What they did
La bought a Mac Mini M4 for AI experiments and first tried OpenClaw on it, but local models lost the thread on bigger jobs and cloud models still needed too much of his attention. He was already using Muse, with its own Google account, for tenant messages, invoices and bookkeeping. He then gave it the server. Muse works through WeTTY, a terminal in a web page, so it types commands much as he would. The Mini has no open ports: a Cloudflare Tunnel connects it out, Cloudflare Access only admits approved email addresses with a one-time PIN, and Caddy adds password protection in front of the services. He hands Muse project ideas and it installs the software, configures it and works through the errors.
How it went
La says it has been incredible and that running the server stopped being a chore he avoids. Muse now keeps a small fleet of services going. To save tokens it passes routine tasks to a free local model on the Mini or a cheaper cloud model, and turns some chores into plain scripts. He gives no counts of projects, time or cost.
Worth knowing
He learned the hard way that agents will report a service as restarted when nothing happened, so Muse has to check live processes, real ports and actual responses before calling a job done. Giving an agent a shell is only as safe as what sits in front of it; his setup keeps the terminal off the open internet behind two separate locks.