YC open-sources QM, a multiplayer agent harness for teams
MIT-licensed, cloud-first, and model-agnostic: run Claude Code, Codex, and OpenCode fleets across Slack and web. Plus: the full 2.78T Kimi K3 on a laptop.
Copy markdown
One deployment, a workspace per person
Every employee and every Slack or web room gets its own scoped memory, files, keychain view, permissions, crons, and durable sandbox, so agents run in parallel without colliding. YC says it runs its own accounting, legal, events, and engineering on QM.
Model-agnostic, no lock-in
QM is a harness for harnesses: point it at Claude Code, Codex, OpenCode, or Pi and switch per task. It's TypeScript on Node (Fastify, Slack Bolt, Lit), MIT-licensed, and cloud-first for Fly or AWS.
Stand up your own org this weekend
`qm init` (via `npm exec --package=@yc-software/qm`) scaffolds an org deployment with its own config and sandbox layer; you supply Node, Postgres, and cloud creds. Agents can open PRs, run tests, watch CI, read logs, and publish custom internal apps to the right people.
Agents that keep working while you sleep
Cron jobs and 'watches' let agents do background work outside live sessions, with security postures dialed from strict human-approval to automatic screening of untrusted data. Useful for on-call triage, digests, and CI babysitting.
Elsewhere: the full 2.78T Kimi K3, on a 64GB laptop
WASTE, a dependency-free C engine under Apache 2.0, keeps K3's 27GB trunk in RAM and streams its 982GB of 3-bit expert weights from NVMe, running the 2.78-trillion-param model at ~0.33 tok/s with a 29GB floor. It ships an OpenAI-compatible server.