Y Combinator has open-sourced QM, a multiplayer agent harness the accelerator built and deployed internally across its accounting, legal, events, and engineering teams. The repository is live on GitHub under the yc-software organization with an MIT license.
QM stands for quartermaster. According to YC’s project page, the name references the person on a ship who coordinates belowdecks to keep things in order. The metaphor maps directly to what the software does: manage identity, policy, scheduling, persistent state, tool access, and sandboxed execution across an organization’s agent fleet.
Why YC Built It
YC had already gone through several internal agent iterations before QM, according to Startup Fortune. The accelerator started with a basic agent loop in Ruby that accessed internal data, extended it with crons and webhook triggers, then provisioned more than 50 Hermes agents as personal assistants for individual employees. Managing a fleet at that scale became difficult, YC says, and OpenClaw pushed the team in a new direction.
The core problem QM addresses is multi-tenancy. A single assistant is straightforward. An assistant shared across a finance employee, a founder, and the engineer who built it creates immediate conflicts around context and credentials. QM isolates each person and each Slack room with scoped memory, files, keychain views, permissions, crons, web apps, and a durable sandbox.
Technical Stack
The README describes the architecture in concrete terms. The core runs TypeScript on Node and uses Fastify for HTTP. Postgres handles session history, memory, and queue state. Slack integration runs as an optional in-process plugin using Bolt. The web UI builds with Vite and renders with Lit.
Every turn routes through a central core that can use multiple models and harnesses to generate responses. One of the agent’s fixed tools is an execute command that runs in the scope’s isolated sandbox, where installed tools persist between sessions. The web UI, admin panel, and public portal are optional plugins over the core’s HTTP API.
QM is not tied to a single model or coding tool. According to the GitHub repository, Pi, OpenCode, Codex, and Claude Code all drive the same core. Every substrate (harness, session store, sandbox, memory) sits behind an interface, so production implementations swap in via a single wiring file. Company-specific configuration, custom tools, skills, and infrastructure live in a deployment directory that the QM CLI validates and deploys.
Security Model
The repository’s security section states that QM follows the pattern established by local coding agents: the agent acts as the person it is working for, using that person’s credentials and permissions, while actions are audited. QM offers configurable security postures ranging from strict human approval to what it explicitly labels a “dangerous” mode with no pauses between tool calls.
Naming the dangerous mode is a deliberate transparency choice, according to Startup Fortune. Admin controls let organizations set org-level configuration, choose which harnesses and models are available, and gate skill promotion across the organization.
The Agent Orchestration Gap
QM enters a crowded but immature market. Google Cloud expanded its Gemini Enterprise Agent Platform with seven-day runtimes and governance gateways in late July. Microsoft is building its own agent harnesses as alternatives to OpenAI and Anthropic. The common thread: organizations deploying multiple agents need orchestration infrastructure that none of the model providers fully offer.
YC’s release carries weight because it comes from a production environment, not a demo. The accelerator ran QM on its own operations before publishing the code. That operational lineage, combined with the MIT license, makes QM immediately forkable and commercially usable for any startup that wants to inspect what YC actually built rather than trust a pitch deck.
The repository is at github.com/yc-software/qm. Ten open issues and 37 pull requests suggest active development, not an archival dump.