Custom Resources

All resources live in the genkit.dev/v1alpha1 API group and are namespaced.

Reference graph

Flow / FlowSet ── pluginConfigRef ──▶ PluginConfig ── credentialsRef ──▶ Secret
   │                                       ▲
   ├── modelRef ──▶ Model ─────────────────┘
   ├── modelSpec (inline, pluginConfigRef still required)
   ├── promptRefs ──▶ Prompt(s)
   ├── prompts[].prompt (inline)
   └── toolRefs ────▶ Tool(s)

A Flow (or each entry in a FlowSet) ties one Model together with one or more Prompts and (optionally) Tools. The Model points at a PluginConfig which points at a Secret.

You can also skip the Model CR entirely by embedding the model definition inline via modelSpec, and skip Prompt CRs by embedding prompt content directly via prompts[].prompt. The PluginConfig CR is still required when using modelSpec because it holds the credentials.

PluginConfig

A provider definition. The type field selects a builder in the runner (anthropic, openai, googleai, vertexai, ollama, bedrock, azureaifoundry).

apiVersion: genkit.dev/v1alpha1
kind: PluginConfig
metadata:
  name: anthropic
spec:
  type: anthropic
  credentialsRef:
    name: anthropic-credentials
  credentialKeys: [ANTHROPIC_API_KEY]
  # extraConfig is opaque — passed through to the runner as JSON.
  # extraConfig:
  #   baseURL: "https://api.anthropic.com"

Short name: gpc.

Model

apiVersion: genkit.dev/v1alpha1
kind: Model
metadata:
  name: claude-opus
spec:
  provider: anthropic
  model: claude-opus-4-6
  pluginConfigRef:
    name: anthropic
  info:
    label: "Anthropic — Claude Opus 4.6"
    supports:
      multiturn: true
      tools: true
      systemRole: true
  defaultConfig:
    temperature: 0.3
    maxOutputTokens: 1024

Short name: gmd.

Prompt

A Dotprompt document — YAML frontmatter followed by a Handlebars body.

apiVersion: genkit.dev/v1alpha1
kind: Prompt
metadata:
  name: greeting
spec:
  content: |
    ---
    model: anthropic/claude-opus-4-6
    temperature: 0.3
    ---
    Greet the user named {{name}} in a single sentence.

Short name: gpr.

Tool

A genkit-go ai.ToolDefinition plus a dispatch target — either an HTTP endpoint or a reference to a Flow.

apiVersion: genkit.dev/v1alpha1
kind: Tool
metadata:
  name: now
spec:
  definition:
    name: now
    description: "Return the current ISO-8601 timestamp."
    inputSchema:  { type: object, properties: {} }
    outputSchema: { type: string }
  implementation:
    http:
      url: "https://my-time-service.internal/now"
      method: POST

Short name: gtl.

Flow

A single HTTP endpoint backed by a runner Pod.

apiVersion: genkit.dev/v1alpha1
kind: Flow
metadata:
  name: greeter
spec:
  image: ghcr.io/xavidop/genkit-runner:v0.4.1
  modelRef: { name: claude-opus }
  promptRefs:
    - { name: greeting }
  toolRefs:
    - { name: now }
  port: 8080
  serviceType: ClusterIP

Exposed at POST /<flow-name> on the Pod’s port (default 8080). Short name: gfl.

Inline model spec

Instead of creating a Model CR and referencing it via modelRef, you can embed the model definition directly in the Flow using modelSpec. modelRef and modelSpec are mutually exclusive — use one or the other.

spec:
  modelSpec:
    provider: anthropic
    model: claude-opus-4-7
    pluginConfigRef:
      name: anthropic-config   # PluginConfig CR still required for credentials
    info:                      # optional
      label: "Anthropic — Claude Opus 4.7"
      supports:
        multiturn: true
        tools: true
        systemRole: true
    defaultConfig:             # optional
      temperature: 0.3
      maxOutputTokens: 1024

Field	Required	Description
`provider`	Yes	Provider identifier (e.g. `anthropic`, `openai`, `googleai`)
`model`	Yes	Model name as recognised by the provider plugin
`pluginConfigRef`	Yes	Reference to a `PluginConfig` CR that holds the credentials
`info`	No	Human-readable label and capability flags
`defaultConfig`	No	Default generation parameters (`temperature`, etc.)

Inline prompts

Instead of creating Prompt CRs and referencing them via promptRefs, you can embed prompt content directly in the Flow using the prompts list with a prompt entry. promptRef and prompt are mutually exclusive within a single list item.

spec:
  prompts:
    - prompt:
        name: greeting         # becomes greeting.prompt on disk
        content: |
          ---
          model: anthropic/claude-opus-4-7
          ---
          Greet the user named {{name}} in a single sentence.

Field	Required	Description
`name`	Yes	Logical name; written as `<name>.prompt` in the runner’s ConfigMap
`content`	Yes	Full Dotprompt document (YAML frontmatter + Handlebars body)

You can mix styles — some items can use promptRef while others use prompt inline — within the same prompts list.

FlowSet

Multiple flows in one Pod. Each flow gets its own per-flow ConfigMaps under /genkit/flows/<flow-name>/ and its own credentials mount.

apiVersion: genkit.dev/v1alpha1
kind: FlowSet
metadata:
  name: assistants
spec:
  image: ghcr.io/xavidop/genkit-runner:v0.4.1
  port: 8080
  flows:
    - name: greeter
      modelRef: { name: claude-opus }
      promptRefs:
        - { name: greeting }
    - name: summarizer
      modelRef: { name: claude-opus }
      promptRefs:
        - { name: summarize }

Routes: POST /greeter, POST /summarizer. Short name: gfs.

Inline model spec and prompts in FlowSet

Each flow entry inside FlowSet.spec.flows supports the same modelSpec and prompts[].prompt inline fields as a standalone Flow. This lets you define the entire FlowSet without creating any Model or Prompt CRs.

spec:
  flows:
    - name: greeter
      modelSpec:
        provider: anthropic
        model: claude-opus-4-7
        pluginConfigRef:
          name: anthropic-config
        defaultConfig:
          temperature: 0.3
      prompts:
        - prompt:
            name: greeting
            content: |
              ---
              model: anthropic/claude-opus-4-7
              ---
              Greet the user named {{name}} in a single sentence.
    - name: summarizer
      modelSpec:
        provider: anthropic
        model: claude-opus-4-7
        pluginConfigRef:
          name: anthropic-config
      prompts:
        - prompt:
            name: summarize
            content: |
              ---
              model: anthropic/claude-opus-4-7
              ---
              Summarize the following text in one paragraph: {{text}}

The same constraints apply as for Flow: modelRef and modelSpec are mutually exclusive per flow entry, and promptRef / prompt are mutually exclusive per list item.

Dataset and Eval

Used together to run scheduled evaluations against a flow. See the samples under config/samples/.

Kind	Short name	Purpose
`Dataset`	`gds`	A set of `{input, reference}` examples for evals
`Eval`	`gev`	Scheduled run of a `Dataset` against a `Flow`

Status conditions

Every CR carries the standard set:

Ready — child resources reconciled successfully.
Reconciling — the controller is currently working on it.
Degraded — a reference is missing or rendering failed.

Inspect with kubectl describe.