500 bytes, on a second such attachment, with nothing wired to the models socket, and on any component other than core/gateway/kong. The gateway imports its routes into its database, which keeps the guard after the attachment is removed, until it is deleted through Kong's Admin API.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/prompt-deny-patterns-kong:1.0.0 - v1.0.0
FeatureRecipe
manual-trigger
Realizes core/manual-trigger on Docker Compose, on core/compute/task's compose Recipe, compose/task; a Release binds it in its Manifest. It emits nothing: the host's Recipe carries the change, giving the task's service the profile manual, so docker compose up leaves it stopped, because a FeatureRecipe can neither add a key to the host's service nor withdraw the run the host would start. A run is started by hand with docker compose run --rm <service>, the service named after the component. A task with a schedule is refused by compose/task whether or not it carries the attachment, as Docker Compose has no scheduler.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/manual-trigger:1.0.0 - v1.0.0
FeatureRecipe
dotenv-secret
Compose realization of core/secret-value, for local development. It delivers origin — from values or infra, whichever carries it — under values.key on the consuming component's inject pipe, and the compose Recipe writes that value into the emitted docker-compose file as the service's environment value, exactly as authored. A literal origin therefore lands in the file as the credential itself. An origin written as a Compose variable reference, such as ${DB_PASSWORD}, lands as that reference instead, and Docker Compose substitutes it when the stack starts — from the shell environment, a file named with --env-file, or a .env file in the project directory — and substitutes an empty string, with a warning, for a variable it cannot resolve and that names no default. The compiler neither reads nor writes a .env file and emits no env_file entry; whoever starts the stack maintains the .env file.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/dotenv-secret:1.0.0 - v1.0.0
FeatureRecipe
fallback-chain
Realizes core/fallback-chain on the AI gateway core/gateway/litellm under Docker Compose; a Release binds it in its Manifest, as compose/fallback-chain. It emits no file: the LiteLLMGatewayConfig bundle the gateway's compose Recipe imports reads the attachment and writes the chain and its bound into the gateway's configuration. Once the gateway gives up on the alias a request asked for, it sends the request to the aliases the alias's entry lists, in order, until one answers, and a fallback that fails follows its own entry in turn. Each alias tried after the first counts one fallback; a fallback's own entry is followed only while fewer than max_hops have been counted, and every alias one entry lists is tried whatever the count. The build halts where chains names an alias the gateway's model map does not publish, on a second such attachment, and on any component other than core/gateway/litellm.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/fallback-chain:1.0.0 - v1.0.0
FeatureRecipe
concurrency-limit-nginx
Component-scope FeatureRecipe for core/concurrency-limit on the NGINX gateway, core/gateway/nginx, under Docker Compose. The cap is a setting inside the server configuration the gateway's own Recipe writes rather than a service of its own, so this recipe emits nothing: the NGINX Recipe reads the attachment and writes a limit_conn keyed on one fixed word, so NGINX counts every request the server is processing as one total, kept in shared memory across its worker processes, and answers a request past max_in_flight with 503. NGINX keeps that one total and no count for each upstream target — the gateway declares dataplane.aggrega.ai/concurrency-cap: total — so this recipe halts on scope model rather than capping the total in its place. It halts on a max_in_flight above 65535, the most limit_conn takes, and when the attachment is carried by any component other than core/gateway/nginx, because no other component's Recipe reads it, and the attachment would render no cap at all.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/concurrency-limit-nginx:1.0.0 - v1.0.0
FeatureRecipe
egress-allowlist
Binds core/egress-allowlist under Docker Compose to a halt naming why compose cannot honour it; a Release binds it in its Manifest. It emits no file and restricts nothing. The Feature lets a component connect only to the components it is wired to, which takes a network policy over the component's outbound connections. Compose has no network policy: every service of a group shares one network and reaches every other service on it, and a service's own fragment has nothing to say where its connections may go. So the build halts on every attachment, naming the Feature, rather than render the component with every connection still open while the Release reads as restricted.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/egress-allowlist:1.0.0 - v1.0.0
FeatureRecipe
compose-bridge-fr
Network bridge for compose, between the networks of one compose file (each boundary's aggrega.ai/namespace, else the file's own network). Runs once per boundary attachment (executionScope: boundary) and renders one nginx-stream gateway named "aggrega-gw" into the boundary's own compose file. Its templates compute everything from the boundary's uses edges, so no address is hand-authored: an edge is bridged when both of its ends are assembled into the gateway's own file and sit on different networks there. Each bridged edge gets a `server { listen <port>; proxy_pass <target> }` block, the gateway joins the networks of every bridged end, and each consumer's variables delivered from the target's address are rewritten to reach it through the gateway.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/compose/compose-bridge-fr:1.0.0 - v1.0.0
FeatureRecipe
storage-alarm
CloudFormation realization of example/storage-alarm. Component-scope: it emits one CloudWatch alarm per attached component, watching the free storage of the RDS instance that component's own Recipe describes. The alarm is a fragment like any other, so it assembles into the same stack as its host component rather than into a template of its own. The alarm's own logical ID carries the identity of the scope that carries the attachment: the host component and the name the attachment is carried under, together. Composing the host alone would be enough to keep the alarm clear of the resource it watches, and not enough to keep this Feature attached twice to one component from colliding with itself — so both halves are composed. The dimension identifying which instance to watch reads the host's resource by the logical ID the standard-library CloudFormation Recipes derive — both halves derive it from the same component instance name, which is what makes them agree. Ref on an RDS instance resolves to that instance's identifier at deploy time; the compiler pipes the reference through without resolving it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/storage-alarm:1.0.0 - v1.0.0
FeatureRecipe
random-password
CloudFormation realization of core/random-password: an AWS::SecretsManager::Secret that generates its own string. The fragment asks Secrets Manager to generate a password into a JSON secret string under values.key when the stack is created, so the rendered template names the secret and carries no value. The fragment assembles into the same stack as the component carrying the attachment. The secret's logical ID composes the host component and the attachment name — or values.name, where one is supplied — with every separator read as a word boundary and each word capitalised, since a logical ID admits only letters and digits. Where values.name is supplied it is also the secret's own name in Secrets Manager; otherwise CloudFormation names the secret. The published provides are a reference that resolves to the secret's ARN at deploy time, and the JSON key holding the password — the two parts a consumer's Secrets Manager dynamic reference names. The password is generated once, when the secret is created. A change to values.length or values.symbols changes the generation settings, and CloudFormation then creates a new secret version with a new password. Secrets Manager is part of the AWS account the stack deploys into, so this realization reads no prerequisite affirmation. The name random-password is shared with the terraform realization of the same Feature, so a Manifest references this one provider-qualified, as cloudformation/random-password:1.0.0. A bare random-password:1.0.0 matches more than one and resolves to none, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/random-password:1.0.0 - v1.0.0
FeatureRecipe
shared-config
CloudFormation realization of core/shared-config. Boundary-scope (executionScope: boundary): it renders once per boundary rather than once per component, emitting a single parameter that carries the settings every component in that boundary shares. The parameter is a fragment like any other, so it assembles into the same stack the boundary's components do — one parameter in the stack, named for the boundary, rather than one per component. At boundary scope the template reads the boundary's own resolved values; the per-component context is not available, because there is no single component the rendering belongs to. The resource's logical ID is derived from the boundary name with a suffix of its own, so it stays clear of every component resource in the stack it lands in.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/shared-config:1.0.0 - v1.0.0
FeatureRecipe
generated-secret
CloudFormation realization of core/generated-secret. Pure wiring: it declares nothing but the capability it binds to and the Feature it implements, and renders no resource of its own. The credential is provisioned and rotated by the target the CloudFormation Recipe manages — a managed database Recipe sets the provider's own managed-password directive — so there is no value to inject and no secret to emit. Its whole contribution is that the slot the component declares is bound, and bound by the form that says the engine owns the credential. The name generated-secret is shared with the terraform and helm realizations of the same Feature, so a Manifest references this one provider-qualified, as cloudformation/generated-secret:1.0.0. A bare generated-secret:1.0.0 matches more than one and resolves to none, and the pull halts on it.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/generated-secret:1.0.0 - v1.0.0
FeatureRecipe
manual-trigger
Realizes core/manual-trigger on core/compute/task's CloudFormation Recipe, cloudformation/task-ecs; a Release binds it in its Manifest. It emits nothing: the host's Recipe carries the change, deploying a task with no schedule as its task definition alone, without starting it. A scheduled task runs on its schedule as it does without the attachment. A run is started by hand by the operator with aws ecs run-task, giving the cluster, the task definition, named after the component, and the subnets, security groups and public IP setting the Release sets under the task's infra.overrides.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/manual-trigger:1.0.0 - v1.0.0
FeatureRecipe
backup-aws-backup
Component-scope FeatureRecipe for core/backup on CloudFormation through AWS Backup: writes a backup vault, a backup plan and a selection naming the attached database's Amazon RDS DB instance, as a fragment assembled into its host's stack. It backs up an instance realized by cloudformation/postgres-rds, cloudformation/mysql-rds or cloudformation/mariadb-rds. Each backup is a snapshot of the instance that AWS Backup takes on infra.schedule, read in UTC and started within an hour of that time, or daily at 05:00 UTC where it is unset, and keeps in the vault for infra.retention_days days. AWS Backup keeps a continuous backup, its point-in-time recovery, only where a rule asks for one, so left out or false it keeps snapshots alone; infra.point_in_time_recovery true asks, for at most 35 days. AWS Backup runs each backup under the existing IAM role infra.role_arn names, which the attachment must set. The build halts, naming the attachment, without infra.role_arn, on a host this realization cannot name — one realized by another Recipe, or a datastore AWS Backup does not protect, such as an OpenSearch Service domain or an Amazon MQ broker — on a schedule AWS Backup cannot keep, such as one starting more than once an hour, on a retention over 36500 days, on a store connection, infra.volume or infra.credentials_secret, since AWS Backup keeps backups in its vault, on infra.run_as_user or infra.run_as_group, since it runs no job in a cluster, and on a continuous backup kept over 35 days or beside infra.keep_automated_backups false. The instance keeps its automated backups beside these, for 7 days or as long as the continuous backup, unless the attachment sets infra.keep_automated_backups to false.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/backup-aws-backup:1.0.0 - v1.0.0
FeatureRecipe
backup-native
Component-scope FeatureRecipe for core/backup on CloudFormation through the managed database service's own automated backup. It renders nothing: the Recipe realizing the attached database configures the backup from the attachment, because the backup is a property of the database that Recipe writes, and a FeatureRecipe writes a fragment of its own and cannot set one on its host. It binds on a database realized by cloudformation/postgres-rds, cloudformation/mysql-rds or cloudformation/mariadb-rds, which keeps infra.retention_days and a daily infra.schedule as the instance's automated backups and refuses what RDS cannot keep. infra.point_in_time_recovery is the host's Recipe's to keep, with Amazon RDS's own default: RDS keeps it for as long as it keeps automated backups, so left out or true is what it gives and false halts by name. The build halts, naming the attachment, on a host whose Recipe configures no backup of its own from core/backup, on infra.role_arn, which AWS Backup reads, on a store connection, infra.volume or infra.credentials_secret, since the service keeps its backups in its own storage, and on infra.run_as_user or infra.run_as_group, since it runs no job in a cluster.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/featurerecipe/cloudformation/backup-native:1.0.0 - v1.0.0
Component
example/gateway/nginx-reviewed
An NGINX gateway in the shape an organization's own architecture review settled on, published once so every team deploys that shape instead of restating it. It extends core/gateway/nginx and adds only what the review produced: the reviewed-tier identity a selector can discriminate on, the data zone the review covers, who to ask about it, and one extra topology socket whose selector admits only components carrying the same reviewed-tier label. Its contract — values, infra and provides — is the base's, inherited exactly as the base authored it, so every Recipe written for core/gateway/nginx serves this component unchanged and it needs no Recipe of its own. Consuming it takes one Manifest mapping line and nothing else.
ОфіційнийgatewayЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/gateway/nginx-reviewed:1.0.0 - v1.0.0
Component
example/storage/external-object-store
An object-storage bucket operated outside this deployment, reached with a static access-key pair. The deployment provisions nothing: it records which bucket it uses and holds the credential it presents. Because the bucket's operator is not a cloud API this deployment talks to, there is no managed secret to delegate the credential to, so the secret half of the key pair is supplied as a value and declared sensitive here. That declaration is what a Terraform realization turns into the engine's own sensitivity attribute on the matching module input.
ОфіційнийstorageЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/storage/external-object-store:1.0.0 - v1.0.0
Component
example/compute/service
A generic HTTP service exposing an application port and a named `metrics` port. Components of this type must be monitored: the requiredType annotation below states that every instance carries an attachment of the exporter Feature, satisfied here by the monitoring annotation the Operator sets rather than by a per-component features: block.
ОфіційнийcomputeЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/compute/service:1.0.0 - v1.0.0
Component
example/compute/greeter
A trivial HTTP greeter service. It exists to demonstrate one capability: the Manifest per-instance recipe override. Two instances of this single component type can resolve to two different Recipes — the developer authors identical intent for both, and Ops decides per instance which physical implementation each gets.
ОфіційнийcomputeЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/compute/greeter:1.0.0 - v1.0.0
Component
example/compute/gcp-credentialed-service
A service that authenticates to Google Cloud with a static service-account JSON key delivered as a mounted file. It declares a `gcp_key` secret slot whose attached credential materializes as a referenceable Secret; the Recipe mounts that Secret as a file and points GOOGLE_APPLICATION_CREDENTIALS at the mounted path, the location the Google client libraries read the key from. The credential is supplied by attaching a referenceable-secret to the slot, so the key material is never authored on the component and the same slot works with any secret backend the target maps it to.
ОфіційнийcomputeЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/compute/gcp-credentialed-service:1.0.0 - v1.0.0
Component
example/compute/db-reader
A minimal application that consumes a database connection. It declares a `database` uses-socket for any stateful-server datastore and injects the matched database's published host onto its injection context. Its Recipe stores that host as a whole-scalar value, so a CloudFormation Recipe can receive the database endpoint as an intrinsic reference — an intra-stack Fn::GetAtt when the database assembles into the same stack, or a cross-stack import when the database assembles into another stack.
ОфіційнийcomputeЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/example/compute/db-reader:1.0.0 - v1.0.0
Component
core/storage/velero
Velero, the cluster backup controller, installed once per cluster in the namespace velero. On the schedules written for it, it copies Kubernetes resources, and the data of the volumes their pods mount, into a bucket reached over the S3 API, and restores them from there. It backs nothing up by itself. A datastore is backed up by attaching core/backup to it in the Release, with the Manifest binding the helm/backup-velero FeatureRecipe, which writes the schedule and its storage location into the velero namespace; they run only where this component is deployed. The Release declares the namespace velero for it. Velero signs its requests to the bucket with the key pair in the credentials file the cloud_credentials slot names; with the slot empty it uses whatever identity its pods' environment offers, such as the role of the node they run on. Realized on Kubernetes by helm/velero-chart, the project's own chart, whose chart repository must be reachable from the build.
ОфіційнийstorageЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/velero:1.0.0 - v1.0.0
Component
core/storage/kasten
Veeam Kasten, a backup controller for Kubernetes, installed once per cluster in the namespace kasten-io. On the policies written for it, it snapshots the resources and volumes of the workloads they select and exports them to a location profile: a bucket reached over the S3 API, or a volume claim. It restores them from there. It backs nothing up by itself. A datastore is backed up by attaching core/backup to it in the Release, with the Manifest binding the helm/backup-kasten FeatureRecipe, which writes the policy and its location profile into kasten-io. The Release declares the namespace kasten-io for this component. Kasten is Veeam's software, run under Veeam's end-user licence agreement, which whoever installs it accepts. Its chart carries a free edition licence, for five worker nodes after the first 30 days, and a Kasten licence of the reader's own goes into a Secret in kasten-io. The compiler validates no licence and contacts no vendor. Kasten's data mover serves its repository over TLS with a certificate and key it reads from two Secrets in kasten-io, kopia-tls-cert and kopia-tls-key, which the operator creates in the cluster. The required data_mover_cert and data_mover_key slots name them, each filled by a credential naming that existing Secret, so no key is in a rendered file; the build halts while either is unfilled. Realized on Kubernetes by helm/kasten-chart, Kasten's own chart, whose chart repository must be reachable from the build.
ОфіційнийstorageЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/kasten:1.0.0 - v1.0.0
Component
core/storage/s3-object-store
One object-storage bucket reached over the S3 API with an access key. A consumer points its S3 client at the published endpoint, signs for the published region and reads the published bucket, which is the bucket the store holds, so a consumer naming a bucket of its own checks it against this one. A consumer holds the key it presents on its own connection. The two key slots are a self-hosted server's own account, which SeaweedFS requires; a managed service issues its keys to its own principals outside the build, and its realization refuses an attached slot by name. Reach for it where the consumer speaks the S3 API with an access key, wherever it runs; and for core/storage/gcs where it uses Google's client libraries with a Google identity, or needs a setting only Cloud Storage has. Realized on Amazon S3, in an AWS account, by terraform/s3-object-store-amazon-s3 through Terraform and by cloudformation/s3-object-store-amazon-s3 in a CloudFormation stack: every public access is blocked, ACLs are disabled, and a consumer brings an IAM principal's access key. Realized on Google Cloud Storage, in a Google Cloud project, by terraform/s3-object-store-gcs, over its XML API: the bucket is never public, a consumer brings a Cloud Storage HMAC key, and an object ACL is refused under uniform bucket-level access. Realized on SeaweedFS, run in a cluster by helm/s3-object-store or on a Compose host by compose/s3-object-store with nothing else in place: the server creates the bucket, serves plain HTTP, and a consumer brings the key pair the slots hold.
ОфіційнийstorageЗалежні: 9Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/s3-object-store:1.0.0 - v1.0.0
Component
core/storage/kubestash
KubeStash, AppsCode's backup operator for Kubernetes, built on restic. It backs up the volumes of a workload into a restic repository in a bucket or on a Kubernetes volume, on the schedules its own resources declare, and restores them. It is installed once per cluster and acts on the KubeStash resources of every namespace. It backs nothing up by itself. A datastore is backed up by attaching core/backup to it in the Release, with the Manifest binding the helm/backup-kubestash FeatureRecipe, which writes a KubeStash BackupConfiguration, its BackupStorage and its RetentionPolicy beside the datastore; the backup runs only where this component is deployed. The Release supplies what KubeStash runs: AppsCode publishes KubeStash under its own terms and runs it only under a licence it issues, so the Release names the chart release in infra.source_version and the catalog names none. The licence reaches the operator by reference to a Secret bound to the required license slot. The compiler validates no licence and contacts no vendor, so a clean build says nothing about one. Realized on Kubernetes by helm/kubestash-chart, AppsCode's own chart, whose chart repository must be reachable from the build; cert-manager must already run in the cluster.
ОфіційнийstorageЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/kubestash:1.0.0 - v1.0.0
Component
core/storage/k8up
K8up, the backup operator built on restic. On the schedules written for it, it copies the files of a namespace's persistent volume claims into a restic repository, here one in a bucket reached over the S3 API or on a volume claim in the cluster, and restores them from there. It is installed once per cluster and acts on the K8up resources of every namespace. It backs nothing up by itself. A datastore is backed up by attaching core/backup to it in the Release, with the Manifest binding the helm/backup-k8up FeatureRecipe, which writes a K8up Schedule beside the datastore; the schedule runs only where this component is deployed. That FeatureRecipe names on each Schedule the Secret holding its repository password, and for a bucket its key. Realized on Kubernetes by helm/k8up-chart, the project's own chart, whose chart repository must be reachable from the build.
ОфіційнийstorageЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/k8up:1.0.0 - v1.0.0
Component
core/security/trustyai-huggingface-detector
TrustyAI's Hugging Face detector: a service that runs one text classifier from the Hugging Face Hub and reports what it detects in the texts a caller sends. For each text posted to its content route with a threshold, it answers with every label the classifier scores at or above it, never the classifier's first label, read as the one for harmless text. It changes nothing and forwards nothing: what to do about a detection is the caller's decision, and core/security/nemo-guardrails is one such caller. The classifier is values.model at values.model_revision. The default, protectai/deberta-v3-base-prompt-injection-v2 (Apache-2.0), labels each text SAFE or INJECTION, and its own model card states that the project which published it is archived and no longer maintained. Another model labels and scores differently, so a caller's labels and thresholds follow it. The verdict misses some attacks and flags some ordinary requests. The weights are fetched where the service runs, never at build time. It authenticates no caller and every text reaches it as sent, so it is published on an in-cluster address only. Its own metrics, in the Prometheus format at /metrics on the same port, are published as `metrics_endpoint`. The Release supplies the image in infra.image and the catalog names none: TrustyAI's published build carries NVIDIA's proprietary CUDA runtime, whose licence agreement limits its use to systems with NVIDIA GPUs, and is published only under a floating tag, whose superseded digests stop resolving. Nothing renders until the Release sets `infra.image`.
ОфіційнийsecurityЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/trustyai-huggingface-detector:1.0.0 - v1.0.0
Component
core/storage/barman-cloud-plugin
The Barman Cloud plugin for CloudNativePG, installed once per cluster in the namespace the CloudNativePG operator runs in. Through it, a PostgreSQL Cluster that CloudNativePG runs archives its write-ahead log and takes base backups into a bucket, and is restored from there. It backs nothing up by itself. A PostgreSQL database realized by helm/postgres-cloudnative-pg is backed up by attaching core/backup to it in the Release, with the Manifest binding helm/backup-cloudnative-pg; the backups run only where this component is deployed. The CloudNativePG operator, 1.26 or later, and cert-manager must already run in the cluster: the plugin talks to the operator over TLS with certificates cert-manager issues. Realized on Kubernetes by helm/barman-cloud-plugin-chart, the project's own chart, whose chart repository must be reachable from the build.
ОфіційнийstorageЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/barman-cloud-plugin:1.0.0 - v1.0.0
Component
core/storage/gcs
One Google Cloud Storage bucket, with the settings only Cloud Storage has, reached through Google's own API. A consumer reads the published bucket name, its gs:// address and its location, and calls the bucket with Google's client libraries as its own Google identity. There is no key slot: Google's API authenticates the caller's own Google identity, a service account or a workload identity, and access to the bucket is granted to that identity by IAM, outside the build. Reach for it where the consumer uses Google's client libraries with a Google identity, or needs a setting only Cloud Storage has, such as its storage class; and for core/storage/s3-object-store where the consumer speaks the S3 API with an access key, wherever it runs. Realized by terraform/gcs: the bucket is never public, because public access prevention is enforced and uniform bucket-level access is on, so access is granted by IAM alone and a request setting an object ACL is refused.
ОфіційнийstorageЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/storage/gcs:1.0.0 - v1.0.0
Component
core/security/trustyai-guardrails-gateway
TrustyAI's guardrails gateway in a model's path: a service speaking the OpenAI chat-completions API on one route, values.route, that sends every request through the orchestrator wired on its orchestrator socket with a fixed detector applied, overwriting any detectors a caller names. When the detector reports the latest message at or above values.threshold, the gateway answers with an ordinary completion carrying values.fallback_message, and the model is not asked to reply, though its server's /tokenize route still counts the refused text; with values.check_output set, a flagged reply is replaced the same way. It reads no system message, earlier turn or tool result, refuses a request whose latest message is a list of content parts or a tool result, and answers a refused streamed request with a stream carrying the detections and no text. When the orchestrator or its detector cannot be reached, the request is refused with an error. Its verdict is the detector's: a probabilistic classifier misses some attacks, refuses some legitimate requests and is known to be evadable. It authenticates no caller and serves no model listing. The Release supplies the image in infra.image and the catalog names none: TrustyAI's build adds the compat-openssl11 package, whose declared licence includes the OpenSSL licence, and is published only under a floating tag. Nothing renders until the Release sets `infra.image`. Realized on Kubernetes by its Helm Recipe and for local development by its Compose Recipe, each running that image as a single stateless service.
ОфіційнийsecurityЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/trustyai-guardrails-gateway:1.0.0 - v1.0.0
Component
core/security/supertokens
An authentication core: the backend service that holds an application's users, their credentials and their sessions, and issues and verifies the tokens that carry a session between requests. It is reached only by an application's own backend, over HTTP, through one of the project's backend client libraries configured with this component's published endpoint. Wire here when sign-in belongs to the application: the application's backend owns the sign-in routes and the screens a user sees, and this component sits behind it as the place the user records live. Wire `core/security/keycloak` instead when sign-in belongs to the platform, as an identity provider applications reach over OpenID Connect and SAML. The core keeps no durable state of its own: it needs a PostgreSQL backend on its required `backend` socket, where it creates and migrates its own tables at startup, and with none wired the build halts. It answers administrative calls over the whole user store to anything that can reach it and issues no API key by default, so it is published on an in-cluster address only; attaching a key to the `api_key` slot requires one. Realized on Kubernetes via its Helm Recipe, which synthesizes the core as a stateless Deployment behind a ClusterIP Service on its HTTP port. Exports its host, port, endpoint and protocol to consumers, where the endpoint is the base URL an application's backend library is configured with.
ОфіційнийsecurityЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/supertokens:1.0.0 - v1.0.0
Component
core/security/presidio-anonymizer
The Presidio anonymizer: a service that replaces spans of a text it is told about. A caller posts a text together with the list of spans to act on — each with its position and its entity type, in the shape the Presidio analyzer reports — and the service answers with the rewritten text and what it did to each span. Each entity type is handled by the operator the request names for it: replaced by a given value, removed, overwritten character by character, hashed or encrypted; with none named, every span is replaced by its entity type in angle brackets, such as <PERSON>. It can also reverse its own encryption when handed the key. It finds nothing. It acts only on the spans it is given, so text a caller does not list reaches the answer unchanged; finding the spans is the work of core/security/presidio-analyzer, and a caller that masks text calls both. Reachability is a deliberate part of the model. The service authenticates no caller: whatever can reach its address can submit text to it, and every text it is sent reaches it unmasked. It is published on an in-cluster address for the component that calls it, and is not meant to be exposed beyond the cluster. core/gateway/litellm takes it on its pii_anonymizer socket. Realized on Kubernetes by its Helm Recipe and for local development by its Compose Recipe, each running the vendor's published image as a single stateless service. Exports its host, port, endpoint and protocol to consumers via the uses topology, where the endpoint is the base URL a caller posts to.
ОфіційнийsecurityЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/presidio-anonymizer:1.0.0 - v1.0.0
Component
core/security/sealed-secrets
A sealed-secrets controller for Kubernetes. sealed-secrets is a cluster-scoped add-on that decrypts SealedSecret custom resources into native Kubernetes Secrets, using a private key the controller holds in the cluster; the matching public key encrypts secret material so it can be committed to source control safely. It extends the cluster with its own SealedSecret custom resource; applications consume it by declaring SealedSecret resources that the controller unseals, rather than by wiring a topology edge, so the component takes no `uses` edge. The one address it exports is where its own metrics are served: the controller serves them in the Prometheus format at /metrics on its metrics port, on by default, and its Recipe publishes that address as `metrics_endpoint`. The controller generates its sealing keypair in the cluster at runtime. sealed-secrets is realized on Kubernetes through a Helm Recipe; it has no standalone local-container form, so no Compose Recipe is provided.
ОфіційнийsecurityЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/sealed-secrets:1.0.0 - v1.0.0
Component
core/security/presidio-analyzer
The Presidio analyzer: a service that finds personal and other sensitive data in text. A caller posts a text, the language it is written in and, optionally, the entity types it is looking for — person names, email addresses, phone numbers, card and bank account numbers, national identifiers and the like — and the service answers with where each match sits in the text, which type it is and how confident the match is. It changes nothing: replacing what it found is the work of core/security/presidio-anonymizer, so a caller that masks text calls both. core/gateway/litellm takes it on its pii_analyzer socket. Recognition is in English only, the one language the published image loads a model for. It is automated recognition, so it misses some sensitive data and marks some text that is not sensitive; other protections belong beside it. It authenticates no caller, and every text reaches it unmasked, so it is published on an in-cluster address for the component that calls it and is not meant to be exposed beyond the cluster. Realized on Kubernetes by its Helm Recipe and for local development by its Compose Recipe, each running the vendor's published image as a single stateless service. Exports its host, port, endpoint and protocol to consumers, where the endpoint is the base URL a caller posts to.
ОфіційнийsecurityЗалежні: 20Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/presidio-analyzer:1.0.0 - v1.0.0
Component
core/security/nemoguard-jailbreak-detect
NVIDIA's NemoGuard JailbreakDetect, served by its NIM: a classifier that reads a text and reports whether it is an attempt to make a model ignore its instructions, with a score. It changes and forwards nothing. core/security/nemo-guardrails points its jailbreak-detection rail at it through its jailbreak_detection socket. Running it requires an NVIDIA AI Enterprise licence or NVIDIA Developer Program membership, and an NGC API key: NVIDIA restricts access to the image and the model, and the container fetches its model from NVIDIA with the key when it starts. The Release names the image in infra.image and the catalog names none, so nothing renders until it is set. Realized on Kubernetes by helm/nemoguard-jailbreak-detect, one replica that requests a GPU. It authenticates no caller, so it is published on an in-cluster address only.
ОфіційнийsecurityЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/nemoguard-jailbreak-detect:1.0.0 - v1.0.0
Component
core/security/openbao
An OpenBao secrets-management server. OpenBao stores, generates, and controls access to tokens, passwords, certificates, and other secrets behind an authenticated HTTP API, and encrypts data at rest on a durable data volume — so it renders as a self-hosted, stateful running server, not a managed-cloud resource. Its operational type is therefore `stateful-server`: the deployment runs and owns the OpenBao process backed by a persistent volume. This component ships a single standalone server that keeps its state in the built-in file storage backend; highly-available, multi-node topologies are a future refinement not modelled here. OpenBao is initialized and unsealed at runtime inside the cluster — its root token and unseal keys are produced by the running server after it starts — so the component declares no credential slot and no secret is materialized at render time. Realized on Kubernetes through its Helm Recipe; it has no standalone local-container form, so no Compose Recipe is provided. Exports its API endpoint host and port to consumers via the uses topology. OpenBao is distributed under the MPL-2.0 license. With `values.metrics_enabled` true the Recipe turns on OpenBao's Prometheus telemetry and publishes its address as `metrics_endpoint`. A scrape presents an OpenBao token, best one whose policy grants read and list on sys/metrics and nothing else.
ОфіційнийsecurityЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/openbao:1.0.0 - v1.0.0
Component
core/security/kong-ai-pii-service
Kong's AI PII service: a service that finds personal and other sensitive data in a text and rewrites it in the same call. A caller posts a text, the categories to look for — person names and places, phone numbers, email addresses, card and bank numbers, national identifiers, credentials and the like — any regular expressions of its own, and how to rewrite what it finds, and the service answers with the rewritten text and what it found. The image is supplied by the operator: Kong distributes the service's images under its own terms, so the Release names in infra.image one it is entitled to pull, and the catalog names none. Nothing renders until the Release sets `infra.image`. Each image carries the language model for the language its tag names, or all of them; a model the service is configured for that its image does not carry is fetched from the spaCy models repository when the service starts, by the running service and never by the build. core/gateway/kong-enterprise takes it on its pii_sanitizer socket, for Kong's AI PII Sanitizer. Recognition is automated, so it misses some sensitive data and marks some text that is not sensitive. Kong's sanitizer presents no credential to it and every text reaches it unmasked, so it is published on an in-cluster address for the component that calls it and is not meant to be exposed beyond the cluster. Realized on Kubernetes by its Helm Recipe, running the image the Release names as a single stateless service over plain HTTP. Exports its host, port and protocol to consumers.
ОфіційнийsecurityЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/kong-ai-pii-service:1.0.0 - v1.0.0
Component
core/security/nemo-guardrails
NVIDIA NeMo Guardrails' server as a guard in a model's path: a service speaking the OpenAI chat-completions API that checks each request against its input rail, forwards what it admits to the upstream model API in infra.upstream_endpoint, and returns the upstream's answer, streamed or not. A request the rail refuses is answered by the guard itself with an ordinary completion, and nothing of it reaches the upstream. The rail sends the latest user message to the detector wired on the detector socket and refuses the request when the detector reports one of values.blocked_labels at or above values.threshold; a detector that cannot be reached, or answers with an error, refuses the request too. It judges the latest user message alone. Its verdict is a probabilistic classifier's: it misses some attacks, refuses some legitimate requests and is known to be evadable. Only a flow in its configuration, redefining the product's own, keeps a request body from switching the rail off, and that closure has to be proved again for any image or flow it was not proved on. NVIDIA's NemoGuard models add rails on the content_safety and jailbreak_detection sockets; the jailbreak rail admits a request when its classifier is unreachable. It authenticates no caller, and with no credential attached to its upstream_key slot a caller can reach the upstream on a credential of its own. The Release supplies the image in infra.image and the catalog names none: TrustyAI's published build carries NLTK's punkt tokenizer model, whose licence is unresolved upstream, and is published only under a floating tag, whose superseded digests stop resolving. Nothing renders until the Release sets `infra.image`.
ОфіційнийsecurityЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/nemo-guardrails:1.0.0 - v1.0.0
Component
core/security/fms-guardrails-orchestrator
The FMS Guardrails Orchestrator: a service that runs text detectors around a model call. A caller posts an OpenAI chat-completions request naming the detectors to apply; the orchestrator checks the latest message, answers a detection with the detections and no reply, and otherwise forwards the request to the upstream server in infra.upstream_host and infra.upstream_port, adding any detection on the reply. It also answers detection-only requests over a text. It is not a guard on its own: a request naming no detectors is forwarded unchecked, so it sits behind a component that names them, such as core/security/trustyai-guardrails-gateway. Its Recipes register the one detector wired on the detector socket under values.detector_id. Answering a detection asks the upstream's /tokenize route, which vLLM serves, to count the refused text, so that text reaches the upstream. It authenticates no caller and forwards none of a caller's headers. The Release supplies the image in infra.image and the catalog names none: TrustyAI's build adds the compat-openssl11 package, whose declared licence includes the OpenSSL licence, compiles in root-certificate data published under a data licence, and is published only under a floating tag. Nothing renders until the Release sets `infra.image`. Realized on Kubernetes by its Helm Recipe and for local development by its Compose Recipe, each running that image as a single stateless service.
ОфіційнийsecurityЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/fms-guardrails-orchestrator:1.0.0 - v1.0.0
Component
core/security/keycloak
Keycloak, a database-backed identity and access-management server. Keycloak serves its account console and token endpoints on an HTTP port and keeps its realms, clients, and users in a PostgreSQL backend wired through its `backend` uses-socket, and it provisions and migrates its own schema on startup against that backend. Keycloak holds no persistent state of its own, so it renders as a stateless server. On Kubernetes there are two realizations and the Manifest chooses between them. `keycloak` synthesizes the server as a single stateless workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `keycloak-chart` renders the codecentric Keycloak.X Helm chart, which additionally carries what the server needs to run as more than one replica, so reach for it where a chart repository is reachable at build time and the server will be scaled out. In both, `infra.external_base_url` becomes the server's hostname and the realms in `values.realm_imports` are imported when it starts. With `values.metrics_enabled` true, both serve the server's own metrics on its management port and publish their address as `metrics_endpoint`; false, that key is empty. The management port, which checks no caller, stays inside the cluster on `keycloak-chart` whatever `infra.overrides.service_type` says; LoadBalancer or NodePort publishes the console alone on `<rendered_name>-external`.
ОфіційнийsecurityЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/keycloak:1.0.0 - v1.0.0
Component
core/security/external-secrets
An external-secrets operator that synchronizes secrets from external secret-management systems into native Kubernetes Secrets. external-secrets is a cluster-scoped add-on that reads material from providers such as HashiCorp Vault, AWS Secrets Manager, Google Secret Manager, or Azure Key Vault and materializes it as in-cluster Secrets, keeping them in sync as the source changes. It extends the cluster with its own ExternalSecret, SecretStore, ClusterSecretStore, and related custom resources; applications obtain their secrets by declaring those resources rather than by wiring a topology edge, so the component takes no `uses` edge, and what it publishes is where Prometheus finds its metrics. A controller reconciles ExternalSecret resources, a webhook validates and defaults the custom resources, and a certificate controller manages the webhook's serving certificate. external-secrets is realized on Kubernetes through its Helm Recipe (the official chart); it has no standalone local-container form, so no Compose Recipe is provided. Each of the three serves its own metrics on a port no Service carries, so the Recipe publishes the pod label selecting all three, the name of their metrics port and the path, for a scraper that discovers those pods directly.
ОфіційнийsecurityЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/external-secrets:1.0.0 - v1.0.0
Component
core/observability/thanos-store
Thanos Store Gateway, the read side of a Thanos metrics stack. The Store Gateway reads historical metric blocks from an S3-compatible object store and serves them to queriers over its StoreAPI on a gRPC port. The object store it reads from is wired through its `objstore` uses-socket: the endpoint the store publishes gives the S3 endpoint in the object-store configuration, and its scheme whether that connection is plain HTTP or TLS. Credentials reach the process as environment variables through the AWS SDK credential chain rather than being embedded in that configuration, so the rendered configuration carries no secret: the access key's secret half is the connection's `credential`, and its identifier the connection's `credential_id` where one is attached, else the `s3_access_key` value. The Store Gateway keeps only an ephemeral local cache of index data — the object store is the system of record — so it renders as a stateless server. Exports its StoreAPI host, port, and protocol to consumers via the uses topology. The gateway serves its own metrics in the Prometheus format at /metrics on its HTTP port, not the published StoreAPI port, and publishes that address as `metrics_endpoint`.
ОфіційнийobservabilityЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/thanos-store:1.0.0 - v1.0.0
Component
core/security/cert-manager
A cert-manager TLS-certificate lifecycle controller for Kubernetes. cert-manager is a cluster-scoped add-on that automates the issuance and renewal of X.509 certificates from issuing sources such as ACME (Let's Encrypt), a private CA, or Vault. It extends the cluster with its own Certificate, Issuer, and ClusterIssuer custom resources; applications request certificates by declaring those resources rather than by wiring a topology edge, so the component takes no `uses` edge, and what it publishes is where Prometheus finds its metrics. The controller reconciles certificate requests, a webhook validates and defaults the custom resources, and a CA injector wires the webhook's CA bundle into the cluster. cert-manager is realized on Kubernetes through a Helm Recipe; it has no standalone local-container form, so no Compose Recipe is provided. Each of the three serves its own metrics, and their Services do not all name the metrics port alike, so the Recipe publishes the pod label selecting all three, the name of their metrics port and the path, for a scraper that discovers those pods directly.
ОфіційнийsecurityЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/security/cert-manager:1.0.0 - v1.0.0
Component
core/observability/thanos-query
Thanos Query, the query front end of a Thanos metrics stack. Query answers PromQL requests and fans them out over the StoreAPI to one or more Store Gateways, deduplicating and merging the results. The Store Gateway it queries is wired through its `store` uses-socket: the gateway's host and port arrive on the injection context and compose the StoreAPI endpoint Query dials. Query speaks the StoreAPI over gRPC, so a component wired there that publishes a protocol other than grpc, or none, halts the build; Thanos Store publishes grpc. Query serves its PromQL API and web UI on an HTTP port and exposes its own StoreAPI on a gRPC port so queriers can be layered. Query holds no persistent state of its own — it reads through the Store Gateway — so it renders as a stateless server. Exports its HTTP host, port, and protocol to consumers via the uses topology. Query serves its own metrics in the Prometheus format at /metrics on its HTTP port and publishes that address as `metrics_endpoint`.
ОфіційнийobservabilityЗалежні: 11Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/thanos-query:1.0.0 - v1.0.0
Component
core/observability/tempo
A Grafana Tempo distributed-tracing server in its single-binary shape: every subsystem runs inside one process backed by a local trace store, so it renders as a single stateful server with a data volume. It receives trace spans over the OpenTelemetry protocol and other tracing wire formats, and answers trace-lookup and search queries over its HTTP API, on its own port, which a dashboards server reads to display traces. It serves its own metrics in the Prometheus format at /metrics on that HTTP port, 3200, on by default, and every Recipe publishes that address as `metrics_endpoint`. Its primary contract to consumers is the OTLP gRPC receiver: a trace producer wires to it and pushes spans there, and the component exports that receiver's host and port. It carries `observability.aggrega.ai/traces: otlp-grpc`, so an OpenTelemetry Collector's `traces` socket takes it and exports spans to that receiver. Realized on helm by two realizations the Manifest chooses between, and on compose by one. `tempo` synthesizes the server as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `tempo-chart` renders the community-maintained Tempo Helm chart, which assembles the server's configuration from named settings rather than from a hand-written file and additionally accepts spans in the Jaeger and Zipkin wire formats, so reach for it where a chart repository is reachable at build time. On compose, `tempo` runs the same image and configuration as its helm namesake, as a service publishing nothing to the host.
ОфіційнийobservabilityЗалежні: 16Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/tempo:1.0.0 - v1.0.0
Component
core/observability/prometheus
A Prometheus metrics server, storing what it scrapes in a local time-series database. Its targets are wired through its `scrape_target` uses-socket, which admits servers, stateful servers and workers, and each connection becomes one scrape job, its label and cadence set on that connection. The target says where it serves metrics: it publishes `metrics_endpoint`, the address Prometheus scrapes, or, on Kubernetes, `metrics_service` or `metrics_pod_selector`, which Prometheus discovers from the cluster. A target publishing none of them halts the build, naming it. A connection's optional `credential` is presented to its target as a bearer token, read from a file the Recipe delivers by reference; a target publishing `metrics_credential_required` true halts the build while its connection carries none. Prometheus publishes its own metrics address as `metrics_endpoint` too. With `values.otlp_receiver` true it also receives metrics pushed over OTLP HTTP, at the address it then publishes as `otlp_http_endpoint`, and that endpoint authenticates no caller. It carries `observability.aggrega.ai/metrics: otlp-http`, which an OpenTelemetry Collector's `metrics` socket selects. On Kubernetes the Manifest chooses between two realizations. `prometheus` synthesizes the server as a workload and fetches nothing, for a build that must complete with no chart repository reachable. `prometheus-chart` renders the prometheus-community Helm chart, whose cluster-wide read grant also lets the server scrape whatever the cluster advertises through scrape annotations.
ОфіційнийobservabilityЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/prometheus:1.0.0 - v1.0.0
Component
core/observability/otel-collector
The OpenTelemetry Collector, in its contrib distribution: a stateless server that receives traces, metrics and logs over the OpenTelemetry protocol (OTLP), on a gRPC port and an HTTP port, and runs each signal through a pipeline of its own. A producer points its OTLP exporter at the endpoint the collector publishes for the transport it speaks. Every signal always has a pipeline fed by the OTLP receiver. A backend wired to the signal's socket — `traces`, `metrics` or `logs`, each taking the components labelled `observability.aggrega.ai/<signal>` — adds one exporter to that pipeline, for the protocol the label names. A Prometheus is reached only with its `values.otlp_receiver` true; the build halts otherwise, naming that line. A signal with nothing wired ends in the collector's debug exporter at basic verbosity: what arrives is summarised in the collector's own log, one line per batch, and sent nowhere else. A producer is not refused for want of a backend, and what it sends is not dropped unseen. Realized on helm by `otel-collector-chart`, which renders the project's own Helm chart and needs its chart repository reachable at build time, and on compose by `otel-collector`, a compose service publishing nothing to the host. Both run the same image with the same configuration, written once in the OtelCollectorConfig bundle they import. With `values.metrics_enabled` true the collector serves its own telemetry metrics, at /metrics on port 8888, on every interface rather than its loopback address, and publishes that address as `metrics_endpoint`; false, that key is empty. The port checks no caller. On helm it joins a ClusterIP Service, and another Service type halts the build.
ОфіційнийobservabilityЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/otel-collector:1.0.0 - v1.0.0
Component
core/observability/langfuse
A self-hosted tracing and evaluation platform for language-model applications, and the home a team keeps its prompts in. An application sends a trace per request, carrying the version of the prompt that produced it, and fetches versioned prompts at run time over the platform's own API; the platform ingests OpenTelemetry spans as well as its own client libraries. Exports its console host, port, endpoint and protocol to consumers. It keeps none of its own durable state. The relational backend, the Redis-compatible cache and the S3-compatible object store are wired as topology edges; the columnar analytics store holding the traces is supplied as pointer strings — its address, migration address, user and database — which carry no default, and the build halts until both addresses are set. On Kubernetes there are two realizations and the Manifest chooses between them. `langfuse` synthesizes the console and the asynchronous worker as one Deployment and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `langfuse-chart` renders the platform maintainers' own Helm chart, which puts the console and the worker in separate Deployments and adds the platform's schema-migration switches, so reach for it where a chart repository is reachable at build time and the two will be scaled independently.
ОфіційнийobservabilityЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/langfuse:1.0.0 - v1.0.0
Component
core/observability/loki
A Grafana Loki log-aggregation server. Loki receives log streams from shippers and agents over its HTTP push API, indexes them by label rather than by full-text content, stores them in a chunk object store, and answers log-query and range requests over the same HTTP API. This component models Loki in its single-binary shape: every Loki subsystem — distributor, ingester, querier, compactor — runs inside one process backed by a local filesystem chunk store, so it renders as a single stateful server with a data volume rather than a set of microservices. A memcached may be wired to the optional `chunk_cache` socket to hold decompressed chunks between queries; the per-connection values on that edge — batch size, parallelism, timeout and entry lifetime — tune how this server drives that one cache, and with nothing wired the server runs without a chunk cache. The socket takes one connection, and a second halts the build, as does a target publishing a protocol other than memcached or memcached-tls, or none. Its contract to consumers is that HTTP endpoint: a log shipper pushes streams to it, and a dashboards server reads them back over the query API. It also receives OpenTelemetry logs at the base address it publishes as `otlp_http_endpoint`, and carries `observability.aggrega.ai/logs: otlp-http`, so an OpenTelemetry Collector's `logs` socket takes it. Realized on helm by `loki`, which sources the official Grafana Loki chart, and on compose by `loki`, a compose service running the same image and publishing nothing to the host. Loki serves its own metrics in the Prometheus format at /metrics on its HTTP port, on by default, and every Recipe publishes that address as `metrics_endpoint`.
ОфіційнийobservabilityЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/loki:1.0.0 - v1.0.0
Component
core/observability/jaeger
A Jaeger distributed-tracing backend in its single-node shape. Jaeger receives trace spans over the OpenTelemetry protocol, keeps them in a local embedded key-value trace store, and serves trace-lookup and search queries over its query UI. Every Jaeger subsystem — the OTLP receiver, the storage writer, and the query service — runs inside one process backed by a local on-disk store, so it renders as a single stateful server with a data volume rather than a set of microservices. Its primary contract to consumers is the OTLP gRPC receiver: a trace producer wires to it and pushes spans there. The OTLP HTTP receiver and the query UI are exposed on their own ports. Realized on Kubernetes via its Helm Recipe, which synthesizes the workload directly from the official Jaeger image; helm-only. Exports its OTLP gRPC receiver host and port to consumers via the uses topology, and carries `observability.aggrega.ai/traces: otlp-grpc`, so an OpenTelemetry Collector's `traces` socket takes it and exports spans to that receiver. The packaged Jaeger server is Apache-2.0. With `values.metrics_enabled` true the Recipe has Jaeger serve its own metrics, at /metrics on port 8888, on every interface rather than its loopback address, and publishes that address as `metrics_endpoint`; false, that key is empty. The port checks no caller.
ОфіційнийobservabilityЗалежні: 14Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/jaeger:1.0.0 - v1.0.0
Component
core/observability/grafana
A Grafana dashboards and visualization server. It keeps its own state — users, saved dashboards, preferences — in a local database, so it renders as a server backed by a persistent volume. It ships standalone: it declares no socket for a metrics or logs source to be wired in as a provisioned data source. Exports its web endpoint host and port to consumers. Grafana serves its own metrics in the Prometheus format at /metrics on its web port, on by default, and every Recipe publishes that address as `metrics_endpoint`. An admin account is always present: its username is a developer value and its password is wired through the optional `admin_password` credential slot. Unattached, the Recipe supplies a deterministic development default instead of a generated password, so the render stays byte-identical across runs. Realized for local development by a Compose Recipe, and on Kubernetes by two realizations the Manifest chooses between. `grafana` synthesizes the server as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `grafana-chart` renders the community-maintained Grafana Helm chart, which additionally carries the server's own provisioning surface — data sources and dashboards declared in the Release through `infra.overrides` — so reach for it where a chart repository is reachable at build time.
ОфіційнийobservabilityЗалежні: 14Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/grafana:1.0.0 - v1.0.0
Component
core/observability/fluentd
A Fluentd log aggregator. Fluentd is the receiving end of a log pipeline: it listens on the `forward` input protocol and accepts log records shipped to it by upstream forwarders (a Fluent Bit or Fluentd forwarder), then processes and routes them to its output. Because it is a sink rather than a forwarder, Fluentd declares no downstream uses-socket; instead it exports the host and port of its forward endpoint through `provides`, so any forwarder that binds this component on its forward socket ships records to it. The aggregator buffers records in memory by default and renders as a stateful server; an optional data volume can back the buffer when durable buffering is requested. Realized for local development by a Compose Recipe, and on Kubernetes by two realizations the Manifest chooses between, which run the same image, configuration, port and buffer. `fluentd` synthesizes the aggregator as native Kubernetes resources and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `fluentd-chart` renders the Fluent project's own Helm chart configured as that aggregator, so reach for it where a chart repository is reachable at build time. The Fluentd server is Apache-2.0.
ОфіційнийobservabilityЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/fluentd:1.0.0 - v1.0.0
Component
core/observability/fluent-bit
A Fluent Bit log processor and forwarder, collecting log records from its inputs and shipping them to a downstream sink wired through its `forward_target` uses-socket: the selected component's host and port arrive on the injection context and are rendered into a `forward` output, so Fluent Bit forwards to whichever sink the topology binds. That output speaks the Fluentd forward protocol, so a sink wired there that publishes a protocol other than forward, or none, halts the build; Fluentd publishes forward. Fluent Bit holds no persistent state, so it renders as a stateless server. Exports its HTTP monitoring endpoint's host and port to consumers, and where Prometheus finds the metrics served there. Realized for local development by a Compose Recipe, and on Kubernetes by two realizations the Manifest chooses between, which differ in where the records come from. `fluent-bit` synthesizes a single forwarding pod with its own declared input and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `fluent-bit-chart` renders the official Fluent Helm chart in its node-collector shape — one pod per node reading that node's container logs, with the cluster read grants the record-enrichment filter needs — so reach for it where records are to be collected from the cluster itself. Both publish the same monitoring endpoint and forward to the wired sink; for its metrics the chart-backed one publishes the Service a scraper discovers every node's collector behind. With the Prometheus Operator, `core/prometheus-exporter` attached to the synthesized realization, port `http` and path `/api/v2/metrics/prometheus`, emits a ServiceMonitor for it.
ОфіційнийobservabilityЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/observability/fluent-bit:1.0.0 - v1.0.0
Component
core/messaging/twilio
A messaging and voice carrier reached over Twilio's REST API, so an application sends a text message or places a call without operating telephony infrastructure. The developer names the modality, text or call, and the direction, outbound, inbound or both; the operator supplies the account coordinate and the API base address. Wire here rather than to a broker: a broker delivers inside the Release, a carrier onto the public networks. Consumers read the base address, the account coordinate, the modality and the direction. Realized by reference: no compute runs here.
ОфіційнийmessagingЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/twilio:1.0.0 - v1.0.0
Component
core/messaging/vonage
A voice carrier reached over Vonage's Voice API, so an application places and answers calls without operating telephony infrastructure. The modality is calls; the developer names the direction, outbound, inbound or both. The operator supplies the API base address and the two addresses the carrier calls back on, for instructions and for events. Wire here rather than to a broker: a broker delivers inside the Release, a carrier onto the public telephone network. Consumers read the base address, both callback addresses, the modality and direction. Realized by reference: no compute runs here.
ОфіційнийmessagingЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/vonage:1.0.0 - v1.0.0
Component
core/messaging/telnyx
A messaging and voice carrier reached over Telnyx's REST API, so an application sends a text message or places a call without operating telephony infrastructure. The developer names the modality, text or call, and the direction, outbound, inbound or both; the operator supplies the API base address. Wire here rather than to a broker: a broker delivers inside the Release, a carrier onto the public networks. Consumers read the base address, the modality and the direction. Realized by reference: no compute runs here.
ОфіційнийmessagingЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/telnyx:1.0.0 - v1.0.0
Component
core/messaging/strimzi-operator
The Strimzi Cluster Operator, declared as an explicit component so an operator-driven Kafka realization can ship its runtime prerequisite alongside the Kafka custom resource it emits, rather than assuming an ambient install. The control plane is the operator that reconciles Kafka custom resources into running brokers. Declare this component in the same Aggregate as an operator-driven Kafka to satisfy the Kafka component's control-plane prerequisite; the operator affirms the prerequisite on the Kafka in the same Release. Realized on Kubernetes by two Helm Recipes. `strimzi-operator-chart` renders the project's own chart, custom-resource definitions and the operator's full roles included, and is the one to reach for; its chart repository must be reachable from the build. `strimzi-operator` is a minimal bootstrap installing the operator alone, and at the image pinned here its role falls short of what the operator needs to run a Kafka cluster. The operator serves its own metrics on its pod. With `values.metrics_enabled` true each Recipe puts their port on a ClusterIP Service in the control plane's namespace and publishes the address as `metrics_endpoint`; false, that key is empty.
ОфіційнийmessagingЗалежні: 6Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/strimzi-operator:1.0.0 - v1.0.0
Component
core/messaging/rabbitmq
A RabbitMQ message broker speaking the AMQP protocol, persisting queues and messages durably. Exports its host and AMQP port to consumers via the uses topology. Self-hosted, it renders as a stateful workload with a persistent volume, by a Helm Recipe (Workload synthesis to native Kubernetes resources), a Compose Recipe (local development), and a controller-driven Helm Recipe that emits a RabbitmqCluster custom resource for the RabbitMQ Cluster Operator. Reach for the synthesis Recipe on a cluster that runs no RabbitMQ control plane, and for the controller-driven one where the control plane is in place and the cluster should own upgrades, rolling restarts and credential rotation. Broker access always requires credentials: on these Recipes the password slot is optional, and each falls back to a deterministic inline development default so the render stays reproducible. Reach for the rabbitmq-amazon-mq Recipes, on terraform and on CloudFormation, where the broker is an Amazon MQ broker in the operator's AWS account. Amazon MQ serves AMQP over TLS alone, on 5671, published as protocol amqps, and requires a password: a core/random-password attachment, without which the build halts. It runs RabbitMQ 4.3, 4.2 and 3.13, so the engine_version default halts there. The self-hosted brokers serve their own metrics, published with `values.metrics_enabled` true; the Amazon MQ Recipes halt on true.
ОфіційнийmessagingЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/rabbitmq:1.0.0 - v1.0.0
Component
core/messaging/rabbitmq-cluster-operator
The RabbitMQ Cluster Operator, declared as an explicit component so a controller-driven RabbitMQ realization can ship its runtime prerequisite alongside the RabbitmqCluster custom resource it emits, rather than assuming an ambient install. The control plane is the controller that reconciles RabbitmqCluster custom resources into running brokers, their client Services and their credential Secrets. Declare this component in the same Aggregate as a controller-driven RabbitMQ to satisfy the RabbitMQ component's control-plane prerequisite; the operator affirms the prerequisite on the broker in the same Release. Realized on Kubernetes via its Helm Recipe. The controller serves its own metrics on its pod. With `values.metrics_enabled` true the Recipe puts their port on a ClusterIP Service in the control plane's namespace and publishes the address as `metrics_endpoint`; false, that key is empty.
ОфіційнийmessagingЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/rabbitmq-cluster-operator:1.0.0 - v1.0.0
Component
core/messaging/nats
A NATS messaging server for pub-sub, request-reply, and queueing. In this lean shape NATS routes subjects in memory with no durable persistence (JetStream file storage is not enabled), so the component renders as a stateless workload rather than a stateful one and declares no storage. Authentication is off by default; when a credential is attached to the optional password slot, each Recipe enables user/password authentication and wires the credential through the credential subsystem. Provider-agnostic: the same component is implemented by a Helm Recipe (Workload synthesis to native Kubernetes resources) and a Compose Recipe (local development). Exports its host and client port to consumers via the uses topology. On Kubernetes there are two realizations and the Manifest chooses between them. `nats` synthesizes the server as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `nats-chart` renders the product's own Helm chart, which additionally carries a sidecar that reloads the server configuration in place, start-up, readiness and liveness probes against the monitoring endpoint, and a drain-then-shut-down sequence that lets connected clients move off before the pod goes away — and it is the realization that honours `monitoring_port`, so reach for it where a chart repository is reachable at build time.
ОфіційнийmessagingЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/nats:1.0.0 - v1.0.0
Component
core/messaging/livekit-server
A LiveKit media server — the selective forwarding unit that carries realtime audio and video between the participants of a room. A client opens a WebSocket signalling session on the API port, then exchanges media over WebRTC, which the server terminates either on one multiplexed UDP port with a TCP port beside it as the fallback path for clients whose network blocks UDP, or on a contiguous range of UDP ports that gives each connection its own socket, which carries traffic only where the pod uses the node's own network. Wire here when the cluster carries realtime voice or video itself rather than handing the media to a carrier it does not run, and the same server serves conventional applications as readily as agent workloads. Consumers receive the server's in-cluster host, its signalling port, the protocol that port speaks, and the API key identifier they pair with their own issued secret; the token-signing secret arrives through the api_secret slot here and as the `credential` on each consumer's connection to this server, and is never published on the topology. Realized by a Helm Recipe that synthesizes a Deployment and a ClusterIP Service over the upstream image and declares no volume: the server holds no durable disk, keeping room and participant state in memory, so more than one replica means wiring the state-store socket to a Redis-compatible datastore rather than adding storage here. With `values.metrics_enabled` true the Recipe turns on the server's metrics, on a port of their own that checks no caller, and publishes their address as `metrics_endpoint`; false, that key is empty.
ОфіційнийmessagingЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/livekit-server:1.0.0 - v1.0.0
Component
core/messaging/livekit-sip
A LiveKit SIP bridge — the service that carries a telephone call into and out of a realtime room. It answers SIP signalling on one port, which it binds for datagram and stream transport alike, negotiates the call there, and then carries the call audio over RTP on a contiguous range of UDP ports it draws a per-call socket from. The room side of the call runs over the bridge's own client session with the media server, and the state of a call in progress lives in the Redis-compatible store the bridge and that server share, which is also how the requests that set a call up pass between them. Wire here when a SIP trunk is terminated inside the cluster and those calls are joined to rooms it already runs, rather than routing the conversation through a carrier's own application platform. Consumers receive the bridge's in-cluster host, its signalling port and the protocol that port speaks; the token-signing secret is the credential the connection to the media server carries and is never published on the topology, and the public key identifier arrives over that same edge, so the halves of one key pair cannot drift apart. Realized by a Helm Recipe that synthesizes a Deployment and a ClusterIP Service over the upstream image and declares no volume: the bridge holds no durable disk. The RTP range carries call audio only where the pod uses the node's own network and the range is open to the trunk on the node's firewall. With `values.metrics_enabled` true the Recipe turns on the bridge's metrics, on a port of their own that checks no caller, and publishes their address as `metrics_endpoint`; false, that key is empty.
ОфіційнийmessagingЗалежні: 14Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/livekit-sip:1.0.0 - v1.0.0
Component
core/messaging/kafka
An Apache Kafka event-streaming broker in KRaft mode, with no ZooKeeper, keeping its commit log on durable storage, so it renders as a stateful server. Its primary contract to consumers is the client bootstrap endpoint: a producer or consumer wires to it, through `host` and `port` or the `bootstrap_servers` list, and reaches the cluster there. Realized on Kubernetes two ways, both through Helm Recipes and chosen in the Manifest: a synthesis Recipe that builds one node, running the broker and controller roles in one process, from the official Apache Kafka image, and a controller-driven Recipe that emits a Kafka custom resource for the Strimzi Cluster Operator to reconcile. Reach for the synthesis Recipe on a cluster that runs no Kafka control plane — it needs no operator installed and no chart repository, so it builds with nothing fetched — and for the controller-driven Recipe where the Strimzi Cluster Operator is already in place and the cluster should own broker upgrades, rolling restarts and partition rebalancing; it runs Kafka 4.2.0 to 4.3.1 only, which the `engine_version` default is not. Reach for terraform/kafka-msk where Kafka is an Amazon MSK cluster in the operator's AWS account: two or three brokers, one per client subnet, serving plaintext on 9092 inside the VPC, or TLS on 9094. There is no Compose realization, and no CloudFormation one: a stack reads no bootstrap address off an MSK cluster without running code of its own. `infra.client_tls`, off by default, serves clients TLS alone and publishes `protocol` `kafka-ssl`: the synthesized broker presents the certificate attached to `server_certificate`, Strimzi one its cluster authority issues, MSK one AWS issues.
ОфіційнийmessagingЗалежні: 16Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/messaging/kafka:1.0.0 - v1.0.0
Component
core/gateway/tyk
Tyk Gateway, a self-hosted API gateway. The gateway terminates client traffic on its data plane and routes it to the services its API definitions name. It keeps its keys, rate-limit counters and session state in a Redis-protocol store wired through its `cache` uses-socket: the store's host and port arrive on the injection context and compose the gateway's storage connection. That store is the gateway's one prerequisite — it needs no management console and no licence key to serve traffic — so the stack a Release stands up is the gateway and a Redis-protocol component and nothing else. The gateway holds no durable state of its own, so it renders as a stateless server. Realized on Kubernetes by a Recipe that points at the chart the gateway's own maintainers publish, and for local development by a Recipe that synthesizes a container from the same published image. It can pin the public key it accepts from an upstream it forwards to, which a Release asks for by attaching core/upstream-key-pinning; that realization's description states what setting one pin costs the upstreams that have none. Exports its data-plane host, port, and protocol to consumers via the uses topology.
ОфіційнийgatewayЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/tyk:1.0.0 - v1.0.0
Component
core/gateway/traefik-hub
Traefik Hub, Traefik's paid gateway: the Traefik proxy run from Traefik Hub's own image, with an API gateway and an AI gateway a licence token unlocks. It serves each model wired to its models socket at /<alias>/v1/chat/completions through Traefik Hub's Chat Completion middleware. The operator supplies the image: Traefik distributes Traefik Hub under its own terms, so the Release names in infra.image one it is entitled to pull, and the catalog names none. Nothing renders until the Release sets `infra.image`. The licence token reaches the gateway by reference to a Secret in the required license slot. A Release may name its callers in infra.callers and their key hashes through the caller_keys slot; every model route then refuses a request with no declared key, and removes the key before forwarding. Through Traefik Hub's AI gateway it realizes core/period-budget, core/prompt-deny-patterns, core/semantic-caching and core/data-masking, each halting the build until infra.ai_gateway_entitled affirms the licence carries it; through Traefik's own failover and retry, core/fallback-chain and core/retry-policy. core/prompt-injection-shield, audit logging, target selection and reply validation halt, naming why. Traefik requires outbound HTTPS to api.traefik.io, where the gateway validates its token. A token lasts a year by default, and the gateway stops once it expires. The compiler validates no token and contacts no vendor, so a clean build cannot say whether the gateway starts, serves or enforces anything. With `values.metrics_enabled` true a ClusterIP Service carries the gateway's metrics port, which checks no caller, and the address is published as `metrics_endpoint`; false, that key is empty.
ОфіційнийgatewayЗалежні: 57Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/traefik-hub:1.0.0 - v1.0.0
Component
core/gateway/litellm
An AI gateway that fronts several models behind one in-cluster, OpenAI-compatible endpoint, routing each request by the alias it asks for. Each model wired to its models socket is served under the connection's alias, with that connection's credential; values.models adds aliases typed by hand, presenting provider_key. No credential reaches the calling application. It publishes one alias as its model, so an agent's model socket takes it in place of a model. Because it holds those credentials, it forwards only the request fields the proxy declares for the route, and refuses the rest naming the field. It answers only its model routes, the model listing, its health checks and key management. Its other optional sockets take the Presidio analyzer and anonymizer, the vector-search store that core/semantic-caching keeps replies in, and the guard that core/prompt-injection-shield sends requests through; wiring one alone changes no request. On Kubernetes, reach for helm/litellm-gateway, the synthesis, where the build must fetch nothing, and helm/litellm-gateway-chart, BerriAI's chart, where its registry is reachable. On Docker Compose, compose/litellm-gateway runs it for local development, without the semantic cache. values.log_masking masks what it lists in the copies of a successful call the gateway hands to its logs, never in what the provider or the caller receives. It excludes core/data-masking. Attached, core/semantic-caching keeps end users apart only by the end user each request names, which the gateway does not verify, and core/period-budget, once spent, refuses the administrative key's model requests too.
ОфіційнийgatewayЗалежні: 61Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/litellm:1.0.0 - v1.0.0
Component
core/gateway/nginx
An NGINX reverse-proxy gateway. NGINX terminates client HTTP traffic and forwards it to an upstream service wired through its `upstream` uses-socket: the upstream host and port arrive on the injection context and are rendered into the NGINX server configuration as the proxy_pass target, so the gateway routes to whichever component the topology binds. NGINX holds no persistent state, so it renders as a stateless server with no volume. Provider-agnostic: the same component is implemented by a synthesized Helm Recipe on the official NGINX image (Kubernetes) and a Compose Recipe (local development). Exports its host and listen port to consumers via the uses topology. The optional guard socket takes the core/security/nemo-guardrails a core/prompt-injection-shield attachment sends every forwarded request through in place of the upstream; wired alone it changes no request.
ОфіційнийgatewayЗалежні: 16Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/nginx:1.0.0 - v1.0.0
Component
core/gateway/kong
Kong, a database-backed API gateway. It routes client traffic by configuration held in the PostgreSQL backend wired to its `backend` socket; a migrations step provisions the schema before the gateway serves. The database is the system of record, so Kong renders as a stateless server. It publishes its proxy host, port and protocol. It also serves models: each one wired to its models socket gets a route through Kong's AI proxy, at /<alias>/v1 under the connection's alias, sent to the endpoint and model name the model publishes and presenting that connection's credential. It publishes values.model's route as its endpoint and that alias as its model, so an agent's model socket takes it in place of a model; with nothing wired both are empty. A core/prompt-deny-patterns attachment puts Kong's AI Prompt Guard on every one of those routes, where it stays once the attachment is removed. On Kubernetes the Manifest chooses between two realizations. `kong` synthesizes a workload with its bootstrap steps ahead of it and fetches nothing, for a build with no chart repository reachable. `kong-chart` renders the official Kong Helm chart, which also migrates the schema on either side of an upgrade, for a gateway upgraded in place. With `values.metrics_enabled` true each Recipe publishes Kong's metrics, /metrics on its Status API port, as `metrics_endpoint`; false, that key is empty. From another address that port answers a GET of /status, /status/ready or /metrics, with no caller check, and refuses every other request, POST /status/ready among them. On Kubernetes it is already open on the pod and the value names it on a Service; on Docker Compose the value newly opens it to every peer on the network.
ОфіційнийgatewayЗалежні: 51Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/kong:1.0.0 - v1.0.0
Component
core/gateway/higress
Higress, an open-source gateway built on Envoy and configured through Kubernetes resources: its routes are ingresses, the services it reaches are its own registry's entries, and its controls are WebAssembly plugins on the routes they act on. It terminates client HTTP on its proxy port, publishing its host, port and protocol. It serves models: each model wired to its models socket is reached under an alias at /<alias>/v1 through Higress's AI proxy, sent to the endpoint and model name the model publishes, presenting the connection's credential from its Secret; over https it checks that the model's certificate chains to the system's trust anchors and names its host. Connections naming one alias in serves share its requests. It publishes the route of values.model as its endpoint and that alias as its model, so an agent's model socket takes it in place of a model; with nothing wired both are empty. On those routes it masks requests, keeps a token budget in its cache socket's store, records each model request in its access log and sets each route's retry, refusing by name what Higress cannot keep; it declares no callers. Reply validation, the fallback chain, the semantic cache, the prompt-injection shield and latency-based target selection halt the build naming what Higress lacks. It pulls its plugin modules from Higress's plugin registry at run time, which the cluster must reach and the build cannot check. It is realized on Kubernetes by `higress-chart` alone, and publishes the pod label, port name and path a scraper discovers its gateway's metrics by.
ОфіційнийgatewayЗалежні: 56Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/higress:1.0.0 - v1.0.0
Component
core/gateway/kong-enterprise
Kong Gateway Enterprise, Kong's paid edition of its database-backed API gateway. Like core/gateway/kong it routes client traffic by configuration held in the PostgreSQL backend wired to its backend socket, and serves each model wired to its models socket at /<alias>/v1 through Kong's AI proxy. The image is supplied by the operator: Kong distributes the enterprise image under its own terms, so the Release names in infra.image one it is entitled to pull, version tag included, and the catalog names none. Nothing renders until the Release sets `infra.image`. The licence reaches the gateway by reference to a Secret bound to the required license slot, never as a value. The compiler validates no licence and contacts no vendor, so a Release that builds cleanly says nothing about one. Kong reads the licence as a node starts, and a node with no valid licence behaves as with an expired one: its entity configuration becomes read-only and proxy traffic is served with the configuration already in place. Kong sells its AI plugins as its AI Gateway Enterprise offering, an entitlement no build can see, so an AI Feature halts the build until infra.ai_gateway_entitled affirms it. Realized on Kubernetes by kong-enterprise-chart, over the official Kong Helm chart; there is no compose realization. With `values.metrics_enabled` true the Recipe puts the gateway's Status API port on a ClusterIP Service of its own and publishes /metrics there as `metrics_endpoint`; false, neither. The port is already open on the pod: from another address it answers a GET of /status, /status/ready or /metrics, with no caller check, and refuses every other request, POST /status/ready and any route the image adds among them.
ОфіційнийgatewayЗалежні: 56Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/kong-enterprise:1.0.0 - v1.0.0
Component
core/gateway/envoy-gateway
An Envoy Gateway ingress gateway, programmed through the Gateway API. Its control plane watches the cluster's GatewayClass, Gateway and route resources and configures an Envoy data plane, which terminates client HTTP traffic — and HTTPS too, where the Release names a certificate Secret through tls_secret — and load-balances it to the Services those routes name. This Component installs the control plane and claims the gateway class and the gateway itself. An application reaches it by declaring a route naming the gateway as its parent, and the gateway is named after the component the Aggregate declares. Realized on Kubernetes through a Helm Recipe; it has no standalone local-container form, so no Compose Recipe is provided. It publishes the pod label, port name and path a scraper discovers its proxies' metrics by. It also serves models through Envoy AI Gateway. Wired to the AI gateway controller on its extension socket, it calls that controller as its extension server, and the Release must place it in envoy-gateway-system or the build halts. Each model on its models socket is served under /v1 at the gateway's address to a request naming the connection's alias, with that connection's credential. The gateway publishes that address as its endpoint, qualified with envoy-gateway-system so an agent in any namespace reaches it, and values.model as its model. While a model is wired, every route on the gateway takes the 50Mi per-connection request buffer Envoy AI Gateway configures. It realizes core/period-budget in tokens, core/retry-policy, core/concurrency-limit per model, and core/audit-logging with no cost or verified caller.
ОфіційнийgatewayЗалежні: 58Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/envoy-gateway:1.0.0 - v1.0.0
Component
core/gateway/envoy-ai-gateway-operator
The Envoy AI Gateway control plane: a controller that turns the project's AI routes, model backends and policies into Envoy configuration, which Envoy Gateway takes from it by calling the controller's extension server, and a mutating webhook that adds the AI external processor to each data-plane pod Envoy Gateway creates. It publishes the address of its extension server for Envoy Gateway to call, and as `metrics_endpoint` the one its own metrics are served at. Its Helm Recipe installs the project's own chart and, beside it, the six AI Gateway custom-resource definitions. The controller runs under the cluster role the project's chart grants, and that grant spans the whole cluster: the controller may read, create, change and delete every Secret, Pod, Service, Deployment and DaemonSet in every namespace, and every resource of the Gateway API, Envoy Gateway, AI Gateway and inference-pool API groups. The webhook's serving certificate is a cert-manager Certificate, signed by a self-signed Issuer the Recipe renders, so cert-manager must already run in the cluster, unless infra.webhook_tls_secret names a Secret holding the certificate, in which case cert-manager is not needed. The controller expects the Envoy Gateway it extends in the `envoy-gateway-system` namespace. A QuotaPolicy applied by hand is not enforced: its rate-limit service is installed by nothing here; a token budget on this path is core/period-budget on Envoy Gateway. MCP session identifiers are encrypted under a key derived from the seed in the Secret infra.mcp_session_seed_secret names; with none named it is the chart's published default seed, a public value, so name one before an MCP route is served through this controller.
ОфіційнийgatewayЗалежні: 14Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/envoy-ai-gateway-operator:1.0.0 - v1.0.0
Component
core/gateway/contour
A Contour ingress gateway. Contour is a Kubernetes-native ingress controller: its control plane watches the cluster's Ingress and HTTPProxy resources and programs an Envoy data plane, which terminates client HTTP and HTTPS traffic and load-balances it to the backing services those resources name. Because its routing input is the cluster's own Ingress objects rather than a wired upstream, the gateway takes no `uses` edge; applications reach it by declaring Ingress resources against its ingress class. Contour is realized on Kubernetes through a Helm Recipe; it has no standalone local-container form, so no Compose Recipe is provided. Exports the Envoy data-plane host and listen port to consumers via the provides contract. Each Envoy pod serves its own metrics on a port no Service carries, so it also publishes the pod label selecting those pods, the name of their metrics port and the path, for a scraper that discovers them directly. The control plane's own metrics are not published.
ОфіційнийgatewayЗалежні: 15Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/contour:1.0.0 - v1.0.0
Component
core/gateway/cloudflare-ai-gateway
Cloudflare's AI Gateway, a hosted service in front of model providers: a caller sends its model requests to the gateway, which forwards them to the provider. Nothing runs in the Release and the compiler never calls Cloudflare. It publishes the address of one of the gateway's endpoints, composed from the Cloudflare account and the gateway's id, and the model a caller asks for, which the gateway forwards unchecked. On terraform its Recipe creates the gateway in the account and is the one that carries controls: reach for it where the Release is to create the gateway and set them. On helm and compose its Recipes reference a gateway that already exists and carry no Feature: reach for those where the gateway is created and configured outside the Release. The terraform Recipe realizes core/period-budget, in US dollars over a moving window, and core/retry-policy. Not realized: core/semantic-caching, as the cache answers identical requests only; core/fallback-chain, as a request falls back only inside a route it names; core/audit-logging, core/data-masking, core/prompt-injection-shield and core/target-selection. It holds no credential. What a gateway does when it serves a request is Cloudflare's documented behaviour and no build observes it. It carries no model.aggrega.ai/api label, so an agent's model socket does not take it.
ОфіційнийgatewayЗалежні: 11Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/cloudflare-ai-gateway:1.0.0 - v1.0.0
Component
core/gateway/azure-api-management
The AI gateway of Azure API Management, a hosted service: a caller sends its model requests to an API on an API Management instance, which forwards them to the model API behind it. Nothing runs in the Release and the compiler never calls Azure. It publishes the API's address and the model a caller asks for, which the gateway forwards unchecked. On terraform its Recipe adds the API, its backend and its policy to an instance that already exists, and is the one that can carry controls: reach for it where the instance exists and the Release is to configure the API. On helm and compose its Recipes reference an API that already exists and carry no Feature: reach for those where the API is configured outside the Release. On terraform it realizes core/semantic-caching, core/period-budget, core/prompt-injection-shield and core/retry-policy in the API's policy, and halts core/audit-logging by name. Not realized: core/response-validation, core/target-selection, core/data-masking, core/concurrency-limit and core/fallback-chain. A caller presents its own subscription key; the one credential the gateway holds is the optional provider_key, written as a reference to a Key Vault secret. What the gateway does with a request is Microsoft's documented behaviour and no build observes it. It carries no model.aggrega.ai/api label, so an agent's model socket does not take it.
ОфіційнийgatewayЗалежні: 11Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/azure-api-management:1.0.0 - v1.0.0
Component
core/gateway/apisix
Apache APISIX, an API gateway. APISIX terminates client HTTP traffic on its proxy port and publishes that port's host, port and protocol. While nothing is wired to its models socket it routes by configuration held in the etcd store wired to its `etcd` socket, written through an Admin API reachable in-cluster only and authenticated by the key attached to `admin_key`; both are then required. It holds no state of its own. It also serves models: each one wired to its models socket gets a chat route, and a completion route unless a period budget or a prompt guard is attached, through APISIX's AI proxy, at /<alias>/v1 under the connection's alias, sent to the endpoint and model name the model publishes and presenting that connection's credential. Those routes are rendered into a file the gateway reads, so with a model wired it reads no etcd and runs no Admin API, and an `etcd` connection or an `admin_key` attachment halts the build. It publishes values.model's route as its endpoint and that alias as its model, so an agent's model socket takes it in place of a model; with nothing wired both are empty. On Kubernetes the Manifest chooses between two realizations. `apisix` synthesizes a workload with a hand-written configuration file and fetches nothing; it serves no model and halts on a models connection. `apisix-chart` renders the project's own Helm chart, which assembles that configuration from named settings; it serves the models, their callers and the controls attached to them. With `values.metrics_enabled` true, both serve the gateway's own metrics, asking no credential, and publish their address as `metrics_endpoint`; false, that key is empty.
ОфіційнийgatewayЗалежні: 57Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/apisix:1.0.0 - v1.0.0
Component
core/gateway/agentgateway
agentgateway, an open-source gateway for model, tool and agent traffic, serving the OpenAI chat-completions API in front of the models wired to it. It terminates client HTTP traffic on its data-plane port, publishes that port's host, port and protocol, and runs from one configuration file. Each model wired to its models socket is served under the connection's alias: a request names the alias as its model at the gateway's one /v1 base URL, and is sent to the endpoint and model name the model publishes, presenting that connection's credential. It publishes that base URL as its endpoint and values.model as its model, so an agent's model socket takes it in place of a model; with nothing wired it serves no model and both are empty. A Release may name its callers in infra.callers, their keys through the caller_keys slot; every model request must then present a declared key as `Authorization: Bearer <key>`. Two realizations run the same image, pinned by digest. On Kubernetes, `agentgateway-chart` renders the project's standalone Helm chart, fetched from the project's registry, and realizes data masking, prompt deny patterns, the period budget, audit logging and the retry policy; the fallback chain, target selection, the semantic cache, response validation and the prompt-injection shield halt the build there, each naming why. Under Docker Compose, `agentgateway` runs it as one service and realizes none of them, nor callers, so an attachment halts the build naming its type. With `values.metrics_enabled` true each Recipe publishes the address of the gateway's metrics, always on, on a port that checks no caller, as `metrics_endpoint`; false, that key is empty.
ОфіційнийgatewayЗалежні: 58Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/gateway/agentgateway:1.0.0 - v1.0.0
Component
core/devtool/adminer
Adminer — a lightweight single-file database administration web UI. Connects to any stateful-server database wired through its `databases` uses-socket: the database host arrives on the injection context and the Recipe renders it as ADMINER_DEFAULT_SERVER so the login screen targets the right server. Login credentials are entered in the browser; the component holds no credential slots of its own.
ОфіційнийdevtoolЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/devtool/adminer:1.0.0 - v1.0.0
Component
core/devtool/harbor
A Harbor container registry, declared as one component and deployed as one unit: an API and token-service control plane, an OCI blob store, a web console, an asynchronous worker tier, a vulnerability scanner, a relational backend and a key-value cache, so the registry stands up with nothing else declared beside it. An Aggregate declaring it gets a registry container clients log in to, push to and pull from, and a console to administer it, both behind the one in-cluster address `provides` publishes. Six keys and shared secrets have no default and the build halts until each slot is filled: token_key, the key the token service signs with and its certificate; core_secret, jobservice_secret and registry_http_secret, the secrets its services authenticate to each other and sign upload state with; csrf_key, the key the control plane signs request-forgery tokens with; and secret_key, the key it encrypts stored credentials with. Each is read from a Kubernetes Secret by name, so none reaches the rendered artifact. The passwords — the administrator's, in its slot, and the backend's and the blob store's account, in `infra` — take development values the Recipe states when left unattached and unset, for evaluation and development. Those values are published in this catalog: a Release serving real traffic replaces them. With `values.metrics_enabled` true the Recipe turns on the Prometheus metrics of the control plane, worker tier, blob store and exporter and publishes their four Services as `metrics_service`; false, those keys are empty. The switch also opens the blob store's debug listener — profiling, execution traces and runtime variables, no caller check — to every in-cluster caller.
ОфіційнийdevtoolЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/devtool/harbor:1.0.0 - v1.0.0
Component
core/devtool/promptlayer
A hosted prompt registry, and the surface a team edits and versions its prompts on. A prompt is authored and versioned in the registry rather than in application source, each version carries a release label such as prod or staging, and an application fetches the prompt behind a label at run time over the platform's API — so moving a label to a new version changes what the application sends without a redeploy. The registry also keeps a log of the requests an application reports to it, which is how a prompt version and the behaviour it produced are read together. Nothing runs in the deployment for this component: the platform is operated by its publisher, and a deployment holds a contract and an API key rather than a process. Its realizations are therefore reference realizations, which emit only the wiring a consumer needs to reach the service and provision no compute. The base address is an operational value so it can be repointed — the platform's own client libraries read the same override — which is what lets a Release address a regional or self-managed installation without a consumer changing. Exports the base address and the release label to consumers via the uses topology, so a consuming component reads which registry to call and which label to ask for from one place rather than each of its own containers carrying a copy. The API key is not exported and is not held here: the environment-variable name a credential arrives under belongs to the consuming application, so a consumer that needs one declares the `credential` slot on the socket that reaches the registry and delivers it from its own container.
ОфіційнийdevtoolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/devtool/promptlayer:1.0.0 - v1.0.0
Component
core/datastore/weaviate
A Weaviate vector database. Stores objects together with their embedding vectors and answers nearest-neighbour and hybrid queries over them, so an application that retrieves reference material by similarity is wired to this component rather than to a general-purpose store. Collections are created through Weaviate's own API at runtime; the component declares the server, not its contents. Where the application supplies its own vectors the vectorizer module is left unset, which is the default; naming a module makes the server compute vectors for incoming objects. Provider-agnostic: the same component is implemented by a Helm Recipe (Workload synthesis to native Kubernetes resources, running the official upstream image with no external chart) and a Compose Recipe (local development). Exports its host, its HTTP port and its gRPC port to consumers via the uses topology. With `values.metrics_enabled` true each Recipe turns on Weaviate's metrics, served at /metrics on port 2112, and publishes that address as `metrics_endpoint`; false, that key is empty. The port checks no caller, even where an API key is attached, and also answers /tenant-activity, listing each collection's tenants with their last read and write.
ОфіційнийdatastoreЗалежні: 53Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/weaviate:1.0.0 - v1.0.0
Component
core/datastore/timescaledb
A time-series database on the PostgreSQL engine: PostgreSQL with the TimescaleDB extension, which adds hypertables — tables partitioned by a time column — and a columnstore for analytical queries over time ranges. Wire here when data arrives ordered by time and is queried over spans of it — metrics, events, telemetry; core/datastore/postgres is the general-purpose relational store, and core/datastore/documentdb the document store on the same engine. It presents the standard PostgreSQL wire contract and exports its host, port, user, database and protocol, so any consumer that speaks PostgreSQL reaches it unchanged. Realized three ways, all self-hosted: a Helm Recipe synthesizing a StatefulSet, a Compose Recipe for local development, and a Helm Recipe emitting a Cluster custom resource for the CloudNativePG control plane. Reach for the synthesized Recipes where nothing is installed in the target cluster; reach for the controller-driven Recipe where that control plane is in place and the cluster can mount an image as a volume. Only that Recipe serves Prometheus metrics; the synthesized Recipes publish those keys empty. On Azure no realization is authored: Azure Database for PostgreSQL flexible server carries the extension's Apache-licensed edition, but nothing its API offers enables it in a database; only CREATE EXTENSION does. The managed path there is core/datastore/postgres on terraform/postgres-flexible-server, with infra.overrides.server_configuration setting azure.extensions to TIMESCALEDB and shared_preload_libraries to timescaledb, and the application running CREATE EXTENSION timescaledb in its database.
ОфіційнийdatastoreЗалежні: 42Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/timescaledb:1.0.0 - v1.0.0
Component
core/datastore/valkey
A Valkey in-memory key-value cache, speaking the Redis wire protocol. Used as a cache, entries live in memory and are not durably persisted, so the component renders as a stateless workload rather than a stateful one and declares no storage. Realized on Kubernetes, on Compose for local development, on Amazon ElastiCache and on Google Cloud Memorystore for Valkey. Exports its host and port to consumers via the uses topology. On Kubernetes there are two realizations and the Manifest chooses between them. `valkey` synthesizes the cache as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `valkey-chart` renders the community-maintained Valkey Helm chart, which hardens the container further, so reach for it where a chart repository is reachable at build time. Reach for terraform/valkey-elasticache, or cloudformation/valkey-elasticache in a CloudFormation stack, where the cache is an ElastiCache service in the operator's AWS account; each halts naming each value ElastiCache cannot apply. Reach for terraform/valkey-memorystore where it is a Memorystore instance in the operator's Google Cloud project, whose VPC network has a service connection policy for Memorystore; it runs without authentication. `infra.client_tls`, off by default, serves TLS alone and publishes `protocol` `valkey-tls`: on Kubernetes with the certificate attached to `server_certificate`, on ElastiCache with one AWS issues, on Memorystore with one from the instance's own authority, which consumers trust through `server_ca`. The Compose realization refuses it by name.
ОфіційнийdatastoreЗалежні: 50Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/valkey:1.0.0 - v1.0.0
Component
core/datastore/redis
A Redis in-memory key-value cache, speaking the Redis wire protocol. Used as a cache, entries live in memory and are not durably persisted, so the component renders as a stateless workload rather than a stateful one and declares no storage. Realized on Kubernetes (Workload synthesis to native Kubernetes resources), on Compose for local development, on Amazon ElastiCache and on Azure Managed Redis. Exports its host and port to consumers via the uses topology. The terms follow the realization. helm/redis and compose/redis declare the AGPL-3.0 licence; terraform/redis-elasticache declares Apache-2.0, the licence of the module it vendors; cloudformation/redis-elasticache and terraform/redis-azure-managed declare none. The Valkey component speaks the same wire protocol, and its helm Recipes declare BSD-3-Clause. From release 8.0 the container image carries the query engine whose vector search a gateway's semantic cache searches its replies with, and it publishes vector_search while it does; the Valkey image the catalog pins does not. Reach for terraform/redis-elasticache, or cloudformation/redis-elasticache in a CloudFormation stack, for an ElastiCache service in the operator's AWS account, where the Release sets infra.engine_version; and for terraform/redis-azure-managed for an Azure Managed Redis instance in the operator's Azure subscription. None publishes vector_search; each halts naming each value its service cannot apply. `infra.client_tls`, off by default, serves TLS alone and publishes `protocol` `redis-tls`: on Kubernetes with the certificate attached to `server_certificate`, on the managed services with their own. The Compose realization refuses it by name.
ОфіційнийdatastoreЗалежні: 49Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/redis:1.0.0 - v1.0.0
Component
core/datastore/postgres
A PostgreSQL relational database. Provider-agnostic: the same component is implemented by Compose Recipes (local development), Helm Recipes (Kubernetes), Terraform Recipes (AWS RDS, Google Cloud SQL, Azure), and CloudFormation Recipes, and by a reference Recipe that provisions nothing and only wires a consumer to a server that already exists. Reach for the synthesis Recipe on a cluster that runs no PostgreSQL control plane: it needs no controller installed and no chart repository, so it builds with nothing fetched. Reach for the controller-driven Recipe, which emits a CloudNativePG Cluster custom resource, where that control plane is already in place and the cluster should own failover, rolling upgrades and the database's credentials. Reach for the reference Recipe where the server already exists outside the deployment and nothing is to be provisioned. On AWS RDS through Terraform, reach for terraform/postgres-rds-terraform-aws-modules, which vendors the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository, and for terraform/postgres-rds, which authors its own body, where nothing may be fetched or read replicas are needed. On Google Cloud SQL, reach for terraform/postgres-cloud-sql, where the database runs in a Google Cloud project whose VPC network has private services access to Cloud SQL. On Azure, reach for terraform/postgres-flexible-server, a flexible server in a delegated subnet of the operator's virtual network. Backup is the core/backup attachment: the managed Recipes set the service's automated backups from it, 7 days with none attached; on Kubernetes the backup product the Manifest binds realizes it; Compose halts on one.
ОфіційнийdatastoreЗалежні: 48Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/postgres:1.0.0 - v1.0.0
Component
core/datastore/pgvector
A vector database on the PostgreSQL engine: PostgreSQL with the pgvector extension, which adds a vector column type, distance operators and approximate nearest-neighbour indexes, so embeddings are stored beside ordinary rows and queried by similarity in SQL. The server runs the pgvector project's own image, which carries the extension, and the database the component creates has it enabled. Wire here when an application retrieves by similarity and keeps its vectors in PostgreSQL; core/datastore/postgres is the general-purpose relational store, and core/datastore/qdrant, core/datastore/weaviate and core/datastore/chroma are vector stores with APIs of their own. It presents the standard PostgreSQL wire contract (POSTGRES_USER / POSTGRES_PASSWORD / POSTGRES_DB) and exports its host, port, user, database and protocol, and an endpoint joining them into a connection URI that carries no password, so a socket selecting vector stores admits it and a consumer reaches it with a PostgreSQL client. Realized two ways, both self-hosted and both fetching nothing: reach for the Helm Recipe on Kubernetes, which synthesizes a StatefulSet with a persistent volume, and for the Compose Recipe for local development. Its Recipes are verified by their rendered output, not by a started server storing and querying vectors. RDS for PostgreSQL offers pgvector on PostgreSQL 12-19, but no setting of the database instance enables it in a database, only a CREATE EXTENSION statement, so no RDS Recipe realizes this component. Use core/datastore/postgres on its RDS Recipes, the application running CREATE EXTENSION vector in its database; vector-store sockets do not admit it, as nothing the build sees enables the extension.
ОфіційнийdatastoreЗалежні: 53Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/pgvector:1.0.0 - v1.0.0
Component
core/datastore/qdrant
A Qdrant vector database. Stores embedding vectors alongside their payloads and answers nearest-neighbour queries over them, so an application that retrieves reference material by similarity is wired to this component rather than to a general-purpose store. Collections are created through Qdrant's own API at runtime; the component declares the server, not its contents. It is realized on Kubernetes and, for local development, on Docker Compose, and exports its host, its HTTP port and its gRPC port to consumers via the uses topology. On Kubernetes there are two realizations and the Manifest chooses between them. `qdrant` synthesizes the server as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `qdrant-chart` renders the product's own Helm chart, which additionally carries its maintainers' start-up script, data-volume ownership step and pre-stop pause, so reach for it where a chart repository is reachable at build time. That chart also renders an interaction check — a Pod annotated as a test hook, whose script deletes and recreates a collection named `test_collection` on the running server — which the chart offers no setting to decline; the chart realization's own description states what it does and when it runs. While api_key is filled, a scrape of `metrics_endpoint` presents on its connection's `credential` the key attached to read_only_api_key, or the api_key where that slot is empty. The read-only key still reads every collection: it makes a scraper read-only, not metrics-only. `aggrega lint --release` reports such a credential as differing from api_key, as expected.
ОфіційнийdatastoreЗалежні: 54Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/qdrant:1.0.0 - v1.0.0
Component
core/datastore/opensearch
An OpenSearch search and analytics engine. Indexes documents for full-text and analytical queries and indexes embedding vectors beside them, answering similarity searches through its k-NN plugin, so a component that retrieves by similarity can be wired here. Indexes are created through OpenSearch's own API at runtime; the component declares the server, not its contents. It runs as one node. Exports its host, port, protocol, an endpoint joining them and the account a client signs in as. The security plugin stays on with one account, the administrator, whose password is the attached credential. Realized self-hosted by the product's own Helm chart on Kubernetes, fetched at build time, and by a Compose Recipe for local development. Each runs the product's image with no demonstration account or certificate, each node generating its own transport key when it starts; the API is served over plain HTTP, not encrypted, reachable only inside the Release, and with no credential attached the password is a development default. Realized on Amazon OpenSearch Service by terraform/opensearch-service and, in a CloudFormation stack, cloudformation/opensearch-service: each a domain in the operator's AWS account and VPC that serves HTTPS alone on port 443 and answers similarity searches through the service's own k-NN support, and each halts unless core/random-password, with values.symbols true, supplies the password. Reach for them where the engine is the operator's managed AWS service; reach for the self-hosted Recipes on a cluster or a Docker host. The Recipes are verified by their rendered output, not by a started server.
ОфіційнийdatastoreЗалежні: 55Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/opensearch:1.0.0 - v1.0.0
Component
core/datastore/mysql
A MySQL relational database. Provider-agnostic: the same component is implemented by Helm Recipes (Workload synthesis to native Kubernetes resources), Terraform Recipes (AWS RDS, Google Cloud SQL), CloudFormation Recipes, and Compose Recipes (local development). On AWS RDS through Terraform, reach for terraform/mysql-rds-terraform-aws-modules, which vendors the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository; reach for terraform/mysql-rds, which authors its own body, or terraform/mysql-rds-module, which vendors a module kept beside it, where nothing may be fetched, and for terraform/mysql-rds where read replicas are needed. On Google Cloud SQL, reach for terraform/mysql-cloud-sql, which vendors the terraform-google-modules project's published Cloud SQL module, where the database runs in a Google Cloud project whose VPC network already has private services access to Cloud SQL; the instance gets a private address in that network and no public one. Backup is the core/backup attachment, written in the Release. The managed Recipes configure the service's own automated backups from it, bound to terraform/backup-native or cloudformation/backup-native, and keep them for 7 days with none attached. The Helm Recipe is backed up by the backup product the Manifest binds; the Compose Recipe has none, so an attachment halts there.
ОфіційнийdatastoreЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/mysql:1.0.0 - v1.0.0
Component
core/datastore/mysql-operator
The MySQL Operator for Kubernetes control plane, declared as an explicit component so a controller-driven MySQL realization can ship its runtime prerequisite alongside the InnoDBCluster custom resource it emits, rather than assuming an ambient install. The control plane is the controller that reconciles InnoDBCluster custom resources into running MySQL servers and the Services that reach them. Declare this component in the same Aggregate as a core/datastore/mysql-innodb-cluster to satisfy that component's control-plane prerequisite; the operator affirms the prerequisite on the cluster in the same Release. The controller image is pinned to a published release tag, so a Release that sets nothing still names one version rather than whatever the target resolves; a Release pulling from its own mirror overrides it. Realized on Kubernetes by two Helm Recipes. `mysql-operator-chart` renders the project's own chart, custom-resource definitions and the controller's full roles included, and is the one to reach for; its chart repository must be reachable from the build. `mysql-operator` is a minimal bootstrap installing the controller alone, and at the image pinned here its role falls short of what the controller needs.
ОфіційнийdatastoreЗалежні: 40Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/mysql-operator:1.0.0 - v1.0.0
Component
core/datastore/milvus
A Milvus vector database. Stores embedding vectors with their scalar fields and answers similarity searches over them, so a component that retrieves by similarity is wired here. Collections are created through Milvus's own API at runtime; the component declares the server, not its contents. It runs as one standalone server holding its metadata store and its data on one volume, with no separate etcd or object store to deploy. Exports its host, port, protocol, an endpoint joining them and the account a client signs in as. Authentication is off unless a password is attached; attached, it is switched on for the API and for the management endpoints the metrics port serves, whose health check and metrics stay open, and the password is delivered by reference. Realized by the product's own Helm chart on Kubernetes, fetched at build time, and by a Compose Recipe for local development. Its Recipes are verified by their rendered output, not by a started server.
ОфіційнийdatastoreЗалежні: 53Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/milvus:1.0.0 - v1.0.0
Component
core/datastore/mysql-innodb-cluster
A MySQL InnoDB Cluster, realized as an InnoDBCluster custom resource the MySQL Operator for Kubernetes reconciles. The emitted custom resource requests a count of MySQL server instances and a count of MySQL Router instances, and the controller acts on that request inside the target cluster — so what this component describes is what is asked for, never what ends up running. What it publishes is root access at an in-cluster address: the controller establishes the root account from the Secret the custom resource names, and creates neither a logical database nor an application user, so this component declares neither and whoever holds root creates the schemas and accounts an application needs. Reach for core/datastore/mysql where the application wants a logical database and a login role created for it, and where nothing may need installing first — its synthesized Recipe needs no controller and no chart repository, so it builds with nothing fetched. Reach for this component where the MySQL Operator control plane is already installed in the target cluster, which core/datastore/mysql-operator declares and installs.
ОфіційнийdatastoreЗалежні: 40Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/mysql-innodb-cluster:1.0.0 - v1.0.0
Component
core/datastore/memcached
A Memcached in-memory key-value cache. Non-persistent by design: entries live in memory and are not durably stored, so the component renders as a stateless workload rather than a stateful one and declares no storage. Realized by a Helm Recipe (Workload synthesis to native Kubernetes resources), a Compose Recipe (local development), on Amazon ElastiCache and on Google Cloud Memorystore for Memcached. Exports its host and port to consumers via the uses topology. Reach for terraform/memcached-elasticache, or cloudformation/memcached-elasticache in a CloudFormation stack, where the cache is an ElastiCache service in the operator's AWS account; each halts naming each value ElastiCache cannot apply, and takes infra.engine_version as a full version ElastiCache offers, such as 1.6.22. Reach for terraform/memcached-memorystore where the cache is a Memorystore instance in the operator's Google Cloud project, whose VPC network has a private services access connection: one node, its address published, with no TLS and no authentication. Google is deprecating that service, which it shuts down on January 31, 2029. Encrypted client connections are an operator setting, `infra.client_tls`, off by default. Turned on, the cache serves TLS alone on its port and publishes `protocol` `memcached-tls`, so a consumer reading `protocol` learns that it must connect over TLS. On Kubernetes the server presents the Secret attached to the `server_certificate` slot, mounted as files, and asks no client for a certificate; ElastiCache presents a certificate AWS issues; the Compose and Memorystore realizations refuse the setting by name.
ОфіційнийdatastoreЗалежні: 49Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/memcached:1.0.0 - v1.0.0
Component
core/datastore/mariadb
A MariaDB relational database. Provider-agnostic: the same component is implemented by Helm Recipes (Workload synthesis to native Kubernetes resources), Terraform Recipes (AWS RDS), CloudFormation Recipes, Compose Recipes (local development), and a controller-driven Helm Recipe that emits a MariaDB custom resource for the mariadb-operator control plane to reconcile. Reach for the synthesis Recipe on a cluster that runs no MariaDB control plane — it needs no controller installed and no chart repository, so it builds with nothing fetched — and for the controller-driven Recipe where the control plane is already in place and the cluster should own upgrades, rolling restarts, credential rotation and, through infra.ha, replication and failover. On AWS RDS through Terraform, reach for terraform/mariadb-rds-terraform-aws-modules, which vendors the terraform-aws-modules project's published RDS module, where the pull may fetch it from its git repository; reach for terraform/mariadb-rds, which authors its own body, or terraform/mariadb-rds-module, which vendors a module kept beside it, where nothing may be fetched, and for terraform/mariadb-rds where read replicas are needed. Backup is the core/backup attachment, written in the Release. The managed Recipes configure the service's own automated backups from it, bound to terraform/backup-native or cloudformation/backup-native, and keep them for 7 days with none attached. The controller-driven Recipe is backed up through helm/backup-mariadb-operator, the synthesized one by the backup product the Manifest binds; the Compose Recipe has none, so an attachment halts there.
ОфіційнийdatastoreЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/mariadb:1.0.0 - v1.0.0
Component
core/datastore/mariadb-operator
The mariadb-operator control plane, declared as an explicit component so a controller-driven MariaDB realization can ship its runtime prerequisite alongside the MariaDB custom resource it emits, rather than assuming an ambient install. The control plane is the controller that reconciles MariaDB custom resources into running database servers, their client Services and their credential Secrets. Declare this component in the same Aggregate as a controller-driven MariaDB to satisfy the MariaDB component's control-plane prerequisite; the operator affirms the prerequisite on the database in the same Release. Realized on Kubernetes by two Helm Recipes. `mariadb-operator-chart` renders the project's own chart, custom-resource definitions, certificate controller and admission webhook included: reach for it for a full install, where the chart repository is reachable from the build. `mariadb-operator` is a minimal bootstrap installing the controller alone: reach for it where the definitions are already installed in the cluster and the build must fetch nothing. The controller serves its own metrics on its pod. With `values.metrics_enabled` true each Recipe puts their port on a ClusterIP Service in the control plane's namespace and publishes the address as `metrics_endpoint`; false, that key is empty.
ОфіційнийdatastoreЗалежні: 40Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/mariadb-operator:1.0.0 - v1.0.0
Component
core/datastore/ferretdb
A document-database proxy that speaks the MongoDB wire protocol and stores data in a PostgreSQL backend with the DocumentDB extension. It is typed as a stateful server because the endpoint it presents to consumers is a durable document database: standard database consumers bind it through their database socket exactly as they would a native datastore. The proxy process itself holds no local state — its Recipe renders a stateless workload — but the service contract it exposes is a database, so the type label reflects the consumer-facing contract rather than the internal workload form. It fronts a PostgreSQL/DocumentDB backend wired through its `backend` uses-socket: the backend host, port, user, and database arrive on the injection context and are composed into the proxy's PostgreSQL connection. Exports its host, port, and protocol to consumers via the uses topology. With `values.metrics_enabled` true each Recipe makes FerretDB's own metrics, served at /debug/metrics on its debug listener, port 8088, reachable and publishes that address as `metrics_endpoint`; false, that key is empty. That listener checks no caller and also serves /debug/pprof, with CPU profiles and execution traces on request, /debug/vars, /debug/archive and /debug/graphs. On Kubernetes the value moves it from the pod's loopback address onto every interface of the pod and names its port on the Service, newly opening it to every in-cluster caller that reaches the pod; on Docker Compose the value newly opens it to every peer on the network.
ОфіційнийdatastoreЗалежні: 40Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/ferretdb:1.0.0 - v1.0.0
Component
core/datastore/etcd
etcd, a distributed, strongly-consistent key-value store. This component runs a single-node etcd for development and lightweight coordination, holding its data on a persistent volume. It serves the v3 client API over HTTP with no authentication, suited to in-cluster consumers that need a simple consistent key-value store or configuration backend. Provider- agnostic: the same component is implemented by a Helm Recipe (synthesized manifests on Kubernetes) and a Compose Recipe (local development). Exports its host and client port to consumers via the uses topology. etcd serves its own metrics in the Prometheus format at /metrics on its client port, on by default, and every Recipe publishes that address as `metrics_endpoint`.
ОфіційнийdatastoreЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/etcd:1.0.0 - v1.0.0
Component
core/datastore/elasticsearch
An Elasticsearch search and analytics engine. Indexes documents for full-text and analytical queries and indexes embedding vectors beside them in dense-vector fields, answering similarity searches, so a component that retrieves by similarity can be wired here. Indexes are created through Elasticsearch's own API at runtime; the component declares the server, not its contents. It runs as one node. Exports its host, port, protocol, an endpoint joining them and the account a client signs in as. Authentication stays on with the built-in superuser, elastic: its password is the attached credential, which the Kubernetes Recipe reads from its Secret by reference, and with none attached it is a development default that attaching one replaces. The server's own first-start setup, which generates certificates and prints a password, is switched off. The API is served over plain HTTP, not encrypted, and is reachable only inside the Release. Elastic publishes the image under the Elastic License 2.0. Realized on Kubernetes by a Recipe that synthesizes the server from that image, and by a Compose Recipe for local development; the host's vm.max_map_count kernel setting is the operator's to raise. Its Recipes are verified by their rendered output, not by a started server.
ОфіційнийdatastoreЗалежні: 53Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/elasticsearch:1.0.0 - v1.0.0
Component
core/datastore/documentdb
A PostgreSQL-backed document store: PostgreSQL 17 with the DocumentDB extension. It presents the standard PostgreSQL wire contract (POSTGRES_USER / POSTGRES_PASSWORD / POSTGRES_DB) and persists data on a volume, so it is a stateful server. It is the storage backend for a document proxy that speaks a document-database wire protocol on top of it; the proxy connects over the exported host, port, user, and database. Exports its host, port, user, database, and protocol to consumers via the uses topology.
ОфіційнийdatastoreЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/documentdb:1.0.0 - v1.0.0
Component
core/datastore/chroma
A Chroma vector database. Stores embeddings with their documents and metadata in named collections and answers nearest-neighbour queries over them through Chroma's HTTP API, so an application that retrieves reference material by similarity is wired here rather than to a general-purpose store. Collections are created through that API at runtime; the component declares the server, not its contents. Exports its host, port and protocol via the uses topology. The server checks no credential: at 1.5.9, the release this component runs by default, any client that reaches its address reads and writes every collection. Both Recipes therefore publish it only inside the Release — a ClusterIP Service on Kubernetes, the compose network locally — and a Release that exposes it further puts something that authenticates clients in front of it. Realized two ways, both self-hosted and both fetching nothing: reach for the Helm Recipe on Kubernetes, which synthesizes a StatefulSet with a persistent volume, and for the Compose Recipe for local development. Its Recipes are verified by their rendered output, not by a started server storing and querying vectors.
ОфіційнийdatastoreЗалежні: 53Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/chroma:1.0.0 - v1.0.0
Component
core/datastore/cognee
A semantic-memory store for agent deployments. It ingests text and documents and turns them into a typed graph of entities and the relations between them, kept alongside embeddings of the same material, so a consumer asks what relates to what rather than only what looks similar. Construction is the point: a model reads each ingested item, names the entities and relations it finds, and the result is written into the graph, optionally grounded against an ontology the deployment supplies so extracted names resolve onto terms the deployment already uses. The graph, the embeddings and the bookkeeping tables are held by this component itself, on its own persistent volume, which is why it is a stateful server rather than a façade over someone else's store. The model it reads with is wired in over the uses topology and its address is a deterministic pointer string the running server resolves per request; the compiler never contacts a model. Realized on Kubernetes via its Helm Recipe, which runs the official upstream image directly and carries the whole memory configuration — the graph engine, the embedding store, the extraction and embedding models, and the ontology grounding — into the rendered artifact. Exports its host, its port, its base URL and its protocol to consumers via the uses topology.
ОфіційнийdatastoreЗалежні: 52Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/datastore/cognee:1.0.0 - v1.0.0
Component
core/compute/worker
A generic background-worker primitive. Its operational type is `worker` — long-running, stateless compute that pulls from queues or runs background loops with zero inbound application ports — and it adds two optional integer fields, `queue_depth_hint` and `concurrency`, to the shared base schema's documentation knobs. An Aggregate declares this for a component that consumes work rather than accepting inbound application traffic, which is what `server` is for. Declaring no `provides`, it publishes nothing across an edge. On a Kubernetes cluster it is a Deployment. In an AWS account with an ECS cluster and a VPC, terraform/worker-ecs and cloudformation/worker-ecs run it as an Amazon ECS service on Fargate. In a Google Cloud project with a VPC network, terraform/worker-cloud-run runs it as a Cloud Run worker pool; in an Azure subscription with a Container Apps environment, terraform/worker-container-apps as a container app with no ingress. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The image is the Release's own code, so an edge from this component delivers what its `inject:` block dictates: `env` entries become variables everywhere, beside both values' default names, which the component's own `env` entries may replace and, while each is the default, no edge may dictate; `file` entries files on Kubernetes and Container Apps alone; and `arg` or any other mechanism is refused by name, as no realization renders a command line. A `core/secret-value` attachment's value is rendered in the clear on every realization; no realization delivers `core/referenced-secret` or `core/random-password` to it, and the helm realization refuses either by name.
ОфіційнийcomputeЗалежні: 9Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/compute/worker:1.0.0 - v1.0.0
Component
core/compute/knative-serving
The Knative Serving control plane, declared as an explicit component so a scale-to-zero realization ships its runtime prerequisite alongside the custom resources it emits. Its controllers reconcile Service, Configuration, Revision and Route into Deployments, hold requests for a revision at zero replicas, and set replica count from observed concurrency. Declaring it in that same Aggregate is what satisfies the prerequisite. Realized on Kubernetes via its Helm Recipe, a minimal bootstrap installing the four control-plane workloads and neither the custom-resource definitions nor the ingress layer. What it publishes across an edge is where Prometheus finds its own metrics, which are off by default. With `values.metrics_enabled` true the Recipe switches them on, and each of the four processes serves them in pods of its own, on a port no Service carries, so the Recipe publishes the pod label selecting all four, the name of their metrics port and the path, for a scraper that discovers those pods directly; false, those keys are empty.
ОфіційнийcomputeЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/compute/knative-serving:1.0.0 - v1.0.0
Component
core/compute/keda
The KEDA control plane, declared as an explicit component so an event-driven realization ships its runtime prerequisite alongside the custom resources it emits. Its operator reconciles ScaledObject and ScaledJob into horizontal pod autoscalers, and its metrics adapter answers the external-metrics API, so a workload scales on queue depth or event rate down to zero. Declaring it in that same Aggregate satisfies the prerequisite. Realized on Kubernetes by two Helm Recipes. `keda-chart` renders the project's own chart, custom-resource definitions and admission webhooks included, and is the one to reach for; its chart repository must be reachable from the build. `keda` is a minimal bootstrap installing the operator and adapter without the definitions or the webhooks, and at the images pinned here neither process starts, because it sets no WATCH_NAMESPACE. What it publishes across an edge is where Prometheus finds its own metrics. The operator and the adapter each serve them behind a Service of their own, under names fixed for the cluster, so each Recipe publishes both Service names, the name of their metrics port and the path, for a scraper that discovers every endpoint behind them.
ОфіційнийcomputeЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/compute/keda:1.0.0 - v1.0.0
Component
core/ai-tool/search
A read-only tool that answers an agent's question from the public web through a managed search API, so a deployment can reach current information without operating a crawler or an index of its own. The developer names the search provider, defaulting to Tavily, and how many results a query returns, defaulting to five. Declared as a resource: the search service is operated outside the deployment, which consumes it rather than running it. Its effect is pinned read-only in its own schema. Realized on Kubernetes and on Docker Compose as a reference realization, selected in the Manifest: neither provisions anything and no container runs for the tool. On Kubernetes the helm provider emits an ExternalName Service mapping an in-cluster name onto the search API's base address; on Compose no service is defined for it, and the address reaches the consuming service through the injection context instead. Both publish the same base address, provider name and result count, so a consumer wires to either unchanged.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/search:1.0.0 - v1.0.0
Component
core/compute/function
A generic event-driven compute primitive. Its operational type is `function` — event-driven compute that scales to zero — and it adds a required `runtime` value to the shared base schema's optional `display_name` and `description` knobs. An Aggregate whose work is event-triggered and idles at no capacity declares this rather than `server`, which expects inbound application traffic, or `task`, which runs once and ends. Declaring no `provides`, it publishes nothing across an edge. On Kubernetes it is realized two ways, chosen in the Manifest: as a Knative Service, or as a Deployment paired with a KEDA scaling resource; either control plane is a prerequisite the operator affirms in the Release, and the build halts while it is unaffirmed. Where the target is an AWS account, terraform/function-lambda and cloudformation/function-lambda run it as an AWS Lambda function from a container image in Amazon ECR, and emit no trigger and no URL: what invokes it is the operator's. Nothing renders until the Release sets `infra.image`. The image is the Release's own code, so an edge delivers what its `inject:` block dictates, and no declared socket narrows the mechanism: `env` entries become variables beside `FUNCTION_RUNTIME`, `runtime`'s default name, which the component's own `env` entries may replace and, while it is the default, no edge may dictate; `file` entries read-only files on Kubernetes only; and `arg` or any other mechanism is refused by name, as no realization renders a command line. A `core/secret-value` attachment's value is rendered in the clear; no realization delivers `core/referenced-secret` or `core/random-password` to it, and the two helm realizations refuse either by name.
ОфіційнийcomputeЗалежні: 5Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/compute/function:1.0.0 - v1.0.0
Component
core/ai-tool/openapi
A tool built from an API's own description: the developer supplies the location of an OpenAPI document, and the operations that document declares become what the agent may call. An Aggregate adds a whole HTTP API to an agent by pointing at its specification, rather than declaring one tool per endpoint. Marked mutating, since the described operations may include ones that write. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/openapi:1.0.0 - v1.0.0
Component
core/ai-tool/model-invocation
A read-only tool that lets an agent call a model as a step in its own work — turning text into speech, transcribing audio, or whatever task the wired model performs — rather than as the agent's own reasoning engine. It declares a required socket onto a model component, so the model is resolved in the topology. The task the invocation performs is named by the developer and defaults to text-to-speech. Its effect is pinned read-only in its own schema. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The model's address reaches the tool's code under default variable names; an `inject:` block on the connection delivers instead, as variables or files, the names that code reads. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/model-invocation:1.0.0 - v1.0.0
Component
core/ai-tool/filesystem
A tool that gives an agent a file tree it can work in, listing, reading and, where the Aggregate allows it, writing files. Access is chosen by the developer and defaults to read-only, so the ability to write is granted deliberately rather than inherited, and the declared effect can be narrowed to match. Marked mutating by default, because a writable file tree outlives the invocation that changed it. The tree sits on a disk this component carries, so a file written in one call is still there for the next — which is why an agent that needs a working area to survive between calls is wired to this component rather than to a tool that keeps nothing. One instance holds that disk, so every call reaches the same files. Realized on Kubernetes via its Helm Recipe, which synthesizes that one instance with its own disk and a ClusterIP Service fronting it; helm-only. Exports the Service's host, port, endpoint and protocol to consumers via the uses topology, so an agent wires to the tool the same way it wires to any other service. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files outside the tree, on every realization that runs the tool. The image serving the tool is the Release's, because this component packages no upstream product. Nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 13Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/filesystem:1.0.0 - v1.0.0
Component
core/ai-tool/file-parser
A read-only tool that turns a document into text an agent can read, so a deployment can put contracts, reports and scanned records in front of an agent without teaching it each file format. The parsing engine is chosen by the developer, and the extraction mode selects how much structure survives — text only by default, where layout is discarded and prose is kept. Its effect is pinned read-only in its own schema: parsing never alters the source document. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/file-parser:1.0.0 - v1.0.0
Component
core/ai-tool/grpc
A tool that lets an agent call a gRPC service. The developer supplies the location of the service's protobuf definition, and the method set that definition declares becomes what the agent may call, so the callable surface comes from the service's own contract rather than from a hand-written tool description. Marked mutating, since a remote procedure call may change state in the service it reaches. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and those are the coordinates of the tool itself rather than of the service it goes on to call. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/grpc:1.0.0 - v1.0.0
Component
core/ai-tool/event-publisher
A tool that lets an agent emit an event rather than call for a result, so a deployment can hand work off to whatever consumes the stream instead of waiting on a reply. The developer names the topic the event is published to, and may reference a schema the payload is expected to satisfy. It declares a required socket onto a message broker the deployment runs, so the broker it publishes to is resolved in the topology. Marked mutating: a published event is visible to its consumers and cannot be withdrawn. Realized on Kubernetes as a Knative Service: the control plane admits a call while the revision runs no replicas and sets the replica count from the concurrency it observes, so the tool occupies no compute between the calls that invoke it. It publishes the host, port, endpoint and protocol an agent reaches it on. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The broker's address reaches the tool's code under default variable names; an `inject:` block on the connection delivers instead, as variables or files, the names that code reads. A broker serving TLS publishes a protocol saying so, and the connection's `server_ca` slot gives the tool the authority to check it against, as a mounted file. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/event-publisher:1.0.0 - v1.0.0
Component
core/ai-tool/datastore-query
A read-only tool that answers an agent's questions by querying a datastore the topology has wired to it. The query dialect is chosen by the developer and defaults to SQL. It declares a required socket onto a datastore component, so the store it reads is resolved as a topology edge rather than supplied as a connection string in a prompt, and its effect is pinned read-only in its own schema — it cannot be widened to write. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The store's address reaches the tool's code under default variable names; an `inject:` block on the connection delivers instead, as variables or files, the names that code reads. A store served over TLS alone halts the build on a connection writing none. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/datastore-query:1.0.0 - v1.0.0
Component
core/ai-tool/computer-use
A tool that presents an agent with a virtual desktop it can see and control, taking screenshots and issuing pointer and keyboard input, so a deployment can automate applications that expose neither an API nor a web page. The desktop's screen geometry is set by the developer, defaulting to 1024 by 768, because the agent reasons over the pixels it is given. Declared as a server: the desktop session is a long-running workload the agent connects to. Marked mutating, since that input drives real applications. Realized on Kubernetes as a single-replica Deployment fronted by a ClusterIP Service, so one desktop stays resident across an agent's turns and every screenshot comes from the session the previous input acted on. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/computer-use:1.0.0 - v1.0.0
Component
core/ai-tool/email
A tool that gives an agent a mailbox. The developer chooses how much of it the agent gets: sending only, reading only, or both, so a deployment can grant an agent the ability to report outward without also granting it sight of an inbox. Marked mutating, since sending mail acts outside the deployment and cannot be recalled. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/email:1.0.0 - v1.0.0
Component
core/ai-tool/code-interpreter
A tool that executes code an agent has written and returns what it produced, so a deployment can give an agent arithmetic, data manipulation and file transformation without declaring each of those steps as its own tool. The language runtime is chosen by the developer and defaults to Python. Nothing in its contract carries state between executions; each invocation is evaluated on its own. Marked mutating, because executed code can act beyond returning a value. Realized on Kubernetes as a Knative Service, which admits an inbound call while the tool is running no replicas, so the tool occupies no compute between the calls that invoke it. It publishes a host, port, endpoint and protocol, and a consumer wires to those without naming how the tool runs. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as read-only files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/code-interpreter:1.0.0 - v1.0.0
Component
core/ai-tool/browser
A tool that drives a real web browser on an agent's behalf, navigating pages, reading rendered content and interacting with page elements, so a deployment can reach sites that offer no API of their own. Automation runs through a browser-automation engine, defaulting to Playwright, against a browser build defaulting to Chromium, headless unless the Aggregate asks for a visible session. Marked mutating: a browsing session can submit forms and trigger actions on the sites it visits. Realized on Kubernetes as a long-running Deployment fronted by a ClusterIP Service, publishing that Service's host, port, endpoint and protocol so an agent wires to the tool the same way it wires to any other service. An edge from the tool, on a socket the author names, delivers what its `inject:` block dictates, as variables or as files, on every realization that runs the tool. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The component's own `inject:` block takes `env` and `file` entries too, and one reading a value of the tool alone replaces that value's default variable name.
Офіційнийai-toolЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/browser:1.0.0 - v1.0.0
Component
core/ai-tool/agent-invocation
A tool letting one agent delegate a step to another over an agent-to-agent protocol, A2A by default: the caller waits synchronously for a result, asynchronously for a handle it polls, or not at all, fire-and-forget. Its required agent socket resolves the target by topology, not prompt. Marked mutating: a delegated agent may act in the world. Realized on Kubernetes as a Knative Service scaled to zero between calls yet admitting each, it publishes a host, port, endpoint and protocol for a consumer to wire to. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The agent's address reaches the tool's code under default variable names; an `inject:` block on the connection delivers instead, as variables or files, the names that code reads. The component's own `inject:` block adds `env` and `file` entries beside them, and one reading a tool value alone replaces that value's default name. The connection may hold a credential, presented under the agent's published credential_header, its Secret holding the whole value: an API key in a header or cookie, or an HTTP scheme such as Bearer or Basic; a client_secret, for the OAuth 2.0 client-credentials grant, directly or through OpenID Connect; and, for mutual TLS, a client_certificate, mounted as files. A server_ca, also mounted, names the agent's server certificate authority. The build halts on a connection to an agent labelled managed with none of credential, client_secret, client_certificate or unauthenticated: true, and on a credential or client_secret to one labelled external publishing protocol http; to an agent inside the Release, http carries the credential or token unencrypted on the cluster network.
Офіційнийai-toolЗалежні: 2Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-tool/agent-invocation:1.0.0 - v1.0.0
Component
core/ai-pipeline/temporal
Temporal, a durable-execution engine. A multi-step job is written as ordinary program code against one of the project's client libraries; the engine records every step's outcome as an event and hands the next piece of work to a worker the application itself deploys. Because the run is a replayable event history rather than a position in a process, a worker that crashes mid-job is replaced and the job resumes where it stopped, and a step that fails is retried on the policy the code declared. That division is what separates this from a scheduler: the engine owns the run's history and its timers, while the application owns the code each step executes. It keeps none of its own durable state. A relational datastore holds the event histories and a second database on that same connection holds the visibility index the console and the list APIs read, so the store is wired as a topology edge to a datastore component the Aggregate already declares rather than bundled inside this one. The engine is reached over gRPC on the frontend service, which is the address the provides contract publishes; a browser console is deployed alongside it. Realized on Kubernetes through a Helm Recipe that sources the project's own chart. It has no standalone local-container form, so no Compose Recipe is provided. Each of the engine's services serves its own metrics, in pods of its own, so the Recipe publishes the pod labels selecting every one of them, the name of their metrics port and the path, for a scraper that discovers those pods directly.
Офіційнийai-pipelineЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-pipeline/temporal:1.0.0 - v1.0.0
Component
core/ai-pipeline/argo-workflows
Argo Workflows, a workflow engine for Kubernetes. A run is submitted as a Workflow custom resource naming the steps to take and how they depend on one another; the workflow controller watches those resources and reconciles each one, starting a pod per step, retrying a step that fails, fanning a step out over a list of inputs, and recording each step's outputs so a later step can read them. The cluster's own API holds the run, so the engine keeps no separate run database. An API server is deployed alongside the controller and serves the submission and query API and the web console; consumers reach it through the provides contract. Argo Workflows is realized on Kubernetes through a Helm Recipe that sources the project's own chart; it has no standalone local-container form, so no Compose Recipe is provided. With `values.metrics_enabled` true the Recipe serves the workflow controller's metrics, at /metrics, over plain HTTP on a ClusterIP Service of their own, and publishes that address as `metrics_endpoint`; false, that key is empty. The port checks no caller.
Офіційнийai-pipelineЗалежні: 12Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-pipeline/argo-workflows:1.0.0 - v1.0.0
Component
core/ai-model/vertex-ai
Google Vertex AI managed generative model endpoint, realized as an external reference (no in-cluster compute). The regional endpoint is operator-supplied. Authentication is the consumer's concern: workloads use ambient GCP workload identity or inject their own credentials, rather than a static key delivered by the realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/vertex-ai:1.0.0 - v1.0.0
Component
core/ai-model/vllm
vLLM running in-cluster as a self-hosted inference server. A Recipe synthesizes the runtime directly: a weights-pull initContainer resolves the model pointer into a model-cache volume the serving container mounts, and a ClusterIP Service fronts the OpenAI-compatible API. The model value is a deterministic pointer resolved at pod start, never fetched at compile time. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint. Each realization also publishes where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, where one address cannot reach every replica, the Service or the pod labels a scraper discovers them by. On KServe the cluster's serving runtime decides the metrics port and path, so nothing is published there. On a cluster running the Prometheus Operator, `core/prometheus-exporter` attached to the synthesized Kubernetes realization, its port set to `http`, emits a ServiceMonitor for it instead.
Офіційнийai-modelЗалежні: 59Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/vllm:1.0.0 - v1.0.0
Component
core/ai-model/vertex-ai-embedding
Google Vertex AI managed embedding model endpoint, realized as an external reference (no in-cluster compute). The regional endpoint is operator-supplied. Authentication is the consumer's concern: workloads use ambient GCP workload identity or inject their own credentials, rather than a static key delivered by the realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/vertex-ai-embedding:1.0.0 - v1.0.0
Component
core/ai-model/tei
Text Embeddings Inference running in-cluster as a self-hosted embedding server. A Recipe synthesizes the runtime directly: a weights-pull initContainer resolves the model pointer into a model-cache volume the serving container mounts, and a ClusterIP Service fronts the embeddings API. The model value is a deterministic pointer resolved at pod start, never fetched at compile time. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so a consumer reaches the embedding endpoint without hardcoding it. Each Recipe publishes where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, while the Kubernetes realization runs more than one replica, the Service a scraper discovers every replica behind. At the default image tag the api_key slot's key guards them too, so while that slot is filled a scrape presents the key on its connection's `credential` slot. On a cluster running the Prometheus Operator, `core/prometheus-exporter` attached with its port set to `http` emits a ServiceMonitor for it instead.
Офіційнийai-modelЗалежні: 57Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/tei:1.0.0 - v1.0.0
Component
core/ai-model/tgi
Text Generation Inference running in-cluster as a self-hosted inference server. A Recipe synthesizes the runtime directly: a weights-pull initContainer resolves the model pointer into a model-cache volume the serving container mounts, and a ClusterIP Service fronts the API. The model value is a deterministic pointer resolved at pod start, never fetched at compile time. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint. Each Recipe also publishes where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, while the Kubernetes realization runs more than one replica, the Service a scraper discovers every replica behind. On a cluster running the Prometheus Operator, `core/prometheus-exporter` attached with its port set to `http` emits a ServiceMonitor for it instead.
Офіційнийai-modelЗалежні: 57Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/tgi:1.0.0 - v1.0.0
Component
core/ai-model/sagemaker
Amazon SageMaker managed model serving endpoint, realized as an external reference (no in-cluster compute). The regional endpoint is operator-supplied. Authentication is the consumer's concern: workloads use ambient AWS IAM (for example EKS IRSA) or inject their own credentials, rather than a static key delivered by the realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/sagemaker:1.0.0 - v1.0.0
Component
core/ai-model/replicate
Replicate's managed prediction API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Replicate addresses a specific published build of a model rather than a moving name, so beside the model the component exposes the version an Aggregate pins to, and whether a prediction is streamed back as it is produced or returned once it is complete. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/replicate:1.0.0 - v1.0.0
Component
core/ai-model/openai
OpenAI's managed chat-completions API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. It is realized as a direct reference — on Kubernetes an ExternalName Service mapping an in-cluster name onto the managed host, on Compose the endpoint carried to consumers with no service run — which delivers no credential: a consumer that needs the key presents it from the connection it reaches the model on. The endpoint defaults to the managed API's base URL, version path included, so a bare declaration is enough, and it sits on `infra` rather than inside a Recipe so an operator may repoint it at a gateway or proxy without editing the Aggregate. Consumers read the endpoint and the model name over a `uses` edge — the contract every model here publishes — so an agent reaches it with no endpoint in its own code. A gateway reaches it the same way, wired to the gateway's models socket with the key on that connection, so the gateway holds the key and an application behind the gateway never sees it.
Офіційнийai-modelЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/openai:1.0.0 - v1.0.0
Component
core/ai-model/openai-embedding
OpenAI's managed embeddings API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. It publishes the same endpoint-and-model contract the generative models publish, so an application that embeds text reads exactly the keys it would read to generate it, and the output width is a declared value where an Aggregate wants vectors narrower than the model's own default. The endpoint defaults to the managed API's base URL, version path included, and sits on `infra`, so a bare declaration is enough and an operator may still repoint it at a gateway or proxy. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/openai-embedding:1.0.0 - v1.0.0
Component
core/ai-model/openai-compatible
A broker speaking the OpenAI API, reached as an external reference: no compute runs in-cluster, and what answers behind the endpoint is whatever a Release points it at — a proxy running in the cluster, a gateway, or any service serving that API. It is the component to reach for when an application should not name a provider at all: the endpoint is operator-supplied in the Release and the model is an alias the broker resolves, so which provider actually answers can change with neither the Aggregate nor the application changing. The provider name it carries records which broker sits behind it, for a reader and for a selector to match on rather than for the compiler, which pipes the strings and reads neither. Consumers read the endpoint and the model name over a `uses` edge, the contract every model here publishes. It holds no credential: a consumer that must present a key to the service behind the endpoint declares that key as `credential` on the socket that reaches the endpoint and presents it itself.
Офіційнийai-modelЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/openai-compatible:1.0.0 - v1.0.0
Component
core/ai-model/ollama
Ollama running in-cluster as a self-hosted inference server. A Recipe synthesizes the runtime directly: a weights-pull initContainer resolves the model pointer into a model-cache volume the serving container mounts, and a ClusterIP Service fronts the OpenAI-compatible API. The model value is a deterministic pointer resolved at pod start, never fetched at compile time. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint. On Kubernetes there are two realizations and the Manifest chooses between them. `ollama` synthesizes the runtime as a workload and fetches nothing, and it is the one to reach for where the build must complete with no chart repository reachable, or where more than one serving replica is wanted — it gives each replica its own model cache. `ollama-chart` renders the community-maintained Ollama Helm chart, which additionally renders a horizontal autoscaler from the replica bounds below, so reach for it where a chart repository is reachable at build time. That realization backs its single serving replica with one model cache and refuses a higher replica count rather than rendering one that cannot schedule. Both realizations run liveness and readiness probes against the serving endpoint, so a rollout reports the server ready only once it answers.
Офіційнийai-modelЗалежні: 58Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/ollama:1.0.0 - v1.0.0
Component
core/ai-model/nvidia-nim
NVIDIA Inference Microservice (NIM) as a self-hosted, GPU-native inference server serving the OpenAI-compatible API. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: every realization runs a NIM image the reader must be entitled to pull, named by the Release in `infra.image`. The catalog names none, and nothing renders until it is set. On Kubernetes three Recipes realize it. `helm/nvidia-nim` assembles the workload itself and fetches nothing at build time: reach for it where no NIM Operator runs. `helm/nvidia-nim-chart` renders NVIDIA's `nim-llm` chart: reach for it where that chart is wanted and NVIDIA's chart repository is reachable from the build. `helm/nvidia-nim-service` emits a `NIMService` for NVIDIA's NIM Operator: reach for it where the operator runs, or is declared beside it as `core/ai-model/nvidia-nim-operator`, and affirm that with `infra.control_plane_present`. On a single Docker host, `compose/nvidia-nim` serves it through the NVIDIA Container Toolkit. Exports its host, port, endpoint, model, and protocol to consumers via the uses topology, so an agent reaches the model without hardcoding an endpoint, and where Prometheus finds the server's metrics: one address as `metrics_endpoint`, or, while it runs more than one replica on Kubernetes, the Service a scraper discovers every replica behind.
Офіційнийai-modelЗалежні: 59Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/nvidia-nim:1.0.0 - v1.0.0
Component
core/ai-model/nvidia-api
The managed NVIDIA API (integrate.api.nvidia.com) generative model endpoint, realized as an external reference (no in-cluster compute). Distinct from the self-hosted NVIDIA Inference Microservice component: this consumes NVIDIA's hosted, OpenAI-compatible API rather than running an inference server.
Офіційнийai-modelЗалежні: 46Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/nvidia-api:1.0.0 - v1.0.0
Component
core/ai-model/nvidia-nim-operator
NVIDIA's NIM Operator, declared as an explicit component so the `NIMService` realization of `core/ai-model/nvidia-nim` can ship its runtime prerequisite in the same Aggregate rather than assuming an ambient install. The control plane is the operator that reconciles NIMService custom resources into running NIM servers. Declare this component beside the NIM model and affirm the prerequisite on the model with `infra.control_plane_present` in the Release. Using it needs an NVIDIA account and an NGC API key, under NVIDIA's licence terms: NVIDIA publishes the operator's chart and image on its own registries, and every NIM server it runs pulls a NIM image the reader must be entitled to pull. The cluster must already run NVIDIA's GPU Operator, which NVIDIA names as the operator's prerequisite; the catalog does not install it and no build checks for it. Realized on Kubernetes by `helm/nvidia-nim-operator`, which renders NVIDIA's own `k8s-nim-operator` chart, its custom-resource definitions and roles included; its chart repository must be reachable from the build. The chart names its roles and accounts itself, so one installation runs per cluster, and the operator watches every namespace.
Офіційнийai-modelЗалежні: 20Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/nvidia-nim-operator:1.0.0 - v1.0.0
Component
core/ai-model/mistral
Mistral AI's managed chat API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none — declaring the component and choosing a model is the whole of what an Aggregate writes. Consumers read that endpoint and the model name over a `uses` edge, the contract every model here publishes, so moving an application onto this provider is a change of component type and nothing else. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/mistral:1.0.0 - v1.0.0
Component
core/ai-model/nemoguard-content-safety
NVIDIA's NemoGuard content-safety model, Llama 3.1 NemoGuard 8B ContentSafety, served by its NIM: a model that reads a conversation, judges the user's message safe or unsafe against a taxonomy of harm categories and names the categories it breaks. core/security/nemo-guardrails points its content-safety rail at it through its content_safety socket. It is core/ai-model/nvidia-nim under a name of its own: it extends that component and takes its values, infra and provides exactly, so every Recipe of the NIM realizes it — helm/nvidia-nim, helm/nvidia-nim-chart and helm/nvidia-nim-service on Kubernetes, compose/nvidia-nim on Docker — and a Manifest maps it to one of them. Running it requires an NVIDIA AI Enterprise licence or NVIDIA Developer Program membership, and an NGC API key: NVIDIA restricts access to the image and the model. The Release names the image in infra.image and the catalog names none, so nothing renders until it is set. NVIDIA states that the model needs 48 GB of GPU memory, on one GPU or two.
Офіційнийai-modelЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/nemoguard-content-safety:1.0.0 - v1.0.0
Component
core/ai-model/litellm
LiteLLM running in-cluster as a self-hosted, OpenAI-compatible model proxy. The published alias in values.model is the name the proxy serves under, and it forwards each request to the provider-qualified upstream model in values.upstream_model, so a consumer reaches one stable in-cluster endpoint and name while the target behind it changes independently. The proxy holds the upstream credential attached to its provider_key slot and presents it on a caller's behalf. Exports its host, port, endpoint, model and protocol, the contract every other model component publishes. Because it holds that credential, its Recipes forward only the request fields the proxy declares for the route, plus any the Release admits through infra.admitted_request_fields, and refuse the rest with an error naming the field. They answer only the model routes, the model listing, the health checks and key management, and refuse every other path, so its token-count, pass-through and other administrative routes, its console and model paths carrying the model name are not available through any. On Kubernetes, reach for helm/litellm, which synthesizes the proxy, where the build must fetch nothing, and for helm/litellm-chart, BerriAI's chart, where its registry is reachable; for local development, the compose Recipe. With `values.metrics_enabled` true each Recipe has the proxy serve its Prometheus metrics, its path allowlist forwarding a GET of /metrics alone to them from port 9090, and publishes that address as `metrics_endpoint`; false, that key is empty. The port checks no caller, and the metrics name the hashed keys, key aliases and user emails the proxy attributes spend to.
Офіційнийai-modelЗалежні: 58Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/litellm:1.0.0 - v1.0.0
Component
core/ai-model/huggingface
Hugging Face's managed Inference API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. It fronts models of several kinds rather than one, so beside the model name the component exposes the task a request is for — generating text, transforming it, or returning features — which is what decides the shape of the answer. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/huggingface:1.0.0 - v1.0.0
Component
core/ai-model/gemini
Google's managed Gemini API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none — declaring the component and choosing a model is the whole of what an Aggregate writes. Consumers read that endpoint and the model name over a `uses` edge, the contract every model here publishes, so an agent reaches it with no endpoint in its own code. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 43Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/gemini:1.0.0 - v1.0.0
Component
core/ai-model/docker-model-runner
A model served by Docker Model Runner, named by its OCI model reference (`ai/smollm2`, say), behind an OpenAI-compatible API. The same reference is served on both targets, so one Aggregate builds for Docker Compose and for Kubernetes and asks both for the same model. The reference is a deterministic pointer the runner resolves when it pulls the model; the compiler never fetches it. On Docker Compose the host's own Docker Model Runner serves it: the compose file declares the model in its top-level `models` element, and Compose pulls and configures it when the stack starts. That needs what no build can check: Docker Compose 2.38 or later, Docker Model Runner enabled on the host, and a model CLI of at least 1.0.6 for `runtime_flags` to reach the engine. On Kubernetes the Docker Model Runner server runs in-cluster, pulls the same reference into a per-replica model cache, and serves it behind a ClusterIP Service. Consumers read the endpoint, model, host, port and protocol over a `uses` edge. Each operator setting works on the target where it means something and is refused by name on the other.
Офіційнийai-modelЗалежні: 57Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/docker-model-runner:1.0.0 - v1.0.0
Component
core/ai-model/databricks
A Databricks model serving endpoint, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Serving is per-workspace, so the host is operator-supplied in the Release while the served endpoint is named separately as a value — which is what lets one workspace serve several models and an Aggregate say which of them it means. Consumers read the endpoint and the model name over a `uses` edge, the contract every model here publishes, so an application reaching a model served from a data platform is wired exactly like one reaching a public API. Authentication is the consumer's concern: a workload attaches its own credential rather than receiving one from this realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/databricks:1.0.0 - v1.0.0
Component
core/ai-model/bedrock
Amazon Bedrock managed generative model endpoint, realized as an external reference (no in-cluster compute). The regional endpoint is operator-supplied. Authentication is the consumer's concern: workloads use ambient AWS IAM (for example EKS IRSA) or inject their own credentials, rather than a static key delivered by the realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/bedrock:1.0.0 - v1.0.0
Component
core/ai-model/cohere
Cohere's managed Command API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none; beyond the model name the component exposes the connectors a request may consult, which is how this provider lets a generation draw on an external source while it is being produced. Consumers read the endpoint and the model name over a `uses` edge, the contract every model here publishes. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/cohere:1.0.0 - v1.0.0
Component
core/ai-model/bedrock-embedding
Amazon Bedrock managed embedding model endpoint, realized as an external reference (no in-cluster compute). The regional endpoint is operator-supplied. Authentication is the consumer's concern: workloads use ambient AWS IAM (for example EKS IRSA) or inject their own credentials, rather than a static key delivered by the realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/bedrock-embedding:1.0.0 - v1.0.0
Component
core/ai-model/azure-openai
Azure OpenAI's managed chat API. Its host belongs to the Azure OpenAI resource an organisation runs rather than to the provider, and a model is addressed by the deployment name given to it inside that resource rather than by the model's own name — both of which this component exposes, together with the API version a request is made against. Consumers read the endpoint and the model name over a `uses` edge, the contract every model here publishes, so an application moved onto this provider reads the same two keys it read before. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model. The helm and compose realizations reference a resource that already exists: no compute runs, the compiler never calls the provider, and the Release supplies the endpoint in infra.endpoint. terraform/azure-openai creates the resource and the model deployment in the resource group, region and subdomain the Release names in infra.overrides, and publishes the endpoint composed from that subdomain. Reach for a reference where the resource exists outside the Release; reach for terraform where the Release is to create it.
Офіційнийai-modelЗалежні: 42Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/azure-openai:1.0.0 - v1.0.0
Component
core/ai-model/azure-openai-embedding
Azure OpenAI's managed embeddings API. Its host belongs to the Azure OpenAI resource an organisation runs, and a model is addressed by the deployment name given to it inside that resource, together with the API version a request is made against and, where an Aggregate wants vectors narrower than the model's own default, the output width. It publishes the same endpoint-and-model contract the generative models publish, so an application that embeds text reads exactly the keys it would read to generate it. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model. The helm and compose realizations reference a resource that already exists: no compute runs, the compiler never calls the provider, and the Release supplies the endpoint in infra.endpoint. terraform/azure-openai-embedding creates the resource and the model deployment in the resource group, region and subdomain the Release names in infra.overrides, and publishes the endpoint composed from that subdomain. Reach for a reference where the resource exists outside the Release; reach for terraform where the Release is to create it.
Офіційнийai-modelЗалежні: 42Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/azure-openai-embedding:1.0.0 - v1.0.0
Component
core/ai-model/azure-managed
An Azure managed online endpoint, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. This is the handle for a model an organisation has deployed behind its own scoring endpoint rather than one of a provider's catalogue models, so the host is per-deployment and operator-supplied in the Release. Consumers read that endpoint and the model name over a `uses` edge, the contract every model here publishes, so an application reaching a privately deployed model is wired exactly like one reaching a public API. Authentication is the consumer's concern: a workload uses the identity its platform gives it, or attaches its own credential, rather than receiving a key from this realization.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/azure-managed:1.0.0 - v1.0.0
Component
core/ai-model/anthropic
Anthropic's managed Claude API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. The endpoint defaults to the managed host, so a bare declaration is enough, and it sits on `infra` rather than inside a Recipe so an operator may repoint it at a gateway or proxy without editing the Aggregate. Consumers read the endpoint and the developer-selected model name over a `uses` edge — the contract every model here publishes — so exchanging this component for another provider's changes nothing on the consuming side. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 42Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/anthropic:1.0.0 - v1.0.0
Component
core/ai-model/ai21
AI21 Labs' managed Jamba API, reached as an external reference: no compute runs in-cluster and the compiler never calls the provider. Its host is fixed, so the realization publishes the endpoint itself and a Release supplies none — declaring the component and choosing a model is the whole of what an Aggregate writes. Consumers read that endpoint and the model name over a `uses` edge, the contract every model here publishes, so an agent reaches it with no endpoint in its own code. The API key is the consumer's to attach: the environment-variable name it arrives under belongs to the application rather than to the model.
Офіційнийai-modelЗалежні: 41Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-model/ai21:1.0.0 - v1.0.0
Component
core/ai-agent/vertex-ai
A Vertex AI agent, realized as an external reference (no in-cluster compute). The agent and its instructions live in Google's service and are addressed by the agent identifier the developer supplies. The regional endpoint is operator-supplied, since the Vertex AI host carries the region. Authentication is the consumer's concern — workloads use ambient GCP workload identity or inject their own credentials, rather than a static key delivered by the realization. The optional data-store socket wires the agent to a retrieval-capable datastore when the Aggregate declares one.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/vertex-ai:1.0.0 - v1.0.0
Component
core/ai-agent/vapi
A voice-agent platform reached over Vapi's REST API, so an application holds a spoken conversation with an assistant Vapi hosts and runs. The developer names the assistant and, optionally, the version to pin; the operator supplies the API base address and the platform-held number a call originates from. Wire here rather than to a carrier: a carrier moves the call, an agent platform conducts the conversation, and many Releases want both. Consumers read the base address, the assistant identifier, the version and the number. Realized by reference: no compute runs here.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/vapi:1.0.0 - v1.0.0
Component
core/ai-agent/smolagents
A smolagents agent, run in-cluster as an application image the Release supplies. smolagents offers two ways for an agent to act, and they are different enough to be different classes rather than a flag: one writes and executes code to reach its answer, the other selects among declared tools. The agent-type value carries that choice, and the step ceiling is deliberately small — a handful of steps rather than a hundred — because the library is built around short reasoning loops. The Release supplies the image containing the smolagents application. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/smolagents:1.0.0 - v1.0.0
Component
core/ai-agent/semantic-kernel
A Semantic Kernel agent, run in-cluster as an application image the Release supplies. Semantic Kernel turns a stated goal into a sequence of calls over the functions an application registers, and how that sequence is produced is a choice rather than a default — a templated plan, the model's own function calling, or a plan built a step at a time — which is what this component's planner value selects. Beside it the component exposes the sampling settings those calls run under, so temperature and nucleus sampling are declared in the Aggregate rather than compiled into the image. The catalog supplies no image for it — Semantic Kernel is published as a library an application is built around, so there is no upstream image to pin — and until a Release sets `infra.image` the build halts naming the container image and renders nothing. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, beside the variables the framework and its client read. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/semantic-kernel:1.0.0 - v1.0.0
Component
core/ai-agent/retell
A voice-agent platform reached over Retell AI's REST API, so an application holds a spoken conversation with an agent Retell hosts and runs. The developer names the agent and, optionally, the version to run; the operator supplies the API base address and the platform-held dialable number a call originates from. Wire here rather than to a carrier: a carrier moves the call, an agent platform conducts the conversation, and many Releases want both. Consumers read the base address, the agent identifier, the version and that number. Realized by reference: no compute runs here.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/retell:1.0.0 - v1.0.0
Component
core/ai-agent/rivet
A Rivet agent, run in-cluster as an application image the Release supplies. Rivet is authored visually: chains are drawn as named graphs and saved together inside a project file, and the runtime executes one of those graphs by name. So this component identifies its work by a file-and-graph coordinate rather than by a prompt — the only agent here that does — and passes the graph's declared inputs as a string map. The Release supplies the image carrying both the runtime and the project file. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/rivet:1.0.0 - v1.0.0
Component
core/ai-agent/pydantic-ai
A Pydantic AI agent, run in-cluster as an application image the Release supplies. Pydantic AI makes an agent's answer a typed, validated object rather than free text: the result type is declared, each reply is checked against it, and a reply that fails the check is asked for again — so what this component exposes is that type, the retry budget, and the system prompt the agent works under. Deploying it gives that agent as a worker wired to a model on the `model` socket, reached through the environment an OpenAI-compatible client reads, so one image runs against any model in the catalog. The catalog supplies no image for it — Pydantic AI is published as a library an application is built around, so there is no upstream image to pin — and until a Release sets `infra.image` the build halts naming the container image and renders nothing. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/pydantic-ai:1.0.0 - v1.0.0
Component
core/ai-agent/phidata
A Phidata agent, run in-cluster as an application image the Release supplies. Phidata separates two kinds of remembering that other frameworks fold together: session storage, which holds the running conversation so a later turn can see an earlier one, and a knowledge base, which holds reference material the agent retrieves from. This component declares both as distinct topology edges — the only agent here that does — and its two values govern whether prior turns are replayed into the prompt and whether replies are formatted as markdown. The Release supplies the image containing the Phidata application. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The variables each connection is delivered under are defaults: an `inject:` block on the connection replaces them with the names the application reads, as variables or as files, and a store connection writing none refuses a store served over TLS alone or over HTTPS, since the application would be handed only its host and port. Neither store connection carries a credential, so a store that requires signing in cannot be reached through one. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/phidata:1.0.0 - v1.0.0
Component
core/ai-agent/promptflow
A Prompt flow application, run in-cluster as an application image the Release supplies. Prompt flow models the work as a flow — a named, executable graph of prompt and Python steps kept beside the source — so what an Aggregate selects is not an agent object but which flow to run. That is why this component's whole value surface is the single flow entrypoint, and why it exposes no step, iteration or delegation knob: everything else about the run is inside the flow definition the image carries. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/promptflow:1.0.0 - v1.0.0
Component
core/ai-agent/openclaw
An OpenClaw agent, run in-cluster as an application image the Release supplies. OpenClaw is a long-lived assistant process that receives requests over messaging channels and drives a model on the requester's behalf, and it declares its providers, channels and agent defaults in a JSON configuration file it reads from its own state directory rather than on a command line. So what this component states is the bound a single turn runs under — whether the agent completes the turn by itself or pauses for the requester between steps, how long the turn may take, and how many steps it may spend — and it leaves provider and channel selection to that configuration. The Release supplies the image and the configuration the image reads. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the image reads, as variables or as files, in place of the endpoint and model variables the Recipes choose by default; the attached key stays. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/openclaw:1.0.0 - v1.0.0
Component
core/ai-agent/openhands
An OpenHands agent, run in-cluster as an application image the Release supplies. OpenHands works on a codebase the way a developer does — reading files, editing them, running commands — so a run is a loop over a workspace rather than a single answer, and the iteration ceiling is the one value this component insists on: it is the bound that stops a loop that is not converging. The task itself is optional here, because it may be handed in at run time instead of fixed in the topology. The Release supplies the image containing the OpenHands application. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/openhands:1.0.0 - v1.0.0
Component
core/ai-agent/openai-agents-sdk
An agent built on the OpenAI Agents SDK, run in-cluster as an application image the Release supplies. The SDK is published as a library an application is built around rather than as a server, so there is no upstream image to pin: the catalog supplies none, and until a Release sets `infra.image` the build halts naming the container image and renders nothing. Deploying it gives that application as a worker wired to a model serving the OpenAI API. The Recipe delivers the model's endpoint, its model name and the connection's credential as the variables the SDK reads, so one image runs against any such model in the catalog without an endpoint in its code. By default the SDK uploads a trace of each run to OpenAI, authenticated with the key it reads as OPENAI_API_KEY, which here is the model connection's credential whatever model it reaches. While `values.upload_traces_to_openai` is false, its default, the Recipe delivers OPENAI_AGENTS_DISABLE_TRACING=1, and that upload is off unless the application turns tracing on in its own code. Set true, the variable is not delivered and each run's traces go to OpenAI. Each tool wired into the optional `tools` socket, such as a Model Context Protocol server reached over HTTP, reaches the application under variables named for the connection's alias, its address as AGGREGA_TOOL_<ALIAS>_URL, which the application hands to the SDK in its own code. An `inject:` block on a connection delivers the names the application reads, as variables or files: on a tools connection in place of the tool's address and header, and on the model connection beside the variables the SDK reads. The component's own `inject:` block takes `env` and `file` entries.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/openai-agents-sdk:1.0.0 - v1.0.0
Component
core/ai-agent/open-webui
A self-hosted Open WebUI installation — the browser platform a person holds conversations in, and the tier that turns a bare model endpoint into a working assistant surface. It reaches its model through the `model` uses-socket and calls it as an OpenAI-compatible API, adding retrieval over the documents a user loads, tool invocation over the tool servers an administrator registers, and per-user and per-group access control. The conversation loop is always started by a person: the platform runs no background loop of its own. It keeps its accounts, conversations and document index on one durable volume, so a realization gives it storage that outlives its container. It exports its host, port and protocol, which is how an ingress tier publishes it. On Kubernetes there are two realizations and the Manifest chooses between them. `open-webui` synthesizes the platform as a workload and fetches nothing, so it is the one to reach for where the build must complete with no chart repository reachable. `open-webui-chart` renders the platform's own Helm chart, which additionally carries single sign-on against several identity providers and per-component log levels, both set through `infra.overrides`, so reach for it where a chart repository is reachable at build time.
Офіційнийai-agentЗалежні: 28Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/open-webui:1.0.0 - v1.0.0
Component
core/ai-agent/mastra
An agent built on Mastra, run in-cluster as an application image the Release supplies: the server `mastra build` produces. The catalog pins no image; the build halts naming `infra.image` until a Release sets it. It runs on `infra.port`, 4111 by default, delivered as Mastra's PORT, with MASTRA_HOST set to every interface. It serves routes under `/api`, such as `/api/agents/<id>/generate` and `/api/v1/responses`, and `/health`, read by the Kubernetes readiness check. They answer any caller unless the application configures authentication, so attaching core/public-ingress or core/http-route halts the build until `infra.routes_authenticated` affirms it does. The address is published as host, port, endpoint and protocol. Mastra serves every registered agent over Agent2Agent; where `values.agent_id` names one, agent_card_url publishes its card's address, for core/ai-tool/agent-invocation to delegate to. The `model` socket takes a model or gateway serving the OpenAI API and delivers its endpoint as OPENAI_BASE_URL and the connection's credential, by reference, as OPENAI_API_KEY, which Mastra reads for an `openai/` model the agent's code names. Each tool wired to the optional `tools` socket, such as an HTTP Model Context Protocol server, reaches the application as AGGREGA_TOOL_<ALIAS>_URL, beside any credential its connection carries. An `inject:` block on any connection, or on the component itself, delivers names the application reads, as variables or files, a tools connection's in place of the tool's address and header. A set `values.agent_id` reaches the application as MASTRA_AGENT_ID by default; an entry of the component's own block reading it alone replaces that name.
Офіційнийai-agentЗалежні: 28Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/mastra:1.0.0 - v1.0.0
Component
core/ai-agent/metagpt
A MetaGPT agent, run in-cluster as an application image the Release supplies. MetaGPT organizes work the way a software team does: each agent is given a role and follows the procedure that role implies, driven from a single stated objective rather than a turn-by-turn conversation. The component's surface is that pair — the role the agent occupies and the objective it is working toward — with no delegation flag or per-agent backstory, because the procedure attached to the role supplies both. The Release supplies the image containing the MetaGPT application. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/metagpt:1.0.0 - v1.0.0
Component
core/ai-agent/llamaindex
A LlamaIndex agent, run in-cluster as an application image the Release supplies. LlamaIndex offers several distinct ways for an agent to reach an answer rather than one — reasoning and acting a step at a time, delegating tool selection to the model's own tool-calling, or planning the whole sequence before any of it runs — and that choice is the value this component insists on, alongside the system prompt the agent works under and a ceiling on how many tool calls a single run may make. Deploying it gives that agent as a worker wired to a model on the `model` socket, so one image runs against any model in the catalog. The catalog supplies no image for it — LlamaIndex is published as a library an application is built around, so there is no upstream image to pin — and until a Release sets `infra.image` the build halts naming the container image and renders nothing. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/llamaindex:1.0.0 - v1.0.0
Component
core/ai-agent/livekit-agent-worker
A LiveKit agent worker — the resident process that joins a realtime room as a participant and runs the agent's own code there: the speech pipeline, or the realtime model, that holds the conversation with whoever else is in the room. It registers with the media server and runs one agent session for each job it is handed; it serves no application traffic, so no other component addresses it. Wire here when the conversation inside the cluster's own rooms is to be carried by software it runs and by agent code written for it, rather than handed to a hosted agent platform. Upstream publishes the agent framework as a language package rather than as a container image, and this component pins no image because there is none to pin. The image is a required operational value carrying no default, so using this component means building an image from the framework together with the agent code, publishing it where the cluster can pull it, and naming that reference in the Release. Nothing renders until that reference is supplied. Consumers receive the dispatch name the worker registers under; the worker publishes no host and no port. Realized by a Helm Recipe that synthesizes a Deployment over the supplied image and declares no Service and no volume for state. The connection to the media server needs no `inject:` block; one it writes delivers further names, as variables or files. The worker serves Prometheus metrics only where the agent's own code gives it a metrics port, which no environment variable or setting here reaches, so it publishes no metrics address, and a worker is not a target the catalog's Prometheus scrapes. The component's own `inject:` block takes `env` and `file` entries.
Офіційнийai-agentЗалежні: 20Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/livekit-agent-worker:1.0.0 - v1.0.0
Component
core/ai-agent/lindy
A Lindy assistant on Lindy's managed platform, realized as an external reference (no in-cluster compute). The assistant lives in Lindy's service and is addressed by the identifier the developer supplies, together with the task prompt it is asked to carry out and an optional persona override. The realization emits only the wiring a consuming application needs, which is the platform endpoint and the assistant identifier to address a request to it. The endpoint carries no default and is operator-supplied, so a Release states the host it reaches the platform at; it is sourced from infra so an operator may also repoint it at a gateway or proxy. Authentication is the consumer's concern: the realization delivers no key, because the environment-variable name belongs to the consuming application rather than to the endpoint.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/lindy:1.0.0 - v1.0.0
Component
core/ai-agent/letta
A stateful-agent platform reached over Letta's REST API, so an application sends a message to an agent Letta hosts and runs, and that agent carries its memory forward from one message to the next. The developer names the agent; the operator supplies the API base address. Wire here rather than to a model: a model answers from the context a caller supplies, an agent platform holds that context itself across messages. Consumers read the base address and the agent identifier. Realized by reference: no compute runs here.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/letta:1.0.0 - v1.0.0
Component
core/ai-agent/gumloop
A Gumloop automation pipeline on Gumloop's managed platform, realized as an external reference (no in-cluster compute). The pipeline is built and stored in Gumloop's service and is addressed by the pipeline identifier the developer supplies, alongside the inputs a run starts from. The realization emits only the wiring a consuming application needs, which is the API endpoint and the pipeline identifier to start a run against it. The endpoint defaults to Gumloop's fixed API host and is sourced from infra so an operator may repoint it at a gateway or proxy. Authentication is the consumer's concern: the realization delivers no key, because the environment-variable name belongs to the consuming application rather than to the endpoint.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/gumloop:1.0.0 - v1.0.0
Component
core/ai-agent/haystack
A Haystack agent, run in-cluster as an application image the Release supplies. Haystack builds retrieval-augmented pipelines declaratively — the pipeline is a YAML document carried inside the application image — so what this component exposes is the path to that document plus the ceiling on how many times the agent may loop before it stops. Alongside the model socket it declares a document-store socket admitting vector, semantic-memory and search engines, so the store a pipeline retrieves from is chosen in the Aggregate rather than fixed in code, and changing it rebuilds no image. The catalog supplies no image for it — Haystack is published as a library an application is built around, so there is no upstream image to pin — and until a Release sets `infra.image` the build halts naming the container image and renders nothing. The document-store connection is delivered to the application only through an `inject:` block on it, under the names the application reads, as variables or as files; on the model connection such a block replaces the default variables. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/haystack:1.0.0 - v1.0.0
Component
core/ai-agent/griptape
A Griptape agent, run in-cluster as an application image the Release supplies. Griptape's distinguishing move is to keep large tool output out of the prompt: a tool's result is parked in task memory and referenced, so the agent reasons over a handle instead of the payload. That is what the off-prompt switch selects, and it is why this component wires a task-memory store as a topology edge alongside the model. The Release supplies the image containing the Griptape application. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The variables each connection is delivered under are defaults: an `inject:` block on the connection replaces them with the names the application reads, as variables or as files, and a task-memory connection writing none refuses a store served over TLS alone or over HTTPS, since the application would be handed only its host and port. The task-memory connection carries no credential, so a store that requires signing in cannot be reached through it. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/griptape:1.0.0 - v1.0.0
Component
core/ai-agent/gemini
A Gemini assistant hosted on Google's generative-language API, realized as an external reference (no in-cluster compute). The model runs in Google's service and the deployment holds a contract and a key; the assistant's behaviour is carried in the system instruction and the per-request options the application sends, including whether search grounding is enabled. The realization emits only the wiring a consuming application needs, which is the API endpoint and the model string to call it with. The endpoint defaults to the fixed generative-language host and is sourced from infra so an operator may repoint it at a gateway or proxy.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/gemini:1.0.0 - v1.0.0
Component
core/ai-agent/dynamiq
A Dynamiq agent, run in-cluster as an application image the Release supplies. Dynamiq composes model calls, retrieval and agent loops into a workflow of typed nodes, and the workflow is Python code the host program constructs — so the unit of work here is a named entrypoint inside that program rather than a prompt, and the Aggregate chooses whether the entrypoint streams each intermediate node result or returns only the final one. Conversation and workflow memory is held by a pluggable backend rather than in the process, which is why this component wires a datastore as a topology edge alongside the model. The Release supplies the image carrying both the framework and the workflow it runs. The catalog pins no image, so nothing renders until the Release sets `infra.image`. The variables each connection is delivered under are defaults: an `inject:` block on the connection replaces them with the names the application reads, as variables or as files, and a state-store connection writing none refuses a store served over TLS alone or over HTTPS, since the application would be handed only its host and port. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/dynamiq:1.0.0 - v1.0.0
Component
core/ai-agent/dify-ssrf-proxy
The egress proxy a Dify installation routes user-directed outbound requests through. A Dify workflow fetches whatever address its author or its end user supplies — a web page to crawl, a file to load, an HTTP request node's target — and a service that made those calls itself would be reachable proof of every address the cluster can see. This component is the chokepoint that removes that: it is a Squid forward proxy configured to refuse any destination in private, loopback, link-local or multicast address space, so a supplied address can reach the public internet and cannot reach the cluster's own services or the node's metadata endpoint. Its allowlist of private destinations is empty by default and is widened only by naming addresses explicitly. Holds no state — its cache is in memory and discarded with the process — so it renders as a stateless server. Exports its host, port and protocol to consumers via the uses topology, which is how each caller learns the proxy address to send through.
Офіційнийai-agentЗалежні: 28Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/dify-ssrf-proxy:1.0.0 - v1.0.0
Component
core/ai-agent/dify-sandbox
The Dify code sandbox — the isolated execution service that runs the Python and JavaScript of a workflow's code nodes. The API server and the task worker hand it a snippet over HTTP and receive the result; the snippet never runs inside either of them, which is what keeps a workflow author's code out of the platform's own process. It holds an authentication key that its callers present, so a workload that merely reaches its port cannot execute code through it, and its own outbound traffic leaves through the forward proxy wired on the `ssrf_proxy` uses-socket rather than going out directly. Keeps no state between calls — each execution is independent and its working directory is discarded — so it renders as a stateless server. Exports its host, port and protocol to consumers via the uses topology.
Офіційнийai-agentЗалежні: 27Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/dify-sandbox:1.0.0 - v1.0.0
Component
core/ai-agent/dify-plugin-daemon
The Dify plugin runtime — the service that installs and runs the plugins a Dify installation's model providers and tools are delivered as. Every model call and every tool invocation an application makes is dispatched through it, so an installation with no plugin runtime can edit a workflow but cannot run one. It keeps its own plugin catalogue and installation records in a dedicated database on the PostgreSQL backend wired through the `backend` uses-socket, coordinates through the Redis-compatible cache wired through `cache`, and calls back into the API server over the internal API on the `api` socket to resolve workspace and credential context. Installed plugin packages and their prepared runtime environments live on its own volume, and it verifies each package's signature before running it, so it renders as a single-replica stateful server with a persistent claim. Exports its host, port and protocol to consumers via the uses topology. It serves its own metrics in the Prometheus format at /metrics/ on that port, on by default and asking no key, and its Recipe publishes that address as `metrics_endpoint`.
Офіційнийai-agentЗалежні: 28Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/dify-plugin-daemon:1.0.0 - v1.0.0
Component
core/ai-agent/bland
A voice-agent platform reached over Bland's REST API, so an application places a spoken call that a conversational pathway Bland hosts and runs conducts. The developer names the pathway and, optionally, the version to run instead of the one the platform marks as production; the operator supplies the API base address. Wire here rather than to a carrier: a carrier moves the call, an agent platform conducts the conversation, and many Releases want both. Consumers read the base address, the pathway identifier and the version. Realized by reference: no compute runs here.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/bland:1.0.0 - v1.0.0
Component
core/ai-agent/claude-agent-sdk
An agent built on the Claude Agent SDK, run in-cluster as an application image the Release supplies. The SDK runs each session in the `claude` command line it bundles, as a process inside that image, so there is no upstream image to pin: the catalog supplies none, and until a Release sets `infra.image` the build halts naming the container image and renders nothing. Deploying it gives that application as a worker wired to a model serving Anthropic's Messages API: the `model` socket takes a model or a gateway whose `model.aggrega.ai/api` is `anthropic`, as `core/ai-model/anthropic` is, and refuses any other. The Recipe delivers the model's endpoint, its model name and the connection's credential as the variables the command line reads. Each tool wired into the optional `tools` socket, such as a Model Context Protocol server reached over HTTP, reaches the application under variables named for the connection's alias: the tool's address as AGGREGA_TOOL_<ALIAS>_URL, and the connection's credential, by reference, as AGGREGA_TOOL_<ALIAS>_CREDENTIAL beside the header to present it under. The SDK reads no tool's address from a variable of its own, so the application names each one in the servers it hands the SDK. An `inject:` block on a connection, or on the component itself, delivers the names the application reads, as variables or files: in place of a tool's address and header, and beside the model's variables. The SDK's Python package is published under the MIT licence. The command line it bundles, in which every session runs, is distributed by Anthropic under its own commercial terms, so the component's `component.aggrega.ai/upstream-licensing` label is `proprietary`.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/claude-agent-sdk:1.0.0 - v1.0.0
Component
core/ai-agent/bedrock
An Amazon Bedrock agent, realized as an external reference (no in-cluster compute). The agent, its instruction and its action groups live in the Bedrock service and are addressed by the agent name the developer supplies. The regional runtime endpoint is operator-supplied, since the Bedrock agent-runtime host carries the region. Authentication is the consumer's concern — workloads use ambient AWS IAM (for example EKS IRSA) or inject their own credentials, rather than a static key delivered by the realization. The optional knowledge-base socket wires the agent to a retrieval-capable datastore when the Aggregate declares one.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/bedrock:1.0.0 - v1.0.0
Component
core/ai-agent/autogpt
An AutoGPT agent, run in-cluster as an application image the Release supplies. AutoGPT pursues a list of goals on its own initiative rather than answering a single request, so this component's surface is that goal list plus the identity the agent adopts while working through it, and a bound on how many steps it may take without being asked again. The Release supplies the image containing the AutoGPT application; the model is wired as a topology edge rather than configured in code. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/autogpt:1.0.0 - v1.0.0
Component
core/ai-agent/anthropic
A Claude assistant hosted on Anthropic's managed API, realized as an external reference (no in-cluster compute). The model runs in Anthropic's service and the deployment holds a contract and a key; the assistant's behaviour is carried in the system prompt and the per-request options the application sends. The realization emits only the wiring a consuming application needs, which is the API endpoint and the model string to call it with. The endpoint defaults to Anthropic's fixed host and is sourced from infra so an operator may repoint it at a gateway or proxy.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/anthropic:1.0.0 - v1.0.0
Component
core/ai-agent/autogen
An AutoGen agent, run in-cluster as an application image the Release supplies. AutoGen composes agents that converse with one another to reach a result, and each agent is constructed with a name and a system message that fixes its part in that conversation — which is what this component exposes. The Release supplies the image containing the AutoGen application; the model the agents talk to is wired as a topology edge rather than configured in code, so the same image runs against any model in the catalog. The catalog pins no image, so nothing renders until the Release sets `infra.image`. An `inject:` block on the model connection delivers the names the application reads, as variables or as files, in place of the model-name variable the Recipes choose by default; the endpoint variable the client SDK reads and the attached key stay. Its Recipes deliver each own value under a default variable name; the component's own `inject:` block takes `env` and `file` entries, and one reading a value alone replaces that value's default name.
Офіційнийai-agentЗалежні: 21Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/autogen:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-worker-streams
The stream-consuming worker tier of an agent workspace — the process that drains the platform's continuous streams and writes what they carry into the relational backend. Three families of stream reach it: the spans an instrumented application or an agent run emits, the platform's own internal events, and the transcript records of an agent session. It reads its work from the stream side of the queue wired through the `queue` socket, writes to the relational backend wired through `backend` using the two database names given on that edge, and uses the cache wired through `cache` for the same read-caching the rest of the platform does. Its `streams` value selects which families one container runs; left unset it runs all three, which is what a single-container installation wants. An Aggregate that needs one family scaled apart from the others declares a second instance of this same component with a value naming just that family — the stream names and consumer groups are unchanged by the split, so two instances share the work rather than duplicating it. Wire here because trace and session ingestion happens off the request path. With nothing running this tier the platform accepts spans and session records into the streams and never drains them, so the console's traces stay empty while nothing reports an error — which is why this is a component of its own and not an optional extra. It serves no port and nothing connects to it, so it publishes no provides contract. Realized on Kubernetes via its Helm Recipe, which synthesizes the tier as a stateless Deployment with no Service, running the platform image under the stream-consumer entry point.
Офіційнийai-agentЗалежні: 20Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-worker-streams:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-worker-queues
The queue-consuming worker tier of an agent workspace — the process that runs the platform's discrete background jobs. Four families of job reach it: the outbound webhooks an installation has configured, the triggers that start a background agent when something happens in a connected application, the asynchronous turns of an agent session, and the evaluation runs a team asks for over a set of test cases. It takes work from the queue wired through the `queue` socket, reads and writes the platform's model in the relational backend wired through `backend` using the two database names given on that edge, uses the cache wired through `cache`, and reaches the model-facing service wired through `services` because a job that runs an agent turn or scores an evaluation needs a model call made. Its `queues` value selects which families one container runs; left unset it runs all four, and an Aggregate that needs one family scaled apart declares a second instance naming just that family. Wire here because background work is what makes an agent workspace more than a chat window: a background agent that runs on a schedule or on an event, and an evaluation over a test set, both exist only on this tier. With nothing running it, such a request is accepted and never performed. It serves no port and nothing connects to it, so it publishes no provides contract. Realized on Kubernetes via its Helm Recipe, which synthesizes the tier as a stateless Deployment with no Service, running the platform image under the queue-consumer entry point.
Офіційнийai-agentЗалежні: 19Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-worker-queues:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-web
The browser console of an agent workspace — the front end a team builds its agents on. It is where an agent is described in chat, its tools and skills are connected, its permissions are set, its version history is compared, and the trace of a finished run is read. It is a server-rendered front end and holds no state. Each page it renders is assembled from the control-plane API wired through the `api` socket and the model-facing service wired through `services`, and the two are separate sockets because they are separate services with separate addresses: agent definitions and traces come from the first, a playground completion from the second. Sign-in runs against the authentication core wired through `auth_core`, which the front end addresses by that core's own base URL. Wire here for the human surface. It carries no data of its own, so it is the one tier of the platform that can be scaled or restarted without consequence, and an Aggregate that only needs the API — automation driving the platform from outside — can leave it out. Exports its host, port, endpoint and protocol to consumers via the uses topology, which is what the ingress tier routes the browser path to. Realized on Kubernetes via its Helm Recipe, which synthesizes the front end as a stateless Deployment behind a ClusterIP Service on its HTTP port. Two kinds of address are emitted separately and deliberately: the in-cluster Service names it calls while rendering a page, and the external base URL the browser itself calls, which the operator supplies because a Service name does not resolve outside the cluster.
Офіційнийai-agentЗалежні: 26Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-web:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-web-mobile
The small-screen front end of an agent workspace — a separate console built for a phone. It serves the part of the workspace that reads and works on a small screen: talking to an agent in chat, reading what a background run did, and approving or refusing an action an agent is waiting on. It is its own service with its own container image, not a responsive mode of the full console, and it holds no state. Each page is assembled from the control-plane API wired through the `api` socket and the model-facing service wired through `services`, and sign-in runs against the authentication core wired through `auth_core` — the same three sockets the full console declares, wired to the same three targets, because the two front ends read one platform. Wire here when people work with their agents away from a desk. It is genuinely optional: an Aggregate that leaves it out loses the small-screen surface and nothing else, and the full console's own mobile-redirect value is what decides whether a small-screen visitor is sent here, so leaving that value off is how an installation without this component stays coherent. Exports its host, port, endpoint and protocol to consumers via the uses topology, which is what the ingress tier routes its own path to — a path distinct from the full console's, since both are reached on one hostname. Realized on Kubernetes via its Helm Recipe, which synthesizes the front end as a stateless Deployment behind a ClusterIP Service on its HTTP port, with the same split between in-cluster and browser-facing addresses the full console's Recipe makes.
Офіційнийai-agentЗалежні: 26Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-web-mobile:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-runner
The agent runner of an agent workspace: the service that actually executes an agent. It takes a run request from the model-facing service, starts the harness the agent is configured for as a child process inside a sandbox, mounts that agent's working directory into the sandbox, and relays the tool calls the harness makes back to the platform through the api socket. Its environment is deliberately narrow, and that is a security property rather than an omission: a locally sandboxed harness can read this container's environment, so the component declares no relational, cache, queue or object-store socket and holds no platform credential beyond the token it requires on every call. An Aggregate that adds such a value here removes that property. Realized on Kubernetes via its Helm Recipe, which synthesizes the runner as a stateless Deployment behind a ClusterIP Service on its HTTP port, with the local sandbox provider enabled and no remote provider configured. It exports its host, port, endpoint and protocol to consumers.
Офіційнийai-agentЗалежні: 26Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-runner:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-services
The model-facing service of an agent workspace: the tier that talks to language-model providers on the platform's behalf. It serves the completion and chat endpoints the console's playground calls, holds the adapter layer that presents one interface over many providers, and starts an agent run by handing it to the runner. It reads the platform's relational model through backend, uses cache and queue as the control-plane API does, keeps the files an agent produces in the object store wired through store, and reaches auth_core to establish who a request belongs to and runner to start, watch and cancel a run. Wire here to invoke a model or start an agent run. That is the line between this service and core/ai-agent/agenta-api, which owns the platform's own data: a caller listing agents or reading a trace talks to that one, and a caller asking for a completion or a run talks to this one. Both are needed, and neither is a superset of the other. Realized on Kubernetes via its Helm Recipe, which synthesizes the service as a stateless Deployment behind a ClusterIP Service on its HTTP port. It exports its host, port, endpoint and protocol to consumers.
Офіційнийai-agentЗалежні: 26Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-services:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-migrations
The schema-migration step of an agent workspace — the run that creates and updates the platform's tables. It applies the migration history the platform ships to both of its databases: the one holding the platform's own model, and the one holding traces. It runs to completion and exits, which is what makes it a different operational shape from every other tier of this platform. It reaches the relational backend through the `backend` socket, using the two database names given on that edge, and it declares nothing else: the run reads no cache, takes no queue work, writes no files, and authenticates no user, so giving it a cache, queue, object-store or authentication socket would be declaring a dependency it does not have. Wire this in because the platform's services do not create their own tables. Without this run the databases stay empty and every service starts, connects, and then fails on its first query — a failure that reads as a broken installation rather than as a missing step, which is why the step is a declared component rather than something a reader is expected to remember. It serves no port and nothing connects to it, so it publishes no provides contract. Because it exits, an Aggregate sequences it ahead of the services that need the schema; it is written to be safe to run again, so re-running it on an already-migrated database applies nothing and succeeds. Realized on Kubernetes via its Helm Recipe, which synthesizes the step as a run-to-completion workload rather than a Deployment, running the platform image under its migration entry point.
Офіційнийai-agentЗалежні: 18Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-migrations:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-cron
The scheduler of an agent workspace — the process that decides when a time-triggered thing is due and puts it on the queue. A background agent set to run every morning, and the platform's own periodic housekeeping, both become work because this tier woke up and enqueued them. It is a resident process carrying its own schedule table rather than a platform-level scheduled object, which is how the platform's own published manifests run it: one container that stays up and fires on its own clock. It reads the schedules from the relational backend wired through `backend`, using the two database names given on that edge, enqueues onto the queue wired through `queue`, and uses the cache wired through `cache` for the same read-caching the rest of the platform does. It performs none of the work itself — the queue-consuming worker tier does — so this tier stays small however much is scheduled. Wire here because a schedule with nothing evaluating it never fires. That failure is silent from every other tier: the agent exists, its schedule is stored, and nothing happens at the appointed time. Exactly one instance of this component belongs in an Aggregate. It is not horizontally scalable — two schedulers reading one schedule table each enqueue the same due item — so scale is a question for the worker tier rather than this one. It serves no port and nothing connects to it, so it publishes no provides contract. Realized on Kubernetes via its Helm Recipe, which synthesizes the tier as a stateless single-replica Deployment with no Service, running the platform image under its scheduler entry point.
Офіційнийai-agentЗалежні: 20Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-cron:1.0.0 - v1.0.0
Component
core/ai-agent/agenta-api
The control-plane API of an agent workspace: the service that holds the agents a team builds, each agent's instructions, tools, skills, permissions and version history, and the record of every run. It serves the REST API the console renders from and an external client calls, and it is the tier that reads and writes the platform's relational model. It holds no state of its own: a relational backend, a cache, a queue and an object store do, each wired through a socket of its own, and it reaches the authentication core wired through auth_core for every sign-in and every session check. Wire here for the platform's own data plane. It is the counterpart of core/ai-agent/agenta-services, which owns completions, chat and the dispatch of an agent run to the runner, so a caller that wants to invoke a model reaches that service rather than this one. Realized on Kubernetes via its Helm Recipe, which synthesizes the service as a stateless Deployment behind a ClusterIP Service on its HTTP port. It exports its host, port, endpoint and protocol to consumers.
Офіційнийai-agentЗалежні: 26Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/agenta-api:1.0.0 - v1.0.0
ArtifactLibrary
TempoConfig
The configuration of a Grafana Tempo server run in its single-binary shape, as one bundle every Recipe synthesizing core/observability/tempo from the grafana/tempo image imports, so the server is configured alike whichever target it runs on. ServerConfig writes the server's whole configuration: the HTTP query API on port 3200, the OTLP receiver for gRPC on 4317 and for HTTP on 4318, the compactor's block retention from the developer's retention_period, and a local trace store with its write-ahead log under /var/tempo. The importing Recipe picks the delivery with the local server_config_delivery — config_map, one ConfigMap holding the configuration under the key tempo.yaml, for a Kubernetes target, or compose_configs, one compose config — and declares as a local the name it shares with the bundle, which stops the build when it is missing or differs.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/tempoconfig:1.0.0 - v1.0.0
Component
core/ai-agent/a2a-remote
An agent another party runs and publishes over the Agent2Agent (A2A) protocol, reached at the address its Agent Card lists. Nothing is provisioned or run. Both realizations are references publishing what a consumer needs: the agent's address, where its card is read, the URL scheme, the header a caller's credential goes under, and where a caller obtains a token. The Release names the address in infra.endpoint, which carries no default: the URL of an interface the card lists for a binding served over HTTP, without any part carrying a key. The card is read at the well-known location on that address's scheme and host, and the header is Authorization, unless infra.agent_card_url and infra.credential_header name others; infra.token_url, infra.oidc_discovery_url and infra.scopes, each optional, copy what the card declares for OAuth 2.0 client credentials or OpenID Connect. No credential is held here: each consumer holds its own on its connection, as it belongs to the consumer, not the address. core/ai-tool/agent-invocation's container presents under that header the whole value a Secret holds, for an API key in a header or cookie or an HTTP scheme such as Bearer or Basic; a token it obtains with a client secret by the OAuth 2.0 client-credentials grant, directly or through OpenID Connect; and, for mutual TLS, a client certificate mounted from a Secret its connection names, the card naming none to publish here. That tool halts the build on a credential or client secret for an http endpoint. Not covered: a key in the query string, an HTTP scheme answering a challenge, such as Digest, and other OAuth 2.0 grants. Reach for helm/a2a-remote on Kubernetes, compose/a2a-remote under Docker Compose.
Офіційнийai-agentЗалежні: 17Опубліковано 2026-11-10oci://hub.aggrega.ai/component/core/ai-agent/a2a-remote:1.0.0 - v1.0.0
ArtifactLibrary
TargetProtocol
A check that every connection on one of a component's uses-sockets reaches a target speaking the protocol the component's client is written for, as one bundle the Recipes of such a component import. Check reads the protocol each connection's target publishes and stops the build on any other, or on none, with a sentence naming the component, the connection, the target and what it publishes, saying why the component needs the protocol, and naming a catalog component that publishes it. It writes no file, so a build whose connections are all right, or with nothing wired, renders what it would without the bundle. The importing Recipe names, as its locals, the socket (check_socket), the protocol (check_protocol, or a JSON list of the protocols its client speaks), why its client needs it (check_reason) and a component publishing it (check_example); the bundle stops the build naming any of the four that is missing.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/targetprotocol:1.0.0 - v1.0.0
ArtifactLibrary
OtelCollectorConfig
The configuration of an OpenTelemetry Collector, as one bundle every Recipe running core/observability/otel-collector imports, so the receivers, processors, exporters and pipelines the collector runs are written once, whichever target it runs on. CollectorConfig writes the collector's whole configuration: the OTLP receiver on the gRPC and HTTP ports the importing Component declares as infra.otlp_grpc_port and infra.otlp_http_port, the health check on port 13133, and one pipeline each for traces, metrics and logs, fed by the OTLP receiver through a memory limiter and a batch processor. Each backend wired to the signal's socket adds one exporter to that pipeline, written for the protocol its observability.aggrega.ai label names: otlp_grpc to its host and port, or otlp_http to its otlp_http_endpoint. A signal with nothing wired ends in the debug exporter at basic verbosity. It stops the build when two listeners share a port, or when a backend publishes no address. With the Component's values.metrics_enabled true it also has the collector serve its own telemetry metrics on every interface, port 8888. The importing Recipe picks the delivery with the local collector_config_delivery — config_map, one ConfigMap holding the configuration under the key relay, for a Kubernetes target, or compose_configs, one compose config — and declares as locals the names and the ports it shares with the bundle, which stops the build when one is missing or differs. The bundle is authored against the collector's contrib distribution at 0.162.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/otelcollectorconfig:1.0.0 - v1.0.0
ArtifactLibrary
LiteLLMGatewayConfig
The configuration a LiteLLM gateway runs on, as one bundle every Recipe realizing core/gateway/litellm imports, so the routing table and the code several Features rely on live here once rather than as a copy in each Recipe. GatewayConfig is one ConfigMap, named after the gateway's rendered name followed by -config, or on Docker Compose one compose config per file. Its first file is the gateway's configuration: the model map, from the hand-typed entries of values.models and the models wired to the models socket, the registration of the request guard the LiteLLMGuard bundle delivers, and the settings values.metrics_enabled and values.log_masking turn on. Beside it, as the attachments and values ask, it carries the caller-identity hook core/context-propagation relies on, the masking module core/data-masking's guardrails are loaded from, and the log-masking module values.log_masking relies on. While core/prompt-injection-shield is attached every map entry is sent to the guard on the guard socket, and while core/concurrency-limit is attached every entry carries its cap. A Feature's realization writes the rest into the configuration through the gateway's pipeline step. The importing Recipe consumes the bundle under the key litellm-gateway-config, the directory its one file, configmap.yaml, is rendered in, and declares as locals the file, module, class and variable names the configuration and its workload share; the bundle stops the build naming any local it reads that the Recipe does not declare. The modules rest on the proxy's internals, so each importing Recipe states the LiteLLM releases they are proven at and halts on any other.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/litellmgatewayconfig:1.0.0 - v1.0.0
ArtifactLibrary
LiteLLMGuard
The request guard and the path allowlist a LiteLLM proxy runs behind, as one bundle every Recipe running that proxy imports, so the text that stops a caller changing where the proxy sends a request, or with which credential, lives here once. RequestGuard carries two texts. A module the proxy loads as a callback forwards only the request fields the proxy declares for the route, the keys the proxy adds on it and any the Release names in infra.admitted_request_fields, and refuses every other field with an error naming it; the fields that would re-target or re-key a call are refused even where the proxy declares them, and a Release admitting one stops the build. An nginx configuration on infra.port forwards only the routes the module checks, the model listing, the health routes, key management and, where the importing Recipe turns them on, the organization routes, and answers every other path 403. With the Component's values.metrics_enabled true it also listens on port 9090, forwarding a GET of /metrics alone to the proxy's own metrics listener on its loopback address, port 9091, with no caller check. The importing Recipe picks the delivery with the local request_guard_delivery — config_map, one ConfigMap for a Kubernetes target, or compose_configs, two compose configs — and declares as locals the names, the address and the ports it shares with the bundle, which stops the build when one is missing or differs. A Feature's realization can make the module refuse more by writing a policy into the proxy's callback settings; a policy only tightens. The bundle is authored against LiteLLM v1.103.1.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/litellmguard:1.0.0 - v1.0.0
ArtifactLibrary
LangflowFlow
The flow a Langflow platform loads at each start, as one bundle every Recipe running core/ai-agent/langflow imports, so the flow, and the addresses its tool connections write into it, are written once, whichever target the platform runs on. FlowFile writes nothing unless the Component's values.flow_definition is set. With no tools connection wired it carries the definition as written. With tools connections wired, each MCP Tools node whose server is named for a connection's alias gets the address the tool publishes as the url of its server configuration and, where the connection carries a credential, the name of the variable holding it under the header the tool publishes: AGGREGA_TOOL_<ALIAS>_AUTHORIZATION, holding Bearer and the credential, under Authorization, and AGGREGA_TOOL_<ALIAS>_CREDENTIAL under any other. The flow is then written back as JSON with its keys sorted. The importing Recipe picks the delivery with the local flow_file_delivery — config_map, a ConfigMap holding the flow as flow.json, for a Kubernetes target, or compose_config, a compose config with every $ doubled — and declares as a local the name it mounts the flow by, which stops the build when it is missing or differs. The bundle is authored against langflow-base 0.11.0.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/langflowflow:1.0.0 - v1.0.0
ArtifactLibrary
KongStatusGuard
The request allowlist a Kong gateway's Status API answers behind, as one bundle every Recipe running that gateway imports, so the line between what the probes and a scraper may ask and what nobody else may lives here once. StatusGuard writes a fragment of nginx configuration that Kong includes in its Status API's own server block. A request from the gateway's loopback address passes as it is. From any other address only a GET of /status, /status/ready or /metrics passes; every other request, POST /status/ready among them, is answered 403 and Kong never sees it. The importing Recipe picks the delivery with the local status_guard_delivery — config_map, one ConfigMap for a Kubernetes target, or compose_configs, one compose config, written only while the importing Component's values.metrics_enabled is true — and declares as locals the names it shares with the bundle, which stops the build when one is missing or differs. The bundle is authored against Kong 3.9.3.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/kongstatusguard:1.0.0 - v1.0.0
ArtifactLibrary
KongModelRoutes
The routes a Kong gateway serves for the models wired to it, written as Kong's own declarative configuration, as one bundle every Recipe running that gateway imports, so the configuration and the rules it is written by live here once. ModelRoutes writes, for each connection on the importing Component's models socket, a service named model-<alias> addressed at the endpoint the model publishes, with a chat route at /<alias>/v1/chat/completions and a completion route at /<alias>/v1/completions, each through Kong's request transformer and AI proxy. A connection carrying a credential has its routes present it as a bearer token read through Kong's environment vault, which names the variable and never the value. With nothing wired to the models socket it writes nothing. The importing Recipe picks the delivery with the local model_routes_delivery — config_map, one ConfigMap for a Kubernetes target, or compose_configs, one compose config — and declares as locals the names and the variable parts it shares with the bundle, which stops the build when one is missing or differs. The bundle is authored against Kong 3.9.3. On the open-source edition it writes Kong's AI Prompt Guard on every route while the importing component carries core/prompt-deny-patterns. A Recipe running Kong Gateway Enterprise declares model_routes_edition: enterprise, and the bundle then writes what the importing component's core/fallback-chain, core/period-budget and core/data-masking attachments ask for: AI Proxy Advanced on a chained alias's routes, AI Rate Limiting Advanced for the whole gateway or on each wired model's service, and the AI PII Sanitizer on each wired model's routes.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/kongmodelroutes:1.0.0 - v1.0.0
ArtifactLibrary
FlowiseFlowSeed
The chatflow a Flowise platform is seeded with and the script that seeds it, as one bundle every Recipe running core/ai-agent/flowise imports, so the steps that create the owner account or sign in as it, then update or import the chatflow, are written once, whichever target the platform runs on. SeedScript writes nothing unless the Component's values.flow_definition is set. Then it carries the definition verbatim, as flow.json, beside a Node.js script, seed.js. The script waits for the platform to answer, registers infra.owner_email as the owner on a fresh instance, signs in, and updates the chatflow at values.chatflow_id or imports the definition at that id, halting with the platform's own status and message when a call is refused. It reads the platform's address, the owner's email and password and the chatflow's id and name from the environment the importing Recipe's step gives it, and the definition from the directory that step mounts the two files in. The importing Recipe picks the delivery with the local flow_seed_delivery — config_map, one ConfigMap holding both files, for a Kubernetes target, or compose_configs, two compose configs, with every $ of the definition doubled — and declares as locals the names and the directory it shares with the bundle, which stops the build when one is missing or differs. The bundle is authored against Flowise 3.1.3.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/flowiseflowseed:1.0.0 - v1.0.0
ArtifactLibrary
ControllerRBAC
Permission bundles for a recipe that installs a controller: the identity the controller runs as, and the grants that let it watch and reconcile the resources it owns. A recipe consumes a bundle with `artifacts: { <key>: { $import: ControllerRBAC#/artifacts/<Bundle> } }`; the version is supplied by the per-Release lockfile and no registry host enters the reference. One import expands into one output file per file in the bundle, each nested under the key it was imported at as `<key>/<filename>`, so objects that are only meaningful together arrive together. Each bundle reads two keys the consuming component declares on its own `infra` surface: `namespace`, the namespace the controller is installed into, and `controller_rules`, the permission list to grant. Carrying the rules as data rather than as text is what lets one bundle serve controllers that watch entirely different resources — the wiring between an identity, a role and a binding is the part every controller shares, and the permission list is the part none of them share. The two bundles compose. A controller confined to one namespace imports NamespacedController alone. A controller that also needs a cluster-scoped grant imports both, at two keys: NamespacedController declares the identity and ClusterScopedBinding grants to that same identity rather than declaring a second one.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/artifactlibrary/controllerrbac:1.0.0 - v1.0.0
Aggregate
core/thanos
A self-contained Thanos metrics read tier over an object-storage backend, packaged as a reusable module: an S3-compatible object store, a Thanos Store Gateway that reads metric blocks from it, and a Thanos Query front end that fans PromQL requests out over the gateway's StoreAPI. The module exports the query front end as `query`, the PromQL endpoint consumers connect to. This module is the read/query tier only: the object store creates the `thanos` bucket, and a writer (a Thanos Receive tier or a Prometheus sidecar) must feed metric blocks into it — a runtime prerequisite, not a module workload. The Store Gateway authenticates to the object store with the store's own key pair, two module parameters an importing Aggregate passes down, `s3_access_key` and `s3_secret_key`; their pinned defaults are development values, for evaluation only. The module attaches `core/referenced-secret` on the object store's two key slots and on the `credential_id` and `credential` the gateway's connection to it carries, so the Manifest of every Release whose Aggregate imports the module maps `core/referenced-secret`, or the build halts.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/thanos:1.0.0 - v1.0.0
Aggregate
core/keycloak
A self-contained identity-server stack: a database-backed identity and access-management server in front of a PostgreSQL backend, packaged as a reusable module. The server provisions and migrates its own schema on startup. The module exports the server as `http`, the identity endpoint consumers connect to. Two credentials are module parameters an importing Aggregate passes down: `db_password`, the one password the server and the backend share, and `admin_password`, the password of the server's bootstrap administrator account. Their pinned defaults are development credentials, for evaluation only. The module attaches `core/referenced-secret` on each credential slot, the server's database password on its connection to the backend, so the Manifest of every Release whose Aggregate imports the module maps `core/referenced-secret`, or the build halts. Because the module fills the backend's credential slot, the backend is realized by a Recipe that reads that slot. `postgres-cloudnative-pg`, whose controller owns the database's credentials itself, halts the build when the slot is filled, so a Manifest importing this module maps the backend to another PostgreSQL realization.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/keycloak:1.0.0 - v1.0.0
Aggregate
core/livekit
A self-hosted realtime voice and video stack packaged as a reusable module: a LiveKit media server, a SIP bridge carrying telephone calls into and out of a room, an agent worker joining a room as a participant, and a Valkey key-value store the server and the bridge share. The module exports the media server as `server`, the endpoint clients open their sessions on, and the bridge as `telephony`, the endpoint a SIP trunk reaches. The agent tier needs one input the module cannot supply. Upstream publishes the agent framework as a language package rather than as a container image, so the catalog pins no image for it, and the Release names an image built from that framework together with its own agent code. Nothing renders until it does. The API secret the three LiveKit components share is `api_secret`, a module parameter an importing Aggregate passes down; it has no default, so the build halts until one is passed. The module attaches `core/referenced-secret` on the media server's credential slot and on the credential each connection to it carries, so the Manifest of every Release whose Aggregate imports the module maps `core/referenced-secret`, or the build halts. Each component publishes an in-cluster address only. The module ships no ingress, no TLS termination, no SIP trunk and no carrier account, which a Release supplies around it, and defines no rooms, dispatch rules or trunk configuration, which are settled at runtime through the media server's own APIs. A UDP media range carries traffic only where the pod uses the node's own network.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/livekit:1.0.0 - v1.0.0
Aggregate
core/harbor
A Harbor container registry packaged as a reusable module. It composes one component, `core/devtool/harbor`, which describes the registry as a whole — its control plane, blob store, console, worker tier, scanner, relational backend, cache and the reverse proxy fronting them — so the module declares no edges of its own and needs nothing declared beside it. One in-cluster address serves container clients and the console, exported as both `registry` and `ui`. A Release sizes the component at `<module-instance>.registry`; its schema defaults cover every `infra` value, so a Release need set none of them. Four credentials are module parameters an importing Aggregate passes down. Three are keys with no default — `csrf_key` (exactly 32 characters), `jobservice_secret` and `registry_http_secret` — so an importer passes each down and its Release fulfills it, or the build halts naming the parameter; `admin_password` has a development default, for evaluation only. The module attaches `core/referenced-secret` on the component's credential slot for each, so the Manifest of every Release whose Aggregate imports the module maps `core/referenced-secret`, or the build halts. The token-signing key, `token_key`, the control plane's shared secret, `core_secret`, and the key stored credentials are encrypted with, `secret_key`, have no default and no module parameter: a Release attaches each on `<module-instance>.registry`, naming a Secret it holds, and its Manifest binds them to `k8s-existing-secret`, or the build halts naming the slot. The registry's remaining passwords take development values stated in its Recipe and published in this catalog: a Release serving real traffic replaces them.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/harbor:1.0.0 - v1.0.0
Aggregate
core/agenta
A self-contained Agenta installation packaged as a reusable module: the workspace in which a team builds agents that do work, by describing the job in chat, connecting the applications the agent needs, setting what it may do on its own, and correcting it through feedback. It composes fourteen single-workload components, from a control-plane API and an agent runner to the PostgreSQL, Redis-compatible, object-store and authentication services they share. It is not a trace store, which core/observability/langfuse is, nor a prompt registry, which core/devtool/promptlayer is. Its credentials, the object store's key pair included, are module parameters an importing Aggregate passes down, each attached through core/referenced-secret on every slot that holds it, on the connection where a component presents it. The key the platform signs its internal tokens with and the key it encrypts stored credentials under have no default; every other default is a development credential every reader holds, for evaluation only. A Release sets infra.external_base_url on the api, services, console and mobile components, or the build halts. Realized on helm: its Agenta components have no compose Recipe. The Manifest of every Release importing it maps core/referenced-secret, or the build halts, and maps the backend to a PostgreSQL Recipe other than postgres-cloudnative-pg, which halts on the filled credential slot. Both Redis-compatible instances run without authentication. The module ships no reverse proxy, no remote sandbox provider, no model provider and no outbound mail; an ingress tier routes to the four components it exports.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/agenta:1.0.0 - v1.0.0
Aggregate
core/apisix
A self-contained API-gateway stack: a dynamic API gateway in front of an etcd configuration store. Packaged as a reusable module composing two single-workload components — the gateway and its store — wired by an internal edge that carries the store's host and client port into the gateway's configuration-store connection. The module exports the gateway as `proxy`, the API endpoint consumers connect to. The gateway administers its routes through an Admin API whose key is the module parameter `admin_key`, with no default: an importing Aggregate passes one down and its Release fulfills it, or the build halts. The module attaches `core/referenced-secret` to the gateway's `admin_key` slot with it, so the Manifest of every Release whose Aggregate imports the module maps `core/referenced-secret`, or the build halts.
ОфіційнийОпубліковано 2026-11-10oci://hub.aggrega.ai/aggregate/core/apisix:1.0.0