Deploy a Flow

This guide deploys a single Flow exposed at POST /greeter, backed by Anthropic Claude.

1. Credentials

kubectl create secret generic anthropic-credentials \
  --from-literal=ANTHROPIC_API_KEY=sk-ant-...

2. Plugin

apiVersion: genkit.dev/v1alpha1
kind: PluginConfig
metadata:
  name: anthropic
spec:
  type: anthropic
  credentialsRef:
    name: anthropic-credentials
  credentialKeys: [ANTHROPIC_API_KEY]

3. Model

apiVersion: genkit.dev/v1alpha1
kind: Model
metadata:
  name: claude-opus
spec:
  provider: anthropic
  model: claude-opus-4-6
  pluginConfigRef:
    name: anthropic
  defaultConfig:
    temperature: 0.3
    maxOutputTokens: 1024

4. Prompt

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.

5. Flow

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 }
  port: 8080

6. Apply and verify

kubectl apply -f .
kubectl get gfl greeter
kubectl describe gfl greeter   # check Conditions

When Ready=True:

kubectl port-forward svc/greeter 8080:8080 &
curl -s -X POST http://localhost:8080/greeter \
  -H 'content-type: application/json' \
  -d '{"name":"Ada"}'

7. Iterate

Edit the Prompt, reapply, and watch the rolling update:

kubectl edit prompt greeting
kubectl rollout status deploy/greeter

You don’t need to bump anything else — the controller recomputes the content hash and patches the Deployment’s Pod template annotation automatically.

Inline model spec and prompts (no Model or Prompt CRs)

If you prefer to keep everything in one manifest — or want to avoid creating separate Model and Prompt CRs — you can embed both the model definition and the prompt content directly in the Flow.

You still need a PluginConfig (and its Secret) for credentials; only the Model and Prompt CRs become optional.

1. Credentials and plugin (same as before)

kubectl create secret generic anthropic-credentials \
  --from-literal=ANTHROPIC_API_KEY=sk-ant-...

apiVersion: genkit.dev/v1alpha1
kind: PluginConfig
metadata:
  name: anthropic-config
spec:
  type: anthropic
  credentialsRef:
    name: anthropic-credentials
  credentialKeys: [ANTHROPIC_API_KEY]

2. Flow with inline model spec and inline prompt

apiVersion: genkit.dev/v1alpha1
kind: Flow
metadata:
  name: greeter
spec:
  image: ghcr.io/xavidop/genkit-runner:v0.4.1
  modelSpec:
    provider: anthropic
    model: claude-opus-4-7
    pluginConfigRef:
      name: anthropic-config
    defaultConfig:
      temperature: 0.3
      maxOutputTokens: 1024
  prompts:
    - prompt:
        name: greeting
        content: |
          ---
          model: anthropic/claude-opus-4-7
          ---
          Greet the user named {{name}} in a single sentence.
  port: 8080

modelSpec and modelRef are mutually exclusive — pick one. Likewise, each item in prompts uses either promptRef (a reference to a Prompt CR) or prompt (inline content), not both.

3. Apply and call

kubectl apply -f plugin-config.yaml -f flow-inline.yaml
kubectl get gfl greeter

kubectl port-forward svc/greeter 8080:8080 &
curl -s -X POST http://localhost:8080/greeter \
  -H 'content-type: application/json' \
  -d '{"name":"Ada"}'