Run an agent office on your local machine using existing AI subscriptions
This tool consolidates terminal-based coding AIs into a single office environment, allowing work to continue while you are away. Each agent has its own memory and mailbox, and a coordinator named Michael assigns tasks and displays progress on a 2D office screen. It operates within the hourly limits of AI services you already pay for, enabling you to use multiple AIs simultaneously without additional costs.
New features in version 0.5.3
The recently released version 0.5.3 adds local voice recognition. On macOS, pressing the Option key, or Control+Alt+Space on Windows and Linux, allows you to speak into any application's text input field. It also includes a feature to record meeting content and convert it to text by distinguishing between 'me' and 'the other party.' Additionally, a single computer can now open multiple offices with different folders and agents simultaneously.
How terminal agent teams are structured
The tool runs various terminal-based AIs, such as Claude Code, Codex, and Gemini CLI, as actual processes. Each agent operates in an independent terminal session and exchanges messages with others via local git repositories. Michael, referred to as the 'GOD agent,' reads all requests, handles routine tasks independently, and asks the user for approval only when human judgment is required, such as for expenses or risky operations. Agents share memory through markdown files and retain previous learning content even after sessions are terminated.
Office screen and task management
Agents appear as avatars on a 2D office screen, where you can see them move to their seats when working and envelopes flying when they send messages to each other. Each agent is assigned a ticket number to track work history, and a kanban board manages task dependencies and schedules. The built-in Monaco IDE allows for file editing and viewing git history, while git worktree support enables agents to work in parallel without conflicts.
Installation and requirements
Installers are provided for macOS, Windows, and Linux, with macOS builds being signed and notarized. Building from source requires Node.js 18 or higher, npm, and a C/C++ toolchain for the node-pty native addon. Users must install at least one supported agent CLI, and the in-app onboarding wizard guides the automatic installation of any missing tools. Using local LLMs requires configuration for Ollama, LM Studio, vLLM, or similar services.
Pre-use considerations
This tool is pre-release software with many features and rapid changes. Official builds send anonymous usage events but do not transmit prompt or code content, and this can be disabled in settings. The source code is under the MIT license, but included pixel art follows a separate license. Since agents perform tasks autonomously, it is recommended to set approval gates for expenses or destructive operations and specify budget limits.