The problem we kept hitting in client work: every agent platform wanted to be the place your documents live. We wanted the opposite — the orchestrator reads your files in place, on your disk, and we never see them.
How it works: you describe work in plain English, Bees proposes the agents, stages, tools and schedule, and nothing runs until you apply it. Agents execute locally in their own run folders. A separate reviewer agent checks output before it advances a stage. Agents stop and ask for approval before external side effects like sending mail.
Coordination is deliberately thin. In a Regular org we sync team membership, process runs and stages, work claims, and file names with relative paths. Not contents. If you want no third party at all, pick a local model and local tools and it runs fully offline.
Model routing is BYO: local models, your own API keys, or the official Claude Code CLI you already have installed and signed into as yourself. We don't intermediate credentials and we add no per-token or per-run fee — that's the whole point.
Rough edges, honestly: {name the two weakest things — I'd lead with whatever your beta testers actually complained about}. Scripted installers exist but Linux distro coverage is the thinnest. Very happy to answer architecture questions, and I'd particularly like pushback from anyone running multi-machine setups.