QM review: self-hosted team workspaces with Claude Code

QM is an open source AI agent harness that the speaker describes as a web application for companies to deploy on their own infrastructure. Users access it through Slack or a browser. In the setup reviewed, agent execution happens in cloud workspaces rather than on a user's laptop, with Postgres storing session history.

Each workspace has an isolated container and persistent storage. Installed environments, files and cloned repositories remain available between sessions. Projects let colleagues share memory, apps, connectors and credentials, while personal spaces keep their own context. The speaker criticizes the web interface for failing to show the active project consistently across tabs.

The review also explains how to use Claude Code as a coding assistant inside QM. Pi is the default harness in the demonstrated deployment. Adding Claude Code required a PUT request to the harnesses endpoint because the admin panel lacked a control for that setting. Administrators decide which harnesses users can select.

QM exposes seven tools and uses shell commands for other work inside its sandbox. The speaker describes strict, auto and dangerous approval modes: strict pauses most calls, auto screens external data and tool results, and dangerous removes those checks. Teams can tighten the organization policy but cannot loosen it. A separate command policy still blocks operations such as recursive deletes and destructive SQL. The review finds shared context useful for teams but calls the UI immature.