I live on a high-desert ranch in Cochise County, Arizona — solar and a Powerwall, a well, cameras, horses, and a small fleet of machines that watch all of it. Last week that system ran about 17,000 scheduled jobs with a 2.7% failure rate, and the ranch covered 96% of its own electricity. I know those numbers because the system records every run it makes; distrusting green dashboards is rule one out here.
I want to be straight about the timeline: this system is young. I started it at the end of May — about seven weeks ago. What surprised me is how fast it compounded: roughly 80 small projects now, and every one of them stands on the same plain substrate that stabilized in the first couple of weeks — a secrets vault that fails closed, a manifest-driven scheduler, a run log in SQLite, a freshness check that catches jobs which quietly stop, and a few conventions for memory and skills. None of it is clever. All of it gets exercised 17,000 times a week.
Today I published it as AI-OS Seed — an alpha, MIT, free. It isn't an agent and it isn't a framework; it's the floor an agent stands on. A build pipeline generates the repo straight from my live tree, scrubbing and auditing every file — so what you clone is what actually runs here, not documentation about it.
If you already run AI-OS Core, the assistant recipe I gave away in June, the seed is its upgrade, not a second system. The install composes into the workspace you have; your memory and files stay put, and your assistant gains an operations floor. It can run jobs on a schedule while you're away, keep a record of every run, notice when something goes quiet, and answer "how is my system doing" in one honest screen. I did exactly that on my own Mac this afternoon — memory intact, first job green on launchd.
Fair warning: it's an alpha, and the README says exactly what I've verified and what I haven't. I've run the install myself on both a Linux box and a real Mac, and CI repeats the whole thing on every push. My own installs caught three real problems the first day; each was fixed and shipped within hours, and the README says what they were. That loop is the argument for shipping this early rather than polishing it in private. The install itself is a five-line paste into Claude Code, and your agent does the rest — showing you every command first.
If you're an operator with a machine that stays on and something real to watch, take a look: https://github.com/cvp1/ai-os-seed. Tell me what breaks — the first stranger to get a green heartbeat is the whole point of shipping this early.