Home lab

Check your home server’s containers every 15 minutes. If one fails, the agent reads its logs, reports a likely cause, and proposes a restart or recreation for you to approve in Slack.

Requires uv, a daemon user with access to the Docker socket, and a Slack app.

The files

~/.config/wardrik/config.yaml

mcp:
  docker:
    command: uvx
    args: [mcp-server-docker]

channels:
  slack:
    type: slack
    token: ${SLACK_BOT_TOKEN}
    signing_secret: ${SLACK_SIGNING_SECRET}
    members: {U024BE7LH: user:admin}

notify:
  targets:
    homelab: {channel: slack, room: "#homelab"}

~/.config/wardrik/policies/homelab.yaml

allow:
  - docker/list_*
  - docker/fetch_container_logs
ask:
  docker/stop_container: This stops a service.
  docker/start_container: This starts a service.
  docker/recreate_container: This recreates a container from its image.
deny:
  docker/remove_*: Nothing is deleted.
  docker/run_container: New containers are started by you.
  docker/build_image: Images are built by you.
budget: {usd: 0.50, per: day}

~/.config/wardrik/agents/homelab.md

---
description: Keeps my home lab's containers running.
tools: [docker/*]
policy: homelab
notify: {to: [homelab], on: [parked, failed]}
triggers:
  - name: check
    every: 15m
    prompt: Check every container.
---
List the containers. If one has exited, keeps restarting, or is unhealthy, read its last 100 log
lines and say in two sentences what went wrong. Then propose the smallest fix: restart it, or
recreate it from its image. If everything is fine, answer "All good."

Try it

systemctl --user restart wardrik    # reads config.yaml and policies/
docker stop jellyfin
wardrik ask homelab "Check every container."

#homelab shows the findings and a start_container call for jellyfin, with Approve and Deny. After editing only an agent file, wardrik reload is enough.

How it works

• notify sends only pending calls and failures, so healthy checks post nothing.
• Docker socket access grants extensive control over the host. The policy requires approval for changes and denies destructive operations.
• Logs are untrusted content, so gated calls require approval after the agent reads them.
• A turn interrupted by a power loss resumes when the daemon starts. See Durable runs.