The OpenClaw Runbook
Run OpenClaw without burning money, quotas or your sanity.
A practical operations guide: what works after running OpenClaw, breaking it, and doing that loop more times than anyone would like to admit. Adapted from digitalknk/openclaw-runbook; not official guidance.
- Chapters
- 8
- Sections
- 100
- Reading time
- 87 min
Eight chapters
Read in order for a complete setup, or jump to whatever is on fire.
- 01
Cost Optimization
Model routing, hardware decisions, cheap models that actually work, and real monthly cost breakdowns. Stop burning money on defaults.
Models · Budget · Hardware12 min - 02
Agent Prompts
Five ready-to-use specialized agents: Monitor, Researcher, Communicator, Orchestrator and Coordinator, with complete model chains and fallback strategies.
Agents · Prompts · Models15 min - 03
Memory Configuration
Cheap embeddings, context pruning with cache-TTL, and compaction that prevents the 'why did it forget that' moments.
Memory · Embeddings · Context8 min - 04
Heartbeat & Task Tracking
A rotating heartbeat that batches checks on a cheap model, plus Todoist integration for full visibility into what your agent is doing.
Heartbeat · Todoist · Visibility10 min - 05
VPS Deployment
A full Hetzner setup with Tailscale SSH, firewall hardening, gateway auth, openclaw backup, systemd service management and production stability patterns.
VPS · Hetzner · Tailscale14 min - 06
Security Hardening
Prompt injection defense, tool policies, gateway lockdown, sandbox mode and security audit workflows. Assume someone will try to steer your agent.
Security · Injection · Tools10 min - 07
Skills & Tools
Build your own skills to the AgentSkills spec, monitor quotas across providers, and see why third-party skills need scrutiny.
Skills · Quotas · Scripts8 min - 08
Config Reference
An annotated openclaw.json walkthrough, a sanitized production config, file structure, custom providers and a post-setup checklist.
Config · JSON · Setup10 min
Before taking blind advice
OpenClaw changes fast, sometimes daily. When something strange happens, check the official resources first; open issues and recent PRs save hours. You don't need expensive hardware or subscriptions, just deliberate configuration and visibility into what's happening.