OpenClaw AI Agent

The idea

The interesting part was not just what an agent could do, but what changes when it is trusted with powerful, real-world tools.

The build

I ran an AI agent on a VPS with access to web, email and productivity services, while thinking carefully about the boundaries around that access.

Problems & investigation

• What permissions an autonomous agent should have
• The risks of joining powerful tools together
• Restricting SSH through Tailscale rather than exposing it publicly

What I learned

• Useful automation needs deliberate boundaries
• Security is part of the design, not an afterthought
• Tool access is a meaningful responsibility

What's next

The project is archived because running costs became too expensive, but the security questions it raised continue to shape future experiments.