YC open-sources qm, the multiplayer agent harness it runs on
MIT-licensed, self-hosted on Fly or AWS, it drives Claude Code, Codex or OpenCode and gives each teammate a sandboxed workspace — the stack YC runs itself on.
Copy markdown
A team agent harness you can actually run
qm gives every person and channel its own scoped memory, files, permissions, sandbox, and crons, so teammates work in parallel without stepping on each other — and can also collaborate with the agent in Slack and a web UI. It's MIT-licensed TypeScript that cleared 4.5k GitHub stars within days of launch.
Deploy it into your own cloud today
One command — `qm init . --org <slug> --target <fly-or-aws>`, run via `npm exec --package=@yc-software/qm` — scaffolds a deployment repo and walks you through infra, web sign-in, connector credentials, optional Slack, and live verification. Everything runs in your own Fly.io or AWS account, not YC's.
Bring your own harness and model
qm doesn't lock you to one agent: Pi, OpenCode, Codex, and Claude Code all drive the same core, and admins choose which harnesses and models are available org-wide. If you already live in Claude Code, you keep it and gain the multiplayer layer on top.
Three security postures, on by default
Set Strict (every tool call needs approval), Auto (the default, with classifier screening), or Dangerous (no screening) per org — but a predeclared command policy applies in every posture, Dangerous included. That's the dial that decides how much your agents can do while nobody's watching.
The catch: thin docs and a no-PR contribution rule
Early adopters flag a partly-generated README and sparse docs, so budget time for setup. Stranger still, qm asks contributors for human-written prose describing changes in markdown rather than pull requests — and ships an "anti-slop" taste skill to keep generated UIs from looking templated.