Claude runtimes

The daily run and the deep dive call Claude, and so do bootstrap, seed and migrate --topics when someone runs them. agents.runtime in homeostat.yaml picks how, which decides where review comments go and whose contract covers them. The plugin, hooks and MCP server never call a model, whatever the runtime.

Runtime	Where prompts go	Credentials	Use it when
api (default)	Anthropic’s Claude API	ANTHROPIC_API_KEY (a Console key)	Most teams. API data is not used for training by default; zero data retention is available on request.
bedrock	Claude in Amazon Bedrock, inside your AWS account	The AWS credential chain: an IAM role in CI (OIDC), SSO, or AWS_* variables	Data must stay inside your AWS boundary and contracts.
vertex	Claude on Google Cloud (Agent Platform), inside your project	Application Default Credentials: workload identity federation in CI, or gcloud auth application-default login	Data must stay inside your Google Cloud boundary.
claude-code	Whatever Claude Code is logged in to	A Claude Code login	Personal use, or a Team/Enterprise plan. A personal subscription is for non-commercial use only.

Cloud setup

agents:
  runtime: bedrock
  bedrock:
    region: eu-west-1        # optional; else AWS_REGION

agents:
  runtime: vertex
  vertex:
    projectId: my-project    # else ANTHROPIC_VERTEX_PROJECT_ID
    region: eu               # global (default), us, eu, or a region

Model names stay as they are (claude-opus-5-5); on Bedrock the anthropic. prefix is added for you, and a full inference-profile ARN is passed through unchanged.

The cloud SDKs are optional, so teams on the default runtime don’t install the AWS or Google libraries. Add the SDK to the CI command:

npx --yes -p homeostat@<version> -p @anthropic-ai/bedrock-sdk@0.34 homeostat daily
npx --yes -p homeostat@<version> -p @anthropic-ai/vertex-sdk@0.20 homeostat daily

Give the CI identity only what it needs: on AWS, bedrock-mantle:CreateInference on the model ARNs you use; on Google Cloud, the Vertex AI User role on the project. Prefer federated identities (GitHub OIDC to an IAM role or to workload identity) over stored keys.

Differences on the clouds

• Neither cloud supports the server-side refusal fallback, so it is not sent; a refused thread is skipped and the run continues.
• Bedrock’s Messages endpoint has no structured outputs, so on both clouds the answer comes back as one forced tool call whose input is checked against the same schema.
• Every model call is logged locally as a model-call event with its runtime, model and size, never its content.