OpenClaw Quick Start — Install in Under 10 Minutes
Everything you need to go from zero to a running AI agent: install, first gateway start, recommended model setup, and a cost breakdown. Runs on Linux, macOS, or Windows (WSL recommended on Windows).
Prerequisites
- Node.js 22 LTS or Node 24 — download from nodejs.org
- An API key from your model provider: Anthropic, OpenAI, or a local Ollama install (free, runs offline)
- A terminal — bash, zsh, or PowerShell on Windows with WSL
Install Steps
-
Install the OpenClaw CLI globally:
npm install -g openclaw -
Initialize your config:
openclaw initYou'll be prompted to choose a provider (Anthropic, OpenAI, Ollama) and paste your API key. Keys are stored locally — they never leave your machine.
-
Start the gateway:
openclaw gateway start -
Install the healthcheck skill and verify everything works:
openclaw skill install healthcheck openclaw doctoropenclaw doctorchecks your config for risky DM policies, missing credentials, and unreachable channels. Run it after any config change.
On a VPS or home server, run the gateway inside tmux or screen so it survives SSH disconnects. The official tmux skill handles this automatically — see Tips below.
Recommended Model Setup
There's no single right answer — it depends on budget, privacy needs, and workload. Here's what works for most use cases:
| Use Case | Model | Hosting | Notes |
|---|---|---|---|
| Daily personal assistant | Claude Haiku 4.5 | Local laptop or $5 VPS | Cheap, fast, capable for most tasks |
| Complex multi-step tasks | Claude Opus 4.6 | VPS or home server | Highest reasoning, higher per-token cost |
| Full-privacy, offline | Ollama (local) | Home workstation with GPU | Zero external calls; quality varies by model |
| Team / multi-channel bot | GPT-4.1 or Claude Sonnet 4.6 | DigitalOcean / Hetzner VPS | Good cost/reliability balance at volume |
What It Costs
OpenClaw itself is free under MIT. Your recurring costs are (1) the model provider and (2) hosting. Approximate monthly ranges as of 2026-04-06:
| Component | Low | Typical | Heavy |
|---|---|---|---|
| Model — Claude Haiku 4.5 | $2 | $8 | $25 |
| Model — Claude Opus 4.6 | $15 | $60 | $200+ |
| Model — Local Ollama | $0 | $0 | $0 (electricity only) |
| VPS hosting | $5 | $12 | $40 |
| Total (typical) | ~$7 | ~$20 | ~$65+ |
Heavy usage assumes 8+ hours/day of active agent time with frequent long-context calls.
Tips & Tricks
- Pin a cheap model as the default. Configure OpenClaw to escalate to Opus only when the agent explicitly requests it. Saves 80%+ on API bills for most users.
- Use the tmux skill.
openclaw skill install tmuxkeeps gateway sessions alive across SSH disconnects. Essential if you're running on a remote VPS. - Let the agent debug itself. Give it read access to its own log path and tell it to diagnose errors. It's usually faster than reading logs manually.
- Use onlycrabs.ai SOUL.md files to give your agent a persistent personality and house rules that survive gateway restarts.
- Schedule daily-brief on cron. A 7 AM summary of calendar + weather + tasks is the highest-ROI automation most users build first.
Troubleshooting
Gateway won't start — "port already in use"
Another OpenClaw process is already bound. Run openclaw gateway stop, then start again. If that fails, find the orphaned process: lsof -i :PORT (Linux/Mac) or netstat -ano | findstr :PORT (Windows).
Skill install fails with "signature mismatch"
ClawHub verifies skill signatures. A mismatch usually means the registry mirror is stale. Update the CLI first: npm install -g openclaw@latest and retry.
My WhatsApp channel isn't responding
Run openclaw doctor. Nine times out of ten it's an allowlist rule that's too strict or an expired session token. Check the allowFrom configuration in your YAML.
The agent is burning through my API quota
Pin a cheaper default model, lower max_tokens, and shorten the system prompt. Audit any skills that make repeated background calls on cron — they're the most common culprit.
How do I switch to a different model provider?
Run openclaw init again. Config is idempotent — your skills and conversation history persist. Only the provider endpoint and API key change.
← Back to OpenClaw hub · See also: Skills Guide · Security Hardening