Every message you send to your AI costs a tiny amount. Like texting — each text costs a fraction of a cent. But if your AI is running 24/7 doing complex tasks, those fractions add up.
The danger: An AI that gets stuck in a loop can burn through your budget fast. Set limits.
| Model | Rough Cost per Message | Best For |
|---|---|---|
| Claude Haiku | ~$0.001 | Quick lookups, simple tasks |
| Claude Sonnet | ~$0.01-0.03 | Most daily work — the sweet spot |
| Claude Opus | ~$0.05-0.15 | Complex analysis, big coding projects |
| GPT-4o | ~$0.01-0.03 | Good general alternative |
It can see everything in your workspace. This is usually what you want — but be aware that anything you put there is visible to your AI.
It can install software, delete files, restart services. This is incredibly powerful but needs guardrails.
It can visit any website, take screenshots, fill out forms, research topics — like having a research assistant.
WhatsApp, Telegram, Discord — it can communicate on your behalf. Make sure you control who it talks to.
It can change how it behaves, update settings, install new features. Usually helpful, but you should know when it happens.
API calls, potentially purchasing things online if you give it payment access. This is why spending limits are crucial.
OpenClaw has an approval system for risky commands. You can configure what needs your permission vs. what the AI can do on its own.
Your AI runs on YOUR machine. Messages pass through your AI provider (Anthropic, OpenAI, etc.) to generate responses, but under API terms they commit to not training on your data. That's very different from consumer AI chat products like claude.ai or chatgpt.com.
This makes the file readable only by you — not even other programs on your computer can see it. Think of it like putting your secrets in a safe that only you have the combination to.
Act fast:
Never let your AI deploy directly to production without showing you first.
Don't deploy directly to production. Test everything on your own computer first.
Always have a way to undo changes if something goes wrong.
If something is running fine, don't let your AI "improve" it without asking you first. Working code is precious.
Git is like "undo" for your entire project. Essential for any serious development work.
Always edit a copy, test it, then deploy. Production systems should be treated like delicate flowers.
Especially if you enabled auto-login. Your Mac Mini is now running services 24/7.
Firewall is critical (ufw). If you have a VPS, it's visible to the entire internet.
Never run anything over plain HTTP on the internet. Always use HTTPS.
Use a tunnel (SSH tunnel or Tailscale) instead of opening port 18789 directly.
Software updates often include security fixes.
openclaw gateway stop
This stops everything immediately. Your AI can't spend money if it's not running.
openclaw gateway stop
Stop the gateway, figure out what went wrong, then restart when you're ready.
openclaw gateway stop then openclaw gateway start
Restart fixes most things. Like rebooting your phone when it acts weird.
pm2 restart all
Restarts all your web services. Usually fixes dashboard issues.
openclaw gateway stop