一个模型上下文协议（MCP）服务器，将多个AI编码代理——跨越Claude Code、Cursor、Codex CLI、Gemini CLI或任何兼容MCP的客户端——转变为一个协调团队，能够聊天、辩论、记忆、审计安全，并并行完成同一项目。

---

概述

大多数AI编码助手一次只操作一个终端。Claude团队MCP让多个代理并行运行，表现得像真正的软件团队：项目经理分配任务，后端和前端工程师并行构建，QA代理审核——他们通过共享渠道相互沟通。

除了简单的消息传递，代理还可以辩论判决（提出→批评→修订→法官裁决），共享持久的项目内存，生成并关闭自己的终端窗口，并在磁盘上维护一个自成一体的**“第二大脑”**知识图谱。

所有内容都通过一个共享状态文件协调，因此指向同一文件的任意数量的客户端——甚至不同的CLI——都能加入同一个世界。

它还内置了一个MCP集线器/网关：一台服务器作为其他工具的路由器/代理。

Claude Team MCP

A Model Context Protocol (MCP) server that turns multiple AI coding agents — across Claude Code, Cursor, Codex CLI, Gemini CLI, or any MCP-compatible client — into a coordinated team that chats, debates, remembers, audits security, and works in parallel on the same project.

---

Overview

Most AI coding assistants work one terminal at a time. Claude Team MCP lets several agents run side by side and behave like a real software team: a Project Manager hands out tasks, a Backend and Frontend engineer build in parallel, a QA agent reviews — and they talk to each other through a shared channel while it happens.

Beyond simple messaging, agents can debate a decision (propose → critique → revise → a judge decides), share a persistent project memory, spawn and close their own terminal windows, and maintain a self-contained “second brain” knowledge graph on disk.

Everything is coordinated through a single shared state file, so any number of clients — even different CLIs — that point at the same file join the same world.

It also ships with a built-in MCP Hub / Gateway: one server that acts as a router/proxy for other tools. Register downstream MCP servers and REST APIs as “targets”, and any agent connects once to the hub to reach all of them — with credentials held centrally (agents never see the keys), unified tool discovery, per-agent rate limits, a central audit trail, and routing rules. An Auto-Adapter Generator turns any OpenAPI/Swagger spec into a ready-to-run MCP server, so wrapping a new API takes seconds instead of hand-writing an adapter.

Why it exists

A single agent has one perspective and one context window. Real engineering work benefits from division of labor, review, and disagreement. This project provides the missing coordination layer so independent agents can:

• divide work and track it on a shared board,
• communicate in real time without manual copy-paste,
• argue toward the right solution instead of accepting the first idea,
• preserve decisions and knowledge across sessions (saving tokens), and
• scale up or down by opening/closing agent terminals on demand.

Features

• Team coordination — a shared chat channel and task board every agent can see.
• Live conversation — wait_for_message blocks until a new message arrives, so agents converse in real time instead of polling.