Y Combinator Open-Sources QM: An MIT-Licensed Multiplayer Agent Harness That Runs In Slack And The Web
Y Combinator open-sourced QM (quartermaster) under the MIT license as a multiplayer agent harness designed for organizational use across Slack and web interfaces QM provides isolated workspaces per person and per room, each with scoped memory, files, keychain, permissions, crons, and durable sandboxes The architecture is harness-agnostic, supporting Pi, OpenCode, Codex, and Claude Code as front-end agents driving a shared core Three security postures (Strict, Auto, Dangerous) control tool call g
Analysis
TL;DR
- Y Combinator open-sourced QM (quartermaster) under the MIT license as a multiplayer agent harness designed for organizational use across Slack and web interfaces
- QM provides isolated workspaces per person and per room, each with scoped memory, files, keychain, permissions, crons, and durable sandboxes
- The architecture is harness-agnostic, supporting Pi, OpenCode, Codex, and Claude Code as front-end agents driving a shared core
- Three security postures (Strict, Auto, Dangerous) control tool call gating, with hard denials for destructive commands enforced across all modes
- QM is positioned for startups and mid-sized companies (10-500 people) with infrastructure expertise, currently deployed internally across accounting, legal, events, and engineering
Why It Matters
QM represents a shift from personal AI assistants to organizational multi-agent systems, addressing the complexity of scaling agent deployments across teams. For AI practitioners, it offers a production-grade reference architecture for building scoped, isolated agent workspaces with configurable security postures. The harness-agnostic design also provides a practical model for avoiding vendor lock-in in enterprise AI deployments.
Technical Details
- Architecture: Central headless core built in TypeScript on Node.js with Fastify for HTTP, backed by Postgres for durable state (user data, session history). Optional plugins include a Slack integration (Bolt), web UI (Vite + Lit), admin panel, and public portal
- Isolation Model: Each user and room receives independent scoped memory, file storage, keychain views, permissions, scheduled crons, web apps, and sandboxed execution environments where installed tools persist across turns
- Security Postures: Strict mode pauses every tool call for human approval; Auto mode (default) screens provenance-labelled external data through a classifier before reaching the model; Dangerous mode removes screening and pauses but retains hard denials for recursive deletes and destructive SQL via a predeclared command policy
- Deployment: Organization-owned deployment repo depends on
@yc-software/qm, initialized withqm inittargeting Fly.io or AWS. Runs in the operator's own cloud account with optional external identity provider support - Agent Harness Support: Pi, OpenCode, Codex, and Claude Code all interface with the same core, enabling vendor-agnostic deployments
Industry Insight
- The multi-agent organizational harness model signals a maturation phase in enterprise AI, where the focus shifts from individual productivity tools to coordinated team-level automation with proper isolation and governance
- The three-tier security posture system (Strict/Auto/Dangerous) provides a practical template for organizations balancing automation speed against risk, particularly valuable for regulated industries like fintech and legal operations
- QM's harness-agnostic approach demonstrates that building a shared core with pluggable agent front-ends is a viable strategy for reducing vendor dependency, a pattern that may become standard for enterprise AI infrastructure as deployments scale beyond pilot projects
Disclaimer: The above content is generated by AI and is for reference only.