QM
Multiplayer agent harness for work
About QM
QM is a multiplayer agent harness, short for quartermaster, that Y Combinator built for internal use and then open-sourced. It is the system YC says it runs its own accounting, legal, events and engineering work on, including the ongoing development of QM itself. The distinction it draws is between a personal assistant and a company tool: most harnesses are designed for one person, and you can stretch one across an organisation, but the administration gets complicated quickly. QM starts from the organisation instead. Every employee and every project gets an agent as needed, and every one of them is scoped - its own memory, files, keychain view, permissions, crons, web apps and durable sandbox.
The design has a documented history rather than a launch narrative, which is unusually useful when judging it. YC's own write-up describes two earlier attempts: a basic agent loop in Ruby with access to internal data, which people got value from on day one but which was limited in scope, and then more than fifty individually provisioned Hermes agents, which turned out to be hard to manage at even that size. QM is the attempt to keep the flexibility of the second while recovering the simplicity of the first, and the requirement that it be self-hosted was deliberate. It is MIT licensed with no hosted tier, so your costs are your own infrastructure and your own model bills. It is also not model-locked: the same core drives several harnesses including Pi, OpenCode, Codex and Claude Code. Set expectations accordingly on maturity - the repository is only weeks old and carries several hundred open issues.
You do not check out the source. Deployment starts by pointing a coding agent at qm init, which walks through infrastructure, credentials, Slack integration and deployment configuration and materialises the setup for either Fly or AWS in your own account. Administrators can reach a new deployment before email is configured using temporary admin login links. Once it is up, the two surfaces are Slack, where QM runs as an in-process plugin, and a web interface, both backed by a headless Node and Fastify core on Postgres with optional admin panels and public portals extending the API surface. Work happens in personal scopes for individual employees and shared scopes for channels and projects, with identity unified across both. Background work runs without anyone present through crons, watches and webhooks. Execution safety is a configurable posture rather than a fixed policy, with strict, auto (the default) and dangerous levels changing how much gets approved before it runs, and the repository ships a security document covering the threat model and known limitations.
- •Personal and Shared Scopes - Each person and each room gets its own memory, files, keychain view, permissions, crons, web apps and durable sandbox, with one identity across both
- •Slack and Web - An in-process Slack plugin plus a web interface over the same headless core, so the agent is where the work already happens
- •Admin Control - Organisation-level configuration and security settings, with three approval postures (strict, auto, dangerous) governing what runs unattended
- •Background Work - Crons, watches and webhooks for scheduled and event-triggered jobs that do not need anyone in the conversation
- •Shared Skills With Granular Permissions - Capabilities published across the organisation without handing everyone the same access
- •Self-Hosted and Model Neutral - qm init deploys into your own Fly or AWS account, and the same core drives Pi, OpenCode, Codex and Claude Code
Small companies and startup teams that want a fleet of agents with real administration rather than fifty individually managed personal assistants, and that are willing to run infrastructure to get it. The scoping and permissions model is the reason to choose it over a personal harness, and the fact that YC runs its own back office on it is a stronger quality signal than most launch claims, because it is a system its authors depend on daily. Teams with a data residency or vendor requirement benefit from it being self-hosted with no hosted option to opt out of. It is a poor fit for a solo developer, who will get more from a personal harness with none of the deployment work, and for anyone who wants to try something in five minutes, because getting to a running instance means a cloud account, a Slack app and credentials. The repository's age and open issue count mean early-adopter tolerance is part of the price.
Pricing
No pricing page found
Free and MIT licensed with no hosted tier. qm init deploys into your own Fly or AWS account, so the running cost is your own infrastructure plus whatever models you point it at.
From the vendor pricing page, 2026-10-08
