Open Office

A visible workspace for AI agents to collaborate as a single team.

Supports Claude, Codex, Gemini, Copilot, Cursor, Aider, OpenCode, Pi & Sapling — one team 🚀

Quick Start | Features | Architecture | Contributing

Features

• 8 AI backends, 23 roles — Claude, Codex, Gemini, Copilot, and more in one pipeline
• Team-based delivery — Leader coordinates planning, coding, review, and release
• Parallel collaboration — Worktree isolation with auto commit, merge, and undo
• Flexible code review — Mix Claude Code, Codex, or any tool you prefer
• Preview & feedback — Live preview, rating, and continuous learning
• Persistent memory — 4-layer memory across sessions and agents
• Telegram Control — Manage your agent team remotely
• Native desktop app — Tauri-based app with system notifications
• Token cost tracking — Real-time usage per agent and team

Quick Start

npx bit-office

Learn More

• Team workflow — how agents plan, execute, and deliver (view)
• Collaboration system — multi-agent orchestration and worktree isolation (view)
• Memory design — four-layer persistent memory architecture (view)

Run from Source

Prerequisites

• Node.js 18+
• pnpm
• At least one AI CLI installed (see Supported Backends)

Setup

git clone https://github.com/longyangxi/open-office.git
cd open-office
pnpm install
pnpm dev

Desktop App (Tauri)

Open Office also ships as a native macOS desktop app powered by Tauri.

Prerequisites

• Node.js 18+, pnpm
• Rust toolchain — install via: curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

Dev Mode

pnpm dev:desktop

Build Release

pnpm build:desktop

Produces Open Office.app and .dmg at:

apps/desktop/src-tauri/target/release/bundle/macos/Open Office.app
apps/desktop/src-tauri/target/release/bundle/dmg/Open Office_0.1.0_aarch64.dmg

Supported Backends

Open Office auto-detects installed AI CLIs at startup. Each backend has its own instruction file convention and capability set.

✅ = actively tested in production workflows. Other backends have not yet verified end-to-end.

Backends with ambiguous binary names (agent, pi, sp) use version-probe detection to avoid false positives.

Architecture

open-office/
├── apps/
│   ├── web/            # Next.js PWA + PixiJS pixel office + control UI
│   ├── gateway/        # Runtime daemon: events, channels, policy, orchestration
│   └── desktop/        # Tauri v2 native shell (macOS .app/.dmg)
└── packages/
    ├── memory/         # Four-layer persistent memory (L0–L3)
    ├── orchestrator/   # Multi-agent execution engine
    └── shared/         # Typed command/event contracts (Zod schemas)

Channels: WebSocket (always on), Ably (optional), Telegram (optional)

Tech Stack

• Frontend: Next.js 15, React, PixiJS v8, Zustand
• Desktop: Tauri v2 (Rust + system WebView)
• Backend: Node.js daemon, WebSocket
• Memory: Four-layer JSON store (session → agent → shared), Jaccard dedup
• Protocol: Zod-validated event schemas
• Integrations: Ably, Telegram, external process detection

Contributing

Issues and PRs are welcome. If you're exploring AI-native dev tooling, workflows, or interfaces, Open Office is a great playground for experiments.

Acknowledgments

Pixel office art inspired by pixel-agents by @pablodelucca.

License

MIT - feel free to use, modify, and distribute.

If Open Office helps your workflow, consider giving it a star!