The biggest mistake beginners make is trying to wire everything at once. Don't. Get the base system running with one channel and one model. Prove it works. Then layer in the rest. OpenClaw rewards clean sequencing. It punishes chaos with interest.
What you need before you start
- A machine or VPS where OpenClaw will run (Ubuntu 22.04/24.04 LTS recommended for VPS)
- At least 2GB RAM (4GB recommended) and 10GB+ disk
- Node.js v20 or newer (check with node --version)
- An API key from an LLM provider (Anthropic, OpenAI, or Google)
- Docker and Docker Compose (if using the container route, which is recommended)
Step 1: Install OpenClaw
There are two main paths. Pick one.
Option A: Direct install (quick start)
Run npm install -g openclaw then openclaw init. The init wizard walks you through the basics.
Option B: Docker install (recommended for VPS)
Clone the repo, configure your .env file, and run the setup script. Docker gives you isolation, persistence, and a clean environment that won't mess with your host OS.
Full Docker steps are in the official Docker install docs.
Step 2: Configure your LLM provider
OpenClaw is not an LLM itself. It gives tools and structure to whatever model you connect. You need an API key from at least one provider.
- Anthropic (Claude): Most common starting point. Sonnet for general use, Haiku for cheap/fast tasks.
- OpenAI (GPT): Solid choice if you already have an OpenAI account.
- Google (Gemini): Good for web-grounded tasks.
- Local models (Ollama, LM Studio): Free but slower. Good for privacy-first setups.
Add your key to the .env file (never hardcode it in config files).
Step 3: Connect your first channel
This is where OpenClaw starts feeling real. You connect it to a chat app you already use, and suddenly you have an AI assistant living in your Telegram, WhatsApp, or Discord.
- Telegram: Easiest to set up. Create a bot via BotFather, add the token. Done in minutes.
- WhatsApp: Uses QR pairing. Works well but sessions can expire and need re-pairing.
- Discord: Good for team setups. Create a bot in the Discord developer portal.
- Signal, iMessage, Slack: Also supported. Setup varies by platform.
Pick one. Test it. Make sure messages flow both ways before adding more.
Step 4: Run the health check
After setup, run openclaw doctor. This is the command most beginners skip and later regret skipping. It checks your config, connectivity, and common errors automatically.
Also useful: openclaw status for a quick overview, and openclaw gateway probe to verify the gateway is reachable.
Step 5: Install your first skill
Skills are what make OpenClaw more than a chat interface. They are modular packages that teach your agent new capabilities: web search, calendar management, email handling, coding workflows, and many more.
Install one from ClawHub with a single command. Start with something simple and immediately useful. Read the source code before installing (this matters for security).
Step 6: Secure your setup
OpenClaw has deep system access. Treat that with respect from day one.
- Run in Docker, not directly on your main machine
- Bind the gateway to localhost (127.0.0.1), not 0.0.0.0
- Set a strong gateway token in .env
- Firewall port 18789 on your VPS
- Store API keys in .env with restricted permissions (chmod 600)
- Run
openclaw security audit
Full details in the security best practices guide.
10 mistakes that trip up beginners
- Wrong API key: Using an OpenAI key when configured for Anthropic (or vice versa). Gateway shows "0 tokens used."
- Old Node.js: Needs v20+. Older versions cause obscure syntax errors.
- Missing build tools: On Linux, you need build-essential, git, and python3 for npm native modules.
- Port conflict: Default port 18789 blocked by another service.
- Not running in Docker: Running directly on your host OS exposes more than intended.
- DM policy left open: Anyone can message your agent.
- Skipping openclaw doctor: Misses easily fixable config issues.
- Missing Docker volumes: Data lost on container restart.
- Gateway exposed without auth: Port 18789 publicly accessible.
- Never starting fresh sessions: Accumulated context increases token costs. Use /new before heavy tasks.
What to do next
- Install one skill you actually need
- Test one real workflow end to end
- Read the security guide before adding more tools
- Check common mistakes if something feels off
Need help from people who already use this stuff?
Want help getting your setup stable faster?
Join My AI Agent Profit Lab for practical help from people who have already walked through the same setup friction and can save you hours of trial and error.
FAQ
How long does OpenClaw setup take?
The basic install takes 15 to 30 minutes. A full VPS deployment with Docker, a channel, and a few skills takes about an hour. Budget extra time if you are new to Docker or terminal commands.
Do I need a VPS to run OpenClaw?
Not strictly. You can run it on a Mac, a Linux desktop, or a Raspberry Pi. But for always-on availability, a cheap VPS (2GB+ RAM) with Docker is the most reliable path.
What is the most common beginner mistake?
Trying to connect every channel, tool, and skill on day one. Get the base system stable first with one channel and one model. Layer everything else after.
Which AI model should I use?
Anthropic Claude Sonnet is the most common starting choice for a good balance of quality and cost. You can switch models later. Avoid using the most expensive models (like Opus) as your default for routine tasks.
How do I keep costs under control?
Use a mid-tier model as default, start fresh sessions with /new before heavy tasks, match models to task complexity, and monitor your API usage. OpenClaw gives you the knobs. Turn them.