YC open-sources QM to bring multiplayer AI agents into workplace

QM is an open-source multiplayer agent harness built specifically for team collaboration. Instead of treating AI as an isolated single-player chatbot, it turns intelligent agents into shared, company-wide infrastructure.

Most modern agent tools behave like personal assistants. Bringing one into a company setting usually leads to chaos around shared context and leaking credentials. QM tackles this with Scopes: a single central engine powering isolated operational rooms. Every employee receives a private workspace with scoped memory, dedicated file storage, credentials, scheduled tasks, and a durable sandbox. At the same time, coworkers can collaborate with the very same agent in Slack channels, group chats, and cross-functional projects. The agent dynamically switches identities, contexts, and access levels depending on which room it enters.

What it does

QM works across both personal workflows and team-wide operations through Slack and a dedicated web app.

On the personal side, the agent can study your writing style from past messages and triage your inbox on a schedule, drafting replies and applying labels automatically. It can also operate as a background developer inside existing repositories, running test suites, opening pull requests, monitoring CI pipelines, and inspecting error logs in its private sandbox.

On the multiplayer side, teammates can drop the agent into a shared Slack channel to track project milestones and publish automated progress updates. When an ad-hoc operational tool is needed, teams can have QM generate internal web applications and publish them directly to authorized colleagues.

Skills inside QM are scope-aware. An individual workflow can be shared with a specific teammate, vetted by an admin for company-wide deployment, or imported directly from a public Git repository. Background crons, file watchers, and inbound webhooks ensure scheduled jobs execute even when everyone is offline.

How the architecture works

QM runs a headless TypeScript core built on Node and Fastify. The internal agent loop uses Pi, allowing teams to route tasks across different model providers without vendor lock-in. A Postgres database serves as the persistence layer for sessions, memory routing, and background queues.

Execution takes place inside durable per-scope sandboxes. Unlike ephemeral containers that vanish after a single run, QM keeps installed tools and system dependencies intact inside each room's sandbox. The web frontend is built with Lit and Vite, while Slack integration runs via an in-process Bolt plugin.

Security follows the design of local coding agents: the agent acts strictly on behalf of the person invoking it, inheriting their credentials while logging every action. Administrators can set global security postures. Strict mode halts every single tool call for manual approval. The default Auto mode automatically blocks private network access. A Dangerous mode bypasses approval prompts while preserving hard guardrails that block destructive shell commands or recursive file deletions. Data sharing defaults to Isolated, meaning personal memory and files never bleed into shared channels unless explicitly opted in.

How to deploy it

QM is distributed under the MIT License and designed to run entirely within your own cloud account using your own API keys.

Getting started requires Node.js and a running Postgres instance. The project provides an interactive initialization flow via npm, allowing operators to target deployments to Fly.io or AWS while configuring authentication, Slack apps, and connector credentials. Developers can also spin up local instances to test the web interface, Slack bot, or both surfaces together.

Teams can either pin the published npm package and manage configurations in a separate directory, or maintain a private source fork to modify the core engine, CLI, or sandbox images directly. Third-party hosted versions are also supported for teams that prefer not to manage infrastructure.

Why it matters

Most enterprise AI setups today either wrap a single-user chatbot in single sign-on or sell an expensive, closed-source platform that locks your data in. QM takes a pragmatic engineering approach: it uses straightforward Unix-style sandboxes and explicit scopes to separate personal privacy from collaborative work.

In our view, QM is well worth testing for technical startups and software teams looking to deploy agents across their workflow without leaking sensitive internal data. Its provider-agnostic model routing and durable execution environments make it far more capable than standard conversational bots. The trade-off is operational overhead: running QM requires real infrastructure management and database administration, making it ideal for developer-led teams rather than non-technical organizations.