QM
A multiplayer agent harness that runs in Slack and on the web inside your own cloud, with your own models and keys, and whose own repository is maintained by the agent it builds.
GitHub avatar of yc-software, not the project’s own logo — taken from github.com on 2026-10-02.
What it is
QM is a multi-user agent harness that lives where the work already happens — Slack and the web — and runs in your own cloud against your own models and keys. It ships as an npm package under the @yc-software scope with Node 24.15 or newer, and the repository documents the pieces that make several people and several agents share one session: a model gateway, persistent subagent sessions, background ownership, swarms, sandboxes that survive, and a deploy directory.
Who built itThe repository’s 776 commits are led by Regan Bell (384) and Joshua France under two names (237 and 66), then Sina Matian (26), Eve Bouffard (18), Gilad K (6), Tom Firth (4) and a score of others. A large part of the record is not human at all: 120 of the 246 co-author trailers name QM itself, the agent the repository builds.
How it is put together
The parts · 5Everything is arranged around the session being shared. A model gateway sits between the workspace and whichever provider is configured; subagent sessions persist so a delegated task survives the turn that started it; ownership of background work is tracked explicitly so two people cannot silently collide in the same sandbox; and the deploy directory is documented as a first-class concept because the harness is meant to be run in an account you control. The web UI is a plugin rather than a separate application, which is why it accounts for the second-largest block of files.
- plugins/web-ui
- 406 files and 3.34 MB — the web surface, shipped as a plugin so the harness and its interface version together.
- test/
- 690 files and 7.75 MB, the largest part of the repository: an agent that several people share has to be tested as a system rather than as a library.
- docs/model-gateway.md and docs/persistent-subagent-sessions.md
- The two documents about holding several conversations at once — one model provider behind a single interface, and delegated sessions that outlive the turn that created them.
- docs/background-ownership.md and docs/sandbox-preservation.md
- What keeps concurrent work from colliding: explicit ownership of background tasks, and sandboxes that survive between turns instead of being rebuilt.
- docs/porter.md and docs/deploy-directory.md
- The deployment story, 29 KB each: the harness is meant to run in your own cloud, so the deploy directory is documented as a concept rather than a folder.
Choices, and what they beat
Run in the reader’s own cloud with their own models and keys over a hosted service with a shared key
It is the first line of the README after the one-line description, and the reason a model gateway and a deploy directory are documented as first-class parts of the design.
Live in Slack and on the web instead of in a terminal over a CLI-only agent
The product is described as multiplayer for work — several people in one session — and the 406-file web UI plugin is what that costs.
Persist subagent sessions and track ownership of background work over treating each delegation as a single turn
Both are documented at length because a shared workspace has to survive one person’s task outliving the message that started it, and two people working at once.
Read fromREADME, docs/ and the repository tree of yc-software/qm, read 2026-10-02.
Build log
4 stages- 01
The agent maintains its own repository
The repository was created on 2026-07-29 and has been pushed to every day since: 776 commits, of which 549 landed in September alone, across 2,355 files. The authorship is the unusual part. 246 commits carry a co-author trailer, and 120 of them name QM — the harness the repository builds — with Claude models accounting for most of the rest (Fable 5 on 27, Opus 5 on 9 and one more at 1M context, Fable 5.1 once, Claude four times), plus OpenAI four times, Codex twice and Cursor once. Human trailers name a dozen people. Regan Bell holds 384 commits and Joshua France 303 under two names.
- 02
Built for several people at once, and it shows in the file count
Of the 2,355 files, 690 are tests at 7.75 MB and 406 are a web UI plugin at 3.34 MB — the two largest parts of the repository by a wide margin. That shape follows from the claim: an agent harness meant to be shared by a team has to be tested as a system rather than as a library, and it needs an interface of its own rather than borrowing the terminal. The README’s headline is the deployment position rather than the feature list: run it in your own cloud, with your own models and keys.
- 03
The documentation is a list of the hard parts
The
docs/directory reads as an inventory of the problems this kind of tool runs into, one document each:porter.mdanddeploy-directory.mdat 29 KB apiece,swarms.mdat 16 KB,sandbox-preservation.mdat 12 KB,background-ownership.mdanddocument-inputs.mdat 9 KB,model-gateway.mdandpersistent-subagent-sessions.mdat 7 KB. A model gateway and persistent subagent sessions are what let one workspace hold several conversations at once; background ownership and sandbox preservation are what stop two people’s work from overwriting each other. - 04
Eight hundred open issues
The repository carries 808 open issues against 1,899 forks and 15,314 stars, and it is published as an npm package under the
@yc-softwarescope with a CI badge and an MIT licence. The issue count next to the commit rate — 549 commits in a single month — describes a project moving faster than it is closing, which is worth stating plainly rather than reading as either health or neglect.
Adjacent records
All records →No. 127
headcount
An agent organization shaped like a company — a chief executive over sixteen independently installable departments and 172 skills, where a skill is a folder of Markdown that loads itself when a request matches its description, one tree installs in both Claude Code and ChatGPT because only the manifests differ, and the 184 outside authorities that settle a question rather than decorate an answer — a regulator, a standards body, primary law — sit in a catalog beside the skills they answer for, each labeled with what an agent may do with it.
No. 122
Rome
A self-hosted agent runtime that treats the environment around a model as the thing worth growing: a Rome App packages an interface, executable actions, on-demand skills and a private database as git-tracked code an agent can reuse later, while the runtime gives every delegated subagent a child session of its own, forks a turn without letting it mutate its source, and fails a turn rather than quietly swapping the model that produced the conversation.
No. 111
OpenBot
An open-source platform from CopilotKit that gives each AI coworker a computer of its own — a container with Chromium, a workspace volume and its own logins — where a coworker is any endpoint speaking AG-UI, and every browser, file, MCP or shell action passes through one gateway that decides it against a CEL policy, writes an audit row and only then acts, or refuses and names the rule.