Y Combinator has open-sourced QM, an internal tool the startup accelerator built to run AI agents across its own company — from accounting and legal work to events and engineering — according to a post from YC and the project’s own site, qm.ycombinator.com. The software is released under the permissive MIT license, with its source code available on GitHub.
What QM does
QM is what YC calls a “multiplayer agent harness”: rather than giving each employee a single personal chatbot, it gives every worker an isolated workspace — with its own memory, files, permissions, and sandbox — while letting teams share AI agents and automations through Slack and a web interface. Shared “skills” can be granted to other employees and promoted company-wide by an administrator, and agents can be triggered on a schedule or by webhook rather than only on demand.
The system is not tied to one model vendor: it supports several coding and agent harnesses, including Claude Code, OpenAI’s Codex, and the open-source OpenCode, so organizations can swap the underlying model without rebuilding their workflows. It also ships with three graduated security postures — from requiring human approval on every tool call to a more autonomous default mode — that govern how freely an agent can act on external data before a person signs off.
Built from internal use
According to YC, QM grew out of an earlier, Ruby-based agent loop the company used internally, refined after it ran more than 50 agents as personal assistants for staff. The accelerator says it now uses QM to build QM itself, alongside day-to-day accounting, legal, and event-planning tasks. The GitHub repository drew roughly 5,000 stars within days of release.
Why it matters
The release adds YC to a small but growing list of organizations publishing the agent infrastructure they use internally instead of keeping it proprietary, giving other companies a free, self-hostable alternative to commercial “agent platform” products for coordinating AI agents at scale.