OpenClaw Installation & Safety Guide: How to Set Up Your Own AI Assistant Without Getting Burned

OpenClaw is a free, open-source AI assistant you install and run on your own device rather than through a website. Once set up, it can chat with you through WhatsApp, Telegram, Slack, or Discord. "Open-source" means the code is free and publicly inspectable, but you still pay the AI provider (Anthropic or OpenAI) for every message it processes.

The one rule that matters most: never install OpenClaw on the laptop or PC you use for banking, work, or personal photos. Use a separate, dedicated device. Everything else in this guide builds on that decision.

Key terms, explained simply

Gateway is the background control centre that manages every request and response. API key is the password that lets OpenClaw talk to Anthropic or OpenAI. Token is the unit providers bill you by (roughly 0.75 words per token). Daemon is a program that keeps running silently in the background. WSL lets Windows run Linux software internally, which OpenClaw needs. Docker runs OpenClaw in an isolated container for extra safety. VPS is a rented cloud computer, usually $5–$20 a month, that works as your dedicated device if you don't have spare hardware.

What you'll need

A dedicated device: a Mac Mini (M1 or newer, roughly $3–4/month to run), an old laptop wiped of personal files, or a cheap VPS from a provider like Hostinger or Hetzner. You'll also need Node.js 22+ (the installer handles this), macOS/Linux/Windows with WSL2, a browser, and an account with Anthropic or OpenAI with a payment method attached. If you plan to use WhatsApp, get a cheap prepaid SIM instead of your personal number.

Should you use Docker?

Docker adds an isolated "container" around OpenClaw, like a locked room inside your house. It's strongly recommended on a VPS or if you'll experiment with third-party plugins. It's less critical if you're already on a completely separate device with no personal data, though it's still good practice. Beginners can skip it initially and add it later using the official Docker install guide.

Installing OpenClaw, step by step

Step 1 – Prepare your device. On Mac, open Terminal (Command + Space, type "Terminal"). On Windows, open PowerShell as Administrator, run wsl --install, restart, then reopen PowerShell and type wsl to enter Linux mode.

Step 2 – Run the installer. Paste this into your Terminal or WSL window and press Enter: curl -fsSL https://openclaw.ai/install.sh | bash

This installs Node.js if needed and launches the setup wizard automatically.

Step 3 – Follow the wizard. Accept the security warning after reading it. Choose "Quickstart" mode. Pick your AI provider (Anthropic or OpenAI) and paste in your API key. Start with a mid-tier model like Claude Sonnet or GPT-4o rather than the priciest options. When asked about running as a daemon, choose Yes so it keeps running after you close the terminal.

Step 4 – Verify it works. Run openclaw doctor (checks for problems), openclaw status (confirms the gateway is active), and openclaw dashboard (opens the visual control panel in your browser).

Connecting Telegram

Telegram is the easiest and most secure messaging channel to start with. Your bot is entirely separate from your personal account; nobody who messages it sees your phone number.

Create the bot: In Telegram, message @BotFather, send /newbot, choose a name and a username ending in "bot." BotFather returns a bot token, copy it and store it somewhere other than the OpenClaw device itself.

Find your user ID: message @userinfobot and note the number it replies with.

Connect and pair: if you didn't select Telegram during setup, run openclaw onboard, choose Telegram, and paste the bot token. Then message your bot in Telegram; it will send you a pairing code. Back in Terminal, run openclaw pairing approve telegram CODE to confirm ownership.

Lock it down by running:

openclaw config set channels.telegram.dmPolicy allowlist
openclaw config set channels.telegram.allowFrom [YOUR_USER_ID]
openclaw gateway restart

If you connect WhatsApp instead, use a dedicated prepaid number only, since every message to that number, including bank codes and family chats, becomes AI input.

Setting up billing without surprise charges

Anthropic and OpenAI both use prepaid credits: you buy a balance, and it depletes as you use the assistant. Running out simply pauses the AI, which is a useful safety net.

The single most important step: turn off Auto-Recharge (Anthropic calls it "Auto-Reload"). If OpenClaw ever loops or a scheduled task misbehaves, auto-recharge will keep refilling your balance and charging your card indefinitely.

For Anthropic: create an account at console.anthropic.com, choose the Build plan, add a payment method, buy $10–$20 in credits, then go to the Auto-Reload section and switch it off. Create your API key under Settings → API Keys.

For OpenAI: create an account at platform.openai.com, go to Settings → Billing, buy a small credit amount, and disable Auto-Recharge (or set the lowest possible recharge limit if it can't be fully disabled). Create your key under API Keys.

Avoid connecting Google Cloud or AWS unless you fully understand their billing, since their spending "limits" are often just notifications, not hard stops. Use a virtual or prepaid card with a fixed balance where possible, and expect $6–$15/month for light personal use, or $50–$200+ for heavy automation and premium models.

Security and privacy checklist

Run through these once after installation, and revisit the first few regularly:

Keep OpenClaw on its dedicated device only. Run it under a normal user account, never as Administrator or root. Turn on consent mode with openclaw config set exec.ask on so it asks before taking actions. Leave the gateway on localhost, never expose it or port 18789 to the public internet. Run openclaw doctor weekly to catch issues early. Check your token with openclaw config get gateway.token and keep it secret. Update monthly with openclaw update --channel stable. Restrict the config folder with chmod 700 ~/.openclaw. Only install plugins from official OpenClaw sources; malicious third-party skills have been found before. Never type passwords, bank details, or national ID numbers into a chat with the agent.

Ongoing maintenance

Weekly: run openclaw doctor, check your provider's usage dashboard, and glance at logs with openclaw logs --follow. Monthly: update OpenClaw and your OS, rotate your gateway auth token, review and trim scheduled tasks, and clear old chat sessions to keep token costs down (long conversations resend their full history, so they cost more per message).

If something feels wrong

Stop the gateway immediately with openclaw gateway stop, or unplug the device as a last resort. Check openclaw logs --follow and run openclaw doctor to diagnose the issue. If you suspect a real compromise, wipe the device and reinstall from scratch, restoring your configuration from backup.

The bottom line

OpenClaw gives you a capable, private AI assistant, but it comes with real system access and real billing risk. Keep it isolated on its own device, keep a human approving actions through consent mode, keep Auto-Recharge off, and check in with openclaw doctor regularly. Follow those habits and you'll have an assistant that's genuinely useful, on your own terms.