QM is a multiplayer agent harness designed specifically for work coordination within organizations, accessible both through Slack and a web interface. Built in TypeScript, it addresses a fundamental limitation of existing agent systems: most are designed as personal assistants and become complex when scaled to company-wide use. QM instead provides each employee with an isolated workspace where they can work independently while maintaining the ability to collaborate with the agent in shared channels, group messages, and projects.
The core architecture runs on Node.js with Fastify for HTTP, using Postgres for persistence of user data and session history. The system is model-agnostic and harness-agnostic by design, supporting multiple backends including Pi, OpenCode, Codex, and Claude Code without vendor lock-in. Every interaction flows through a central core that can swap between different models and harnesses through a single wiring file. The agent operates with a deliberately small, fixed tool surface, with one critical tool called execute that runs commands in each scope's isolated sandbox—a durable computing environment where installed tools persist across sessions.
QM implements granular scoping across multiple dimensions. Each person and each room maintains its own scoped memory, files, keychain view, permissions, crons, web applications, and durable sandbox. This isolation prevents one user's actions from affecting others while enabling collaborative work. The system supports personal customization alongside organizational governance, with administrators able to set org-level configuration, define security posture, and control which harnesses and models are available.
The feature set encompasses internal knowledge retrieval across notes, email, documents, databases, and the web; custom internal app development with targeted publication; automated inbox triage that learns from writing voice; direct repository integration for running tests, opening pull requests, and monitoring CI; and shared project tracking with automated updates. Skills are scope-owned and shareable through grants, with admin-gated promotion to the organization and the ability to import skill packs from git repositories. Background work runs through crons and watches that execute while no one is actively using the system.
Security follows the model of local coding agents, where the agent acts as the person it serves with their credentials and permissions, with all actions audited. Organizations choose one of three security postures: Strict mode pauses every harness tool call for human approval except turn enders; Auto mode uses a classifier to screen external data and tool results before they reach the model; and Dangerous mode removes content screening and pauses. A predeclared command policy applies across all postures, enforcing approval rules and hard denials for destructive operations.
Deployment is organization-owned through a deployment repository that depends on the qm package. The qm CLI validates and deploys configurations without requiring source checkout. Each deployment runs in the operator's own cloud account, and the system supports both a deployment-directory approach for minimal customization and a private fork approach where organizations maintain the entire codebase with organization-specific customizations isolated in deploy/layers directories.
Since GitGenius' previous check on August 4, 2026, the repository has grown from 11,064 to 11,082 stargazers and from 1,197 to 1,199 forks, while open issues decreased from 120 to 110. The most active contributors tracked include Rowerguy508, Mr-Neutr0n, and 16francej. Issue and pull request response latency averages 19.9 hours with a median of 14 hours across tracked items. The repository connects to related projects including openclaw/openclaw, nousresearch/hermes-agent, and nodejs/node through overlapping contributors.
