baton treats AI agents as a workforce rather than chat sessions. It gives them a ticketing system, a message bus, a shared wiki and a night shift, from one CLI, on macOS, Linux and Windows.
The core is deliberately small. baton writes state; the batond daemon makes state happen. Work is filed with acceptance criteria, claimed under a cross-node mutex, and closed with evidence. For work that must not stop until it is provably done, there is the judge loop: plan draft turns a research document into phases, each carrying only the knowledge that phase needs, and a goal then runs until an independent judge — a fresh model that did none of the work — agrees the criteria are met, checked against the git diff rather than the agent's word for it. workflow run fans one brief out across many repos with bounded concurrency and per-target, git-measured evidence. A role library defines who an agent is; launch --provider decides which engine embodies it — claude, glm, qwen, deepseek or vertex — so a second opinion is a flag, not a rewrite. Cron entries compile to launchd, systemd timers or Task Scheduler, and every fire lands back in the same auditable ledger.
Machines federate over an SSH file-drop wire. No server, no accounts; a fleet is however many machines you point at each other, and all of it is plain markdown and SQLite under ~/.baton.
Past the core sit some two dozen modules — on-device Apple AI, a CDP browser driver, CVE and site-health audits, a video suite, five messaging bridges — each an instrument you install only if the work asks for it. Every download is sha256-pinned and ed25519-signed, and the installer proves the bytes came from us before one is unpacked.