Open-source desktop platform for multi-agent AI collaboration
Give AI agents local filesystem access, shell execution, and a rich toolbox — all on your machine.
中文文档 • Why • Features • Architecture • Quick Start • Docs
Most AI chat interfaces are isolated from your actual work environment. You spend half the time copy-pasting code, file contents, and terminal output between windows.
OpenCowork puts the agent on your machine:
- Direct filesystem access — Agents read, write, and edit files in your project with your approval.
- Shell execution — Run commands, check logs, and manage dev servers without leaving the conversation.
- Full context awareness — Agents explore your codebase on their own. No manual context feeding.
- Human-in-the-loop — Transparent tool-call approval keeps you in control at every step.
- 4-layer Electron architecture — Main process, Preload bridge, Renderer UI (React 19), and provider-agnostic agent runtime.
- TypeScript end-to-end — Type safety from SQLite through IPC to the UI.
- SSH remote support — Agents operate on remote hosts transparently via SSH with xterm.js terminal integration.
Every conversation picks the right mode:
| Mode | Purpose |
|---|---|
clarify |
Ask grounded questions, resolve ambiguity, produce a reviewable plan before any code is written. |
cowork |
Full agent: code search, file I/O, shell, browser, sub-agent delegation, and more. |
code |
Pair programming — focused code generation and surgical editing with Monaco Editor integration. |
acp |
Architecture-control lead: clarify, design, decompose, and delegate implementation to sub-agents. |
- File & Shell — Read, Write, Edit, Glob, Grep, Bash (local and SSH).
- Browser — Built-in webview with navigate, snapshot, click, type, and content extraction.
- Task & Team — Decompose work with TaskCreate/TaskUpdate, spawn parallel sub-agents via Task, and orchestrate Agent Teams with TeamCreate/SendMessage/TeamStatus.
- Plan Mode — EnterPlanMode → write plan → ExitPlanMode for structured, reviewable implementation plans.
- Goal Tracking — Create, track, and complete session-level goals with token budgets.
- Memory System — Layered memory: global SOUL.md / USER.md / MEMORY.md and per-project .agents/ overrides.
- Cron Agent — Schedule recurring or one-shot background agent tasks with multi-channel delivery.
- MCP Client — Connect to Model Context Protocol servers (stdio, SSE, streamable-HTTP) and expose active MCP tools directly to the agent.
- Skill System — Install domain-specific skills from the Skills Market; loaded dynamically and surfaced to the agent at runtime.
- Custom Extensions — Build plugins with declarative HTTP tools, sandboxed JS handlers, and custom HTML renderers.
| Platform | Support |
|---|---|
| Feishu / Lark | ✅ |
| DingTalk | ✅ |
| Discord | ✅ |
| ✅ | |
| Telegram | ✅ |
| WeCom (WeChat Work) | ✅ |
| WeChat Official | ✅ |
| ✅ |
- SQLite — Messages, sessions, projects, tasks, and plans survive restarts.
- Additive schema — Columns are added when absent; no migration files, no data loss.
13+ languages including English, Chinese, Vietnamese, Turkish, and more — all via i18next.
Renderer (React 19) ←→ Preload (contextBridge) ←→ Main Process
│ │
Agent Loop ─ Tool Registry ─ IPC ──────────→ IPC Handlers
│ │
├─ File I/O, Grep/Glob, Bash SQLite (better-sqlite3)
├─ Browser (webview) Shell / SSH (node-pty, ssh2)
├─ Sub-Agents & Teams Messaging Plugins
├─ Plan, Goal, Memory MCP Client
├─ Skills & Extensions Cron Scheduler
└─ MCP Resources File System
- Renderer — React 19 + Tailwind CSS + Zustand stores. Agent loop runs here, tools execute via IPC.
- Preload — Narrow
contextBridgeAPI for secure main↔renderer communication. - Main Process — System access: SQLite, filesystem, shell, SSH, messaging plugins, cron, MCP client.
- Agent Runtime — Provider-agnostic (
js-agent-runtime.ts), streams responses, handles tool calls.
Prerequisites: Node.js ≥ 18, npm ≥ 9
git clone https://github.com/AIDotNet/OpenCowork.git
cd OpenCowork
npm install
npm run dev| Command | Description |
|---|---|
npm run dev |
Start Electron + Vite with hot reload |
npm run build |
Typecheck then build for production |
npm run build:win |
Build Windows installer |
npm run build:mac |
Build macOS .dmg/zip |
npm run build:linux |
Build Linux .AppImage/.deb |
npm run lint |
ESLint with cache |
npm run typecheck |
TypeScript check (main + renderer) |
npm run format |
Prettier auto-format |
Data directory:
~/.open-cowork/— SQLite database, config, agents, skills, commands, and prompts.
- Autonomous coding — Agents refactor, debug, and write code directly in your workspace.
- Scheduled ops — Cron agents monitor logs or system health and report to Feishu / DingTalk / Slack.
- Data research — Scrape web pages, process CSVs, generate reports with charts.
- Remote management — Operate on remote servers via SSH without leaving the app.
Full documentation at open-cowork.dev — built with Fumadocs + Next.js.
We welcome contributions! See AGENTS.md for the development guide, coding conventions, and commit message format.
RoutinAI — Enterprise-grade unified LLM API gateway providing a single, type-safe interface to 100+ models across GPT, Claude, and Gemini families.
GeneralUpdate — Cross-platform auto-update component for .NET applications.
⭐ If this project helps you, please give it a star.
Made with ❤️ by the AIDotNet Team