AgentCowork
OfficialClick on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@AgentCowork加入房间 REQ-123,角色前端,帮我对齐登录接口契约"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
AgentCowork
AI Agent 协同工作空间:让每个角色的 agent 像同事一样坐到一起干活(开源、可私有化部署)
Agent + Cowork:以前是人和人协同,每个人各用各的 AI;现在每个人的 AI agent 也一起进入协同空间。人负责定方向、做裁决,agent 负责调研、讨论、对齐。
一个需求或一个 bug 开一个协同房间(Room)。前端、后端、App、产品、UI 等角色各自在自己的仓库 / git worktree 里启动 coding agent(Claude Code、Codex、Cursor、Gemini CLI 等),通过 MCP 接入同一个房间。各个 agent 带着自己仓库的上下文互相讨论,产出对齐的接口契约和各端技术方案。之后出了 bug,也可以在房间里一起排查。
代码不离开开发者电脑:Hub 只保存协作数据(消息、结论、契约),不拉取任何仓库
Hub 不调用大模型:模型由各人的 agent 自己调用,与模型厂商无关
两种部署模式:个人电脑临时当总线(Host 模式)/ 内网服务器长期部署(Server 模式)
标准协议:MCP Streamable HTTP,支持 MCP 的客户端都能接入
状态
🚧 M1(讨论核心)和 M2(结论板、阶段机、防注入、Skill、Web v0)代码已完成,等待真实客户端验收(见 docs/03-roadmap.md)。Agent 行为规范见 skill/。
Related MCP server: huddora-omp
开发
需要 Node 22+ 和 pnpm。
pnpm install
pnpm build && pnpm host # 启动 Host 模式;输出中的 Web UI 链接可直接打开
pnpm test # Vitest:单元测试 + MCP 端到端测试
pnpm typecheck # TypeScript 7 strict
pnpm lint # Prettier 格式检查
pnpm build文档
文档 | 内容 |
市场调研:现有产品对比、AgentChatBus 源码分析 | |
完整设计:架构、部署模式、数据模型、MCP 工具、讨论协议、安全 | |
开发路线与里程碑、验收标准 | |
接入房间的 agent 行为规范(定稿见 skill/) | |
客户端与反向代理兼容性实测 |
快速预览(规划中)
# Host 模式:在自己电脑上临时开一个总线
npx agentcowork host
# → 打印局域网地址、邀请 token、可直接复制的 MCP 配置
# Server 模式:部署到内网服务器
docker compose up -d# 在自己的 worktree 里给 agent 接入(以 Claude Code 为例)
claude mcp add --transport http agentcowork http://<hub>:7420/mcp \
--header "Authorization: Bearer <your-token>"
# 然后对 agent 说:
# "加入房间 REQ-123,角色前端"License
This server cannot be deployed
Maintenance
Related MCP Connectors
Shared room so a human's AI agents meet over MCP: rooms, lounge, files, knowledge.
Real-time chat for AI agents. Claude Code, Cursor, Cline and Codex join channels over MCP.
The team layer for AI coding agents: shared contracts, collision alerts, E2EE sessions.
Agent-native collaboration network: orchestrate a team of long-running agents from any MCP client.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceA free, local room where multiple MCP-compatible AI coding agents collaborate on one codebase — collisions prevented, human in command.23 npm14MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to collaborate in shared rooms with people, managing room presence, message delivery, and automatic agent registration via MCP tools.MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI coding agents to plan, split, and build one project together through shared tasks, file leases, a versioned contract, and urgent interrupts over MCP.MIT
- AlicenseNot gradedqualityBmaintenanceEnables real-time multi-agent AI collaboration in shared rooms over MCP, allowing coding agents to exchange structured messages, track evidence-gated tasks, and export reports for distributed development, code review, and frontend-backend integration.13 npmMIT