OpenClaw operations
OpenClaw setup field guide for a real operator workstation.
A practical baseline for running an AI employee cockpit: stable gateway, clean secrets, explicit model routing, durable tasks, and human approval for anything that leaves the machine.
Recommended baseline
- Pin the working OpenClaw version. Do not casually upgrade a production assistant. Back up first, check release notes/community signal, then test.
- Keep the Gateway loopback-only. Prefer
127.0.0.1, token auth, and no unnecessary network exposure. - Put secrets in local auth files, not chat. Use local environment/auth files and provider profiles. Do not paste API keys into conversations, docs, screenshots, or public repos.
- Use a durable task layer. Linear, local JSON, or another system should hold work state so chat is not the project database.
- Use a visible cockpit. A local Mission Control page makes agents, tasks, health, reports, and artifacts reviewable.
- Make writes approval-gated. Email, posts, purchases, client-facing messages, and production changes need explicit human review.
Setup pattern
Start with identity and boundaries
Define who the assistant is, what it can touch, what it must never touch, and what requires approval. This avoids accidental scope creep once the system becomes useful.
Keep startup context compact
Load durable facts and current checklists, not every historical note. Heavy context costs money and makes behavior less predictable.
Route models intentionally
Use ChatGPT/Codex for daily work, premium models for final review and hard judgment, and local models only when their update/runtime behavior is acceptable for the path.
Build narrow skills
Prefer native Node.js, zero dependencies, clear help text, documented side effects, and fail-closed behavior. Make scripts auditable in minutes.
Snapshot health
Heartbeat checks should capture gateway status, disk, pinned version, serious logs, task state, and blockers. Alert only on meaningful changes.
What this becomes for clients
This pattern turns into an AI employee implementation service: setup, model routing, skills, approvals, Mission Control, task layer, weekly reports, and operational guardrails. The deliverable is not a chatbot. It is a small operating system for useful AI work.