YC Open-Sources qm, a Multiplayer Agent Harness for Slack and Web
TL;DR
- Y Combinator open-sourced qm under an MIT license, framing it as a multiplayer agent harness for companies rather than a personal assistant.
- Each person and each room gets scoped memory, files, keychain view, permissions, crons, web apps, and a durable sandbox.
- The same core drives Pi, OpenCode, Codex, and Claude Code, so deployments are not locked to a single model vendor.
Y Combinator open-sourced qm this week, and the framing on the GitHub page is what makes it interesting. 'Most agents are designed like personal assistants,' the README says, 'QM is designed for startups.' The pitch is a multiplayer agent harness for work, running in Slack and on the web, with the same identity following a user between the two surfaces.
The technical shape is straightforward. Each person and each room, per the README, gets its own scoped memory, files, keychain view, permissions, crons, web apps, and durable sandbox. The core runs TypeScript on Node with Fastify, Postgres holds session history and durable state, the Slack plugin uses Bolt, and the web UI is built with Vite and Lit. Crucially, qm is not tied to one model vendor: the README says Pi, OpenCode, Codex, and Claude Code 'all drive the same core,' so a deployment can swap harnesses without rebuilding the scaffolding around them.
Why this matters for anyone picking internal AI tooling right now is the shape of the primitive. The industry default for the last two years has been per-user copilots bolted onto individual products. qm is arguing the useful unit is a shared org-wide harness where each employee has an isolated workspace but skills are scope-owned and shareable by grant, with admin-gated promotion across the company. YC's own post on X says the firm uses qm 'across accounting, legal, events, and engineering (including building QM itself!).'
Security is handled through three postures the README calls Strict, Auto, and Dangerous. Auto is the default and puts a classifier in front of external data before it reaches the model, Strict requires human approval on every tool call, and Dangerous turns the screening off. Everything the agent does is audited, and credentials are scoped per person.
The honest caveat is that the README is a pitch, not an operations manual. It does not tell you how sandbox isolation is enforced under the hood, what the ongoing maintenance burden looks like for a small ops team, or how you should decide which harness handles which scope when four are on offer. Take the '1000x' language from YC staff on X as marketing rather than measured claim. But the piece worth watching is the shape: a cloud-first, MIT-licensed, model-agnostic team surface, released by an org whose recommendation carries weight with the founders it funds.
Originally reported by github.com
Read the original article →Original headline: YC-Backed 'qm' Open-Sources a Multiplayer Agent Harness Spanning Slack and Web With Per-User Scopes