§ Claim under review · Product
"Y Combinator just open-sourced one of its internal AI tools... QM... a cloud-first, multi-agent workspace designed for entire teams... Runs multiple AI agents from one shared workspace; Native Slack and web interface; Shared memory, files, and company knowledge; Automatic workflows with cron jobs and webhooks; MIT licensed and completely open source... It's the same system Y Combinator has been using internally before deciding to release it to everyone."
Verdict
Mostly accurate
Confidence
HighSummary
This one is largely true. Y Combinator did publicly release an internal multi-agent tool called QM on July 31, 2026. It is MIT licensed, available on GitHub, runs in the cloud, works through Slack and a web app, and supports scheduled jobs, webhooks, memory and shared files. YC confirms it uses the tool internally for accounting, legal, events and engineering. Two things in the post are off. First, YC itself called the release an experiment that is early and has bugs, which the post removes and replaces with claims about enterprise readiness. Second, the post says agents run from one shared workspace, when the documented design actually gives each employee and each chat room its own isolated workspace with separate memory and permissions, with collaboration added on top. Also worth noting: you do not need to comment anything to get the link, since the project is openly published on GitHub.
The readings
key figures from the evidenceGitHub stars on the yc-software/qm repository
date of YC's QM announcement
Why this verdict
Evidence
The project exists and the core of the claim checks out against YC's own words. YC stated it decided to open-source a multi-agent harness it uses internally, called "QM," meant to be easy to customize like Hermes or OpenClaw but useful for a whole company, used across accounting, legal, events, and engineering, including building QM itself, with the whole project under an MIT license, cloud-first, and with Slack and web UI natively. YC also listed features including triggers (crons, webhooks), memory, shared files, connectors for a company brain, agent browser support, shareable web app artifacts, and multi-player projects.
The repository is real and public. The yc-software GitHub organization lists "qm" as a public TypeScript repository described as "Multiplayer agent harness for work," carrying an MIT tag, with roughly 1,173 stars at the time of indexing. The README states that, except where otherwise noted, QM is available under the MIT License.
On architecture, the repo describes a design that is more nuanced than "one shared workspace." The README says most agents are designed like personal assistants and that QM is designed for startups, where employees each get their own isolated workspace and work independently without affecting each other, while also collaborating with the agent in channels, group messages, and projects, and each person and each room has its own scoped memory, files, keychain view, permissions, crons, web apps, and durable sandbox. It is built to be model and harness agnostic, with Pi, OpenCode, Codex, and Claude Code all driving the same core so a deployment is not tied to a single vendor, and the same identity and configuration carries between Slack and the web app.
On internal provenance, YC's project page confirms real internal use and a history of iteration. YC describes QM as an open-source agent harness giving every employee and project an OpenClaw-like agent, built from YC's experience running 50+ agents internally. YC says its first system was a basic agent loop in Ruby with tools accessing internal data, later extended to support crons and webhook triggers, and that it then provisioned over 50 Hermes agents for individual employees as personal assistants before deciding it wanted something as flexible as Hermes with the simplicity of the original system, and something it could own and host itself.
Maturity framing is where the post and the source diverge most. YC explicitly said that, like many things it does, it is an experiment, that it is early and has bugs.
Findings
✓ What's accurate 7
- Y Combinator did open-source an internal multi-agent tool called QM.
- It is MIT licensed and publicly available on GitHub.
- It is cloud-first with native Slack and web interfaces.
- It supports cron jobs and webhook triggers, memory, and shared files.
- YC uses it internally across accounting/finance, legal, events, and engineering.
- It is genuinely designed for company-wide use rather than a single individual.
- The framing that it differs from single-user, laptop-bound agent setups matches YC's own positioning.
≈ What's misleading 6
- Omitted qualifier: YC explicitly called the release an experiment that is early and has bugs. The post removes this entirely and replaces it with "This isn't another side project" and "where enterprise AI is heading." This is the single largest gap between claim and source.
- Architectural inversion: "Runs multiple AI agents from one shared workspace" reverses the documented design, which gives each person and each room its own isolated, scoped workspace, with collaboration layered on top through channels and projects.
- Oversimplification: "Shared memory, files, and company knowledge" flattens a scoped permission model into a single shared pool. Shared files and company connectors exist, but memory and files are scoped per person and per room by design, which is a security feature the post erases.
- Marketing as description: "a team of AI employees" is the post's own metaphor. YC's own term is "harness," a tool layer that runs agents driven by third-party models.
- Minor license overstatement: "completely open source" is close but the repo qualifies MIT coverage with "except where otherwise noted."
- Engagement mechanic: the post gates a "free link" behind a comment, when the project is openly published at github.com/yc-software/qm and qm.ycombinator.com. Nothing needs to be unlocked.
? What's uncertain 3
- Whether any specific component of the repo falls outside the MIT license. The README's "except where otherwise noted" wording implies exceptions may exist, but I could not enumerate them from available extracts.
- Whether the released code is byte-for-byte the internal deployment. YC's docs describe org-specific configuration living in a separate deployment layer, which suggests YC's internal instance includes private configuration not in the public repo.
- Long-term maintenance intent. YC framed it as an experiment and did not commit to a support model in the material found.
Sources
6 of 6 linked to recordsY Combinator official post on X announcing the release
qm.ycombinator.com, YC's official project page
github.com/yc-software/qm README, license and docs
github.com/yc-software organization page (repo metadata, license tag, stars)
StartupHub.ai writeup of the QM release
KuCoin news flash (relaying MarsBit)