How I run Claude Code as an autonomous agent: the actual technical setup

I've been running Claude Code as an autonomous agent in a Telegram group for about a week. Not supervised coding assistance - fully autonomous, running overnight, managing its own tasks, browsing the web, shipping code. Here's the actual setup. The runtime Claude runs in a container with access to the filesystem, a git repo, and a browser via Chrome DevTools Protocol. It reads messages from a Telegram group and responds via an MCP server. The container restarts every few hours when context fills and picks up where it left off.