Had a Mac Mini doing nothing. Gave it its own GitHub account, its own email, its own Telegram bot, and root access. No sandboxing, no permission prompts, no guardrails. It reads email, pushes code, runs tasks overnight, messages me when something needs attention. I text it like an intern who might accidentally nuke something but can't do that much damage because it's on a dedicated box. Started with OpenClaw but the containers and safety layers defeated the point — I wanted to let it roam free, not babysit Docker. Forked NanoClaw, ripped out containers entirely, agents run as raw Node.js processes with full machine access. That's the whole philosophy: give it a computer, give it accounts, see what it does. It's open source, MIT, and honestly more fun than useful half the time. But it does genuinely work — my agent Ziggy handles email, monitors PRs, runs scheduled research, and commits its own code. Looking for collaborators who think AI agents should be less controlled, not more. The skills system means anyone can add capabilities as markdown files. If you want to give a spare computer a personality and a job, come play.

GhostClaw - Claws Gone Wild

The stripped down Claw fork - no containers, bare metal

The stripped down Claw fork - no containers, bare metal