The OpenClaw Problem, or: How I Stopped Wrapping Docker in VMs and Built a One-Command Setup Instead

A continuation of 128GB of RAM, Zero Internet, and a Year of Building AI Infrastructure Nobody Asked For When I shipped headless agent mode for Cloister, I thought I'd solved the hard problem. VM isolation, consent policies, credential forwarding — all the security layers you'd want between an autonomous AI agent and your actual life. The next step was obvious: run OpenClaw inside it. OpenClaw is an open-source AI agent framework — a persistent assistant that connects to Telegram, WhatsApp, iMessage, and runs tasks on your behalf. It has a gateway (the brain), nodes (execution endpoints), and a sandbox for running arbitrary code. It's powerful. It's also the kind of software whose community has been refreshingly honest about: don't run it on bare metal . So I didn't. I ran it inside a Cloister VM. Inside Docker. Inside a Linux VM. On macOS. If that sentence made you wince, you're ahead of where I was. Three Security Bypasses and a Caddy Proxy The first sign something was architecturall