Girru wants your AI team to start with the projects and memory you already have

Girru’s latest release tackles an awkward first step for agent teams: finding the work and context already on your computer. Version 1.4.1, published November 1 at 02:19 UTC, says its desktop setup discovers accessible projects, memory, skills, MCP connections and SSH identity metadata before creating a team.

The community project calls itself “Your Company of AI Employees.” Its more concrete proposition is a standalone workspace with agent roles, permissions, memory and a shared task board. Girru is the coordination application; Claude Code and Codex are separate tools it can use, while the selected models supply inference. It is not a new model or an extension that must run inside either coding tool.

Finding context before assigning work

The release describes a refreshed discovery view, clearer file labels and improved team editing. It also improves durable background memory imports, workspace reuse and discovery responsiveness. On Windows, it handles temporary file locks when saving discovery progress, and shutdown now waits for filesystem workers to finish.

That is a practical onboarding change, rather than evidence that any number of agents can reliably run an entire company. There is also a documentation boundary: the pinned README still marks broader imports of past chats and memory as in development. The release’s import fixes should not be read as a guarantee that every other AI application’s history can now be migrated.

Existing subscriptions, separate permissions

Girru says it works through locally installed Claude Code and Codex tools signed into the user’s own Claude or ChatGPT plan. Provider terms and usage limits still apply. Local models and cloud model connections are additional routes; calling the workspace local-first does not mean information sent to an external model stays on the machine.

Version 1.4.0, published October 31, introduced the Girru name for the existing Agent Player/Dpro project. Version 1.4.1 says updates preserve accounts, workspaces and local sessions. Its release page offers installers by operating system and CPU with SHA256SUMS manifests. Windows builds are unsigned; macOS builds are ad-hoc signed and not notarized. Some pinned README download links still point to 1.3.2, so the versioned release assets are the relevant download reference.

The project is MIT-licensed. Its feature descriptions come from the maintainers’ English release notes and README, not an AITrending installation or performance test. Discovering MCP connections or SSH identity metadata also warrants reviewing what an agent is allowed to access; discovery itself is not permission for every subsequent action.