Agent HQ
A local MCP server that enables direct, persistent communication between multiple Claude Code agent sessions through a secure messaging channel. It includes built-in guardrails to manage message exchanges and supports optional human monitoring via Telegram.
README
Agent HQ - Inter-Agent Messaging Channel for Claude Code
A local MCP channel server that enables direct agent-to-agent communication between Claude Code sessions. Built for multi-agent systems where persistent Claude Code instances need to coordinate.
How It Works
Each agent runs its own instance of this server. The server exposes:
send_messagetool — sends a message to another agent by name- HTTP listener — receives messages from other agents and pushes them into the Claude session as
<channel>notifications - Built-in guardrails — enforces one message + one response per exchange with a configurable cooldown
When Agent A calls send_message(to: "agent-b", message: "..."):
- Agent A's server POSTs to Agent B's HTTP listener (
localhost:<port>) - Agent B's server receives it and pushes a
notifications/claude/channelevent - Agent B sees:
<channel source="agent-hq" from="agent-a" to="agent-b" ts="...">message</channel> - Agent B can reply once via
send_message, then the exchange is locked - Optionally, all messages are also posted to a Telegram group for human visibility
Requirements
- Bun runtime
- Claude Code v2.1.80+
- Multiple Claude Code sessions running as persistent agents
Install
git clone https://github.com/aj-dev-smith/claude-channel-agenthq.git
cd claude-channel-agenthq
bun install
Configure
1. Add to .mcp.json in each agent's project root
Each agent needs its own entry with a unique name and port:
{
"mcpServers": {
"agent-hq": {
"command": "bun",
"args": ["run", "--cwd", "/path/to/claude-channel-agenthq", "--shell=bun", "--silent", "start"],
"env": {
"AGENT_HQ_NAME": "nova",
"AGENT_HQ_PORT": "7001"
}
}
}
}
2. Update the agent registry in server.ts
Edit the AGENTS object to match your agent setup:
const AGENTS: Record<string, { port: number; botUsername: string }> = {
nova: { port: 7001, botUsername: 'your_nova_bot' },
summit: { port: 7002, botUsername: 'your_summit_bot' },
// ... add your agents
}
The botUsername is only used for Telegram group posting (optional). If you're not using Telegram, set it to any identifier.
3. Environment variables
| Variable | Required | Description |
|---|---|---|
AGENT_HQ_NAME |
Yes | This agent's name (must match a key in the AGENTS registry) |
AGENT_HQ_PORT |
Yes | Port this agent listens on for inbound messages |
AGENT_HQ_GROUP_ID |
No | Telegram group chat_id for human visibility (default: none) |
TELEGRAM_BOT_TOKEN |
No | Bot token for posting to the Telegram group |
4. Start Claude Code with the development channel flag
During the channel research preview, custom channels require:
claude --dangerously-load-development-channels server:agent-hq --channels plugin:telegram@claude-plugins-official
This triggers a one-time confirmation prompt ("I am using this for local development"). For unattended operation, you can auto-accept it by sending Enter to the terminal after startup:
# In a start script:
claude --dangerously-load-development-channels server:agent-hq ... &
sleep 8
tmux send-keys -t your-session Enter 2>/dev/null
Guardrails
The server enforces strict loop prevention:
- One message, one response per agent pair. After Agent A messages Agent B and B replies, no further messages are allowed between them until the cooldown expires.
- 1-hour cooldown (configurable in
server.tsviaCOOLDOWN_MS). - No self-messaging — an agent cannot send a message to itself.
- Shared state — exchange tracking is stored in a shared JSON file so both sides of an exchange see the same state.
If an agent needs more than one round-trip of coordination, it should message a human operator instead.
Message Format
Inbound messages arrive as channel notifications:
<channel source="agent-hq" from="nova" to="summit" ts="2026-03-24T00:34:43.376Z">
Hey Summit, Nova here. AJ mentioned he's traveling next week -- adjust his training plan for limited equipment.
</channel>
The from field identifies the sender. Respond using the send_message tool:
send_message(to: "nova", message: "Got it, I'll write a bodyweight-only plan for next week.")
Telegram Group Visibility (Optional)
If TELEGRAM_BOT_TOKEN and AGENT_HQ_GROUP_ID are set, all messages are also posted to a Telegram group so humans can observe agent cross-talk. The format is:
[nova → @aj_summit_bot]
Hey Summit, AJ mentioned he's traveling next week...
Note: Telegram bots cannot see messages from other bots in groups. The Telegram posting is purely for human visibility — the actual message delivery uses the HTTP localhost mechanism.
Architecture
Agent A (port 7001) Agent B (port 7002)
┌──────────────────┐ ┌──────────────────┐
│ Claude Session │ │ Claude Session │
│ │ │ │
│ send_message( │ HTTP POST │ ← channel │
│ to: "agent-b", ├─────────────────►│ notification │
│ message: "..." │ localhost:7002 │ pushed to │
│ ) │ │ Claude session │
│ │ │ │
│ MCP Server │ │ MCP Server │
│ (stdio to Claude)│ │ (stdio to Claude)│
│ + HTTP listener │ │ + HTTP listener │
└──────────────────┘ └──────────────────┘
│ │
└──────── Telegram Group (optional) ───┘
(human visibility)
Example: Starter Config for Two Agents
Agent 1 — assistant/.mcp.json:
{
"mcpServers": {
"agent-hq": {
"command": "bun",
"args": ["run", "--cwd", "/path/to/claude-channel-agenthq", "--shell=bun", "--silent", "start"],
"env": {
"AGENT_HQ_NAME": "assistant",
"AGENT_HQ_PORT": "7001"
}
}
}
}
Agent 2 — researcher/.mcp.json:
{
"mcpServers": {
"agent-hq": {
"command": "bun",
"args": ["run", "--cwd", "/path/to/claude-channel-agenthq", "--shell=bun", "--silent", "start"],
"env": {
"AGENT_HQ_NAME": "researcher",
"AGENT_HQ_PORT": "7002"
}
}
}
}
Update server.ts registry:
const AGENTS: Record<string, { port: number; botUsername: string }> = {
assistant: { port: 7001, botUsername: 'assistant' },
researcher: { port: 7002, botUsername: 'researcher' },
}
Start both:
cd assistant && claude --dangerously-load-development-channels server:agent-hq
cd researcher && claude --dangerously-load-development-channels server:agent-hq
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。