Prerequisites

You need a Kubernetes cluster and the usual client tooling. For local development a single-node cluster is plenty.

Tooling

• kubectl: configured to talk to your target cluster.
• A Kubernetes cluster: v1.27+. Any distribution works; for local use see Local cluster.
• A container registry: only if you build your own images. The published controller and agent images are public, so a normal install pulls them with no registry, credentials or side-loading of your own.

To build from source

The project is a D monorepo. To build the binaries yourself you need:

• LDC (the LLVM-based D compiler): used for static linking.
• dub: the D package manager and build tool.

See Building.

Credentials

Agents call a model provider, so the agent process needs credentials available in the Pod. There is one supported way to provide them: create a namespace Secret named agent-secrets, then reference its keys from the recipe’s resources.secrets (e.g. {name: ANTHROPIC_API_KEY, ref: ANTHROPIC_API_KEY}). The controller injects each entry as a container env var via secretKeyRef, sourced from that single Secret. See Launch an agent.

Note

Earlier versions suggested mounting your host ~/.claude into the run container for subscription auth during local development. That no longer works: the ai-agents namespace enforces Pod Security Admission at baseline, which rejects hostPath volumes — deliberately, since the Station template’s preserve-unknown-fields passthrough would otherwise leave that escape hatch open. Use an API key in agent-secrets on every cluster, local included.