Want to understand what OpenClaw is, or want the fastest way to start using it? The simplest path is ZenClaw. OpenClaw managed service, 9 seconds to spin up OpenClaw. Sign in, click a button, done. No Node, Docker, certs, or DNS pits to fall into. OpenClaw is the personal AI agent maintained by Peter Steinberger and the open-source community. This post covers the definition, core features, community ecosystem, and the core difference from ChatGPT.
The definition of OpenClaw
OpenClaw is an open-source personal AI agent: lives on your own machine, connects to the messaging apps you already use, calls models like Claude, GPT, and Gemini, keeps data local, and is maintained by Peter Steinberger and the community. Key words broken down:
- Open source — code is public at github.com/openclaw/openclaw. Read, modify, fork.
- Personal agent — unlike SaaS subscriptions (ChatGPT, Claude Web), this is your own agent.
- Persistent — runs 24/7 on your host. Responds when you call it. Can push proactively.
- Cross-channel — Telegram, LINE, WhatsApp, Slack, Discord, and more. No new app to open.
- Multi-model — you pick the backend model.
The mascot is a space lobster named Molty, featured on openclaw.ai.
OpenClaw’s core features
Messaging channels + multi-device + model abstraction + skill system + local state. Five things that make up all of OpenClaw. Breakdown:
Messaging channels
OpenClaw upstream supports Telegram, LINE, WhatsApp, Slack, Discord, and more. Some channels (iMessage, WebChat) are community-contributed or experimental. WhatsApp uses Baileys (a reverse-engineered QR session connection); official docs at docs.openclaw.ai/channels/whatsapp. The ZenClaw managed dashboard currently offers Telegram, LINE, and Microsoft Teams, click-to-connect from the dashboard with no webhook setup.
Multi-device and gateway
OpenClaw runs a gateway on port 18789 by default. iOS, Android, and macOS clients talk to your agent through it, and cross-device state remembers where the conversation left off. The official security docs emphasize that the gateway should bind to 127.0.0.1 and remote access should go through Tailscale or an SSH tunnel, not straight to the public internet.
Model abstraction (LLM routing)
You set the active model in openclaw.json. OpenClaw isn’t locked to a single provider. Common combos: Claude Opus for reasoning, GPT-4o for multimodal, Gemini for long documents. Anthropic’s pricing is at anthropic.com/pricing.
Skill system
Skills are OpenClaw’s plugin system. You can install community skills directly. Common ones: RAG lookup, Shopify order lookup, Google Calendar, Notion sync. Simon Willison’s OpenClaw Docker writeup notes that skills are flexible but deserve careful review.
Local state
OpenClaw stores everything under ~/.openclaw/ by default:
openclaw.json— main config (model, channels, skill toggles)sessions/— per-channel conversation sessionsagents/— the multiple agent files you createcredentials/— messaging tokens, WhatsApp Baileys session (encrypt this)skills/— installed skills
Which means: back up and migrate by saving this whole directory. This Docker common pitfall is the volume not being mounted so ~/.openclaw gets wiped on restart.
Who is Peter Steinberger, and what’s the community like
Peter Steinberger is the core maintainer and community leader of OpenClaw. He’s been active in the iOS open-source world for years (founder of PSPDFKit). OpenClaw is maintained at github.com/openclaw/openclaw, with community across GitHub, Discord, and X/Twitter. Highlights:
- GitHub — main repo at github.com/openclaw/openclaw, issues and stars still climbing
- Docs — official docs.openclaw.ai covers getting started, CLI, channel setup, and skill development
- Contributors — many people contributing, including channel adapters, skills, and model connection layers
- Derivative versions — NVIDIA’s NemoClaw (enterprise-hardened version) and community self-hosted variants
How long does self-hosting OpenClaw take
The official getting started docs say 5–10 minutes, but community reports range from several hours to several weeks. Challenging for non-engineers. Public cases:
- A developer on LinkedIn logged “8 hours across 3 days” to get OpenClaw working
- The GitHub repo ishwarjha/openclaw-setup-guide-i-wish-i-had opens its README with “15 days of tinkering”
- The WhatsApp Baileys session rescan loop in Issue #9096 is a common trap
What you’re actually handling: recent Node versions (see official docs), Docker permissions, the ~/.openclaw volume, TLS certs, firewall rules, skill allowlists, and chasing CVE upgrades. Blink.new stats show OpenClaw had accumulated roughly 138 known CVEs by April 2026.
Why ZenClaw is the best choice for most people
ZenClaw is an OpenClaw managed service: sign in, click “Hire AI Employees Now”, 9 seconds later you have a working OpenClaw instance with the backend on our cloud and certs, upgrades, and security all handled by us. Versus self-hosting:
| Dimension | Self-host OpenClaw | ZenClaw |
|---|---|---|
| Time to launch | Hours to weeks | ✅ 9 seconds |
| Technical skills needed | Node, Docker, DNS, firewall | ✅ None, just a browser |
| Certs / TLS | DIY | ✅ Automatic |
| Upgrades / CVE | Chase yourself | ✅ Automatic |
| Cost | Host + API + your time | ✅ Monthly subscription, see pricing |
| Channels | All (connect yourself) | ✅ Telegram, LINE, Microsoft Teams (click-to-connect) |
| Sandbox | Configure yourself | ✅ NVIDIA enterprise sandbox (plans include NemoClaw sandbox, see pricing) |