OpenClaw
OpenClaw is the most popular self-hosted AI agent framework. You host it on your own machine or VPS, choose your model provider, connect any channel (WhatsApp, Telegram, Discord, email), and extend it with skills. Every guide on this hub is independently researched and annotated — we link back to creators and original sources.
Guides
What Is OpenClaw?
OpenClaw is a personal AI assistant runtime that you host on your own machine or VPS. Unlike hosted products (ChatGPT, Claude Cowork), OpenClaw gives you control over which model provider answers your questions, which channels reach the agent, and which skills your agent can run.
The gateway routes incoming messages to the right agent, applies allowlist rules, and isolates sessions per sender. A user can start with a Claude backend, swap to a local Ollama model mid-project, and keep the same skills and conversation history. The skill ecosystem — via ClawHub — ranges from "get the weather" to "manage GitHub PRs" to "rebalance DeFi positions."
If you've used ChatGPT or Claude and wished you controlled the plumbing, OpenClaw is the direct answer. Compare it with similar agents: IronClaw (security-hardened fork), NemoClaw (NVIDIA-optimised), or see the full comparison table on the homepage.
At a Glance
| License | MIT (free, open source) |
| Install | npm install -g openclaw |
| Requires | Node.js 22.16+ or Node 24 |
| Channels | WhatsApp, Telegram, Discord, iMessage, email, and more |
| Model providers | Anthropic, OpenAI, Ollama (local), and any OpenAI-compatible API |
| Official skills | 53 (see Skills Database) |
| Community skills | 13,700+ on ClawHub (third-party, use with caution) |
| Typical monthly cost | ~$7 (cheap model) to ~$20 (typical) — full breakdown |
OpenClaw Use Cases — Real-World Setups
Concrete things people actually build with OpenClaw, with full setup steps and cost breakdowns.
- Morning brief — daily 7am email + calendar + news digest delivered to Telegram
- Family calendar coordinator — one agent watches everyone's schedule
- Daily journal with mood tracking — privacy-first, runs locally
- Code review automation — PR triage with style + correctness checks
- Dependency updater — weekly npm/pip refresh with changelog summary
- All 12 use cases →
OpenClaw Troubleshooting — Common Errors
The errors people actually hit on day 1 — with copy-paste fixes.
- npm EACCES on global install — don't use sudo with npm
- Gateway failed to start — port in use — find and free the bound process
- Rate limit 429 (Anthropic provider) — backoff and retry strategy
- All troubleshooting entries →
OpenClaw Security — What to Harden First
Self-hosted OpenClaw runs as your OS user with default-allow skills. These are the controls that matter most.
- Skill & tool allowlisting — the single highest-impact control
- Secrets & credentials — never in SOUL.md, always in .env
- Sandboxing — contain the blast radius of an agent mistake
- Prompt injection — the #1 vulnerability for any agent that reads wide
- 15-minute hardening checklist — the OpenClaw-specific essentials
Related on This Site
- IronClaw — security-hardened fork with mandatory skill allowlisting and sandboxed defaults
- NemoClaw — OpenClaw on NVIDIA OpenShell, containerised and policy-controlled
- Decision guide — pick the right agent for your use case
- Weekly News Digest — OpenClaw releases, CVEs, and community updates every Monday