From Broken Docker Containers to a Working AI Agent: The Full OpenClaw Journey

Every blocker, every fix, and why bare metal is the sweet spot for a personal AI agent setup. TL;DR: I spent several weeks self-hosting OpenClaw on a Hetzner VPS. The setup involved more blockers than I expected — a 3-day crash loop caused by a model hallucinating its own config, a Docker isolation wall that prevented the agent from using any tools I didn't bake into the image at build time, and a browser control problem I had completely misunderstood. I migrated to bare metal. Everything works now. Here's the full story with every exact fix. What is OpenClaw and why did I want it? OpenClaw is an open-source AI agent gateway. You connect it to Telegram (or a web chat), point it at a model, and you get a personal AI agent that can browse the web, read files, run code, and respond to messages. The selling point is control: you pick the model, you own the server, you pay for it if you want or run it for free with community models. I wanted a personal AI agent that could take real actions