My OpenClaw fleet
The agents I actually run, on a runtime I didn't write.
OpenClaw is an agent runtime someone else built. I run my own fleet on top of it — gateway on a Mac mini in the corner of the office, each agent with its own identity and skills, skills stored as versioned Markdown in a public repo, secrets in SOPS, backups to GitHub on a cron.
The interesting bit isn’t the runtime — that work has been done. The interesting bit is what I’ve chosen to put on it, and what I’ve learned about running an actual fleet of agents in the house.
What’s running today§
- eve-family — household assistant agent, routed to via a front-door agent.
- william-agent — Type 1 diabetes-aware coach for my son. Carefully scoped. Health-critical messages flag parents, not just him.
- A few agents for the 3D printing hobby — print queue, filament stock, catalogue of what I’ve made.
- scc-* — agents that admin the Stafford Camera Club website.
- home-assistant — bridges the smart home into the rest of the fleet.
- A handful of dev-facing helpers I actually use every day (Jira, Cloudflare, deploy bots).
What I’ve actually built around it§
- Skills as Markdown in a public repo. Anyone can audit what an agent is told. Diffs are reviewable. Future-me thanks present-me.
- SOPS-managed secrets. Per-agent credentials, age-encrypted, kept out of the runtime config.
- A backup script on cron that syncs identity + workspaces back to GitHub. I’ve restored from it.
- A discipline around scope. No agent does more than one job. eve-family routes to the right agent rather than knowing about everything.
What I’d write up properly later§
How the fleet actually feels day-to-day, what’s earned its keep and what’s been quietly switched off, and what I’d want from any agent runtime if I were choosing again now.
That post is coming. I want a few more months of running it before I write down the things I’m sure of.