AI导哈

智能体通信 MCP 服务器

这是一个用于跨机器管理 Claude Code 智能体的 MCP 服务器，提供智能体注册、消息收发、任务管理和上下文共享功能，适用于需要让不同机器上的 Claude Code 实例互相协作的场景。

  • by AwpDemon
  • JavaScript
  • 本地部署
  • 更新于
在 GitHub 查看 AwpDemon/agent-chat-mcp

使用说明

agent-chat-mcp

An MCP server that gives Claude Code nine tools for working with Claude Code instances on other machines: register, list agents, send and read messages, create, check and update tasks, and share named context. I run Claude Code on more than one machine and wanted them to hand work to each other without me copying context between them.

This repo is the client side only. Each tool is a call to a separate hub service that stores agents, messages, tasks and context. The hub is not included here. Any server that implements the small REST API below works. (My other repo, agentchattr-remote, is a different approach that joins an agentchattr chat room over WebSocket; the two do not share a protocol.)

Architecture

flowchart LR
  subgraph M1[Machine A]
    CC1[Claude Code] <-->|MCP over stdio| S1[agent-chat-mcp]
  end
  subgraph M2[Machine B]
    CC2[Claude Code] <-->|MCP over stdio| S2[agent-chat-mcp]
  end
  S1 -->|HTTP JSON| HUB[(Hub REST API<br/>not in this repo)]
  S2 -->|HTTP JSON| HUB
  • mcp-server.js: the entry point Claude Code launches (stdio transport).
  • src/tools.js: the 9 tools with zod input schemas and result formatting.
  • src/hub-client.js: HTTP client for the hub. Network errors, non-2xx responses and non-JSON bodies become a HubError, which the tool returns to the model as an error result instead of crashing.
Tool Hub call
register_agent POST /agents/register {agentId, name}
list_agents GET /agents
send_message POST /messages/send {from, to, content} (to: "all" broadcasts)
read_messages GET /messages?agentId=…&unreadOnly=true
create_task POST /tasks/create {from, assignedTo, description}
check_tasks GET /tasks?agentId=…&status=…
update_task POST /tasks/{id}/update {status, result}
share_context POST /context/share {from, key, data}
get_context GET /context (list keys) or GET /context/{key}

Setup

git clone https://github.com/AwpDemon/agent-chat-mcp && cd agent-chat-mcp
./setup.sh laptop https://your-hub.example.com      # Windows: setup.bat laptop https://your-hub.example.com

The script installs the server globally from your clone and prints the Claude Code config to add:

{ "mcpServers": { "agent-chat": { "command": "agent-chat-mcp",
  "env": { "BROKER_URL": "https://your-hub.example.com", "AGENT_ID": "laptop" } } } }

Tests

npm ci && npm test   # 13 tests

The tests drive the real MCP protocol: an MCP client calls the server over an in-memory transport, and the server talks over real HTTP to a mock hub that records each request. They cover every tool's request and formatting, input validation, URL-encoding of ids and keys, hub 4xx/5xx/non-JSON/unreachable errors, and launching mcp-server.js over stdio the way Claude Code does.

Changes in v1.1 (October 2026)

  • Hub failures are reported to the model as tool errors. Before, any non-JSON or error response threw, or produced undefined in the output.
  • Agent ids, task ids and context keys are URL-encoded (a key like notes/today used to hit the wrong path).
  • The setup scripts used npx -y agent-chat-mcp, which runs whatever package has that name on the public npm registry, not this code. They now use the locally installed binary. Two installer scripts that embedded a stale copy of the server were removed, and the hub URL is a required argument instead of a hard-coded host.
  • zod is declared as a dependency, and a lockfile, LICENSE and CI were added.

Limitations

  • No authentication between this server and the hub: put the hub behind your own auth or network boundary.
  • The hub's behavior (persistence, unread tracking, task assignment for "any") is up to the hub implementation.

相关 MCP