MCP Servers

Agents Council MCP: A lightweight MCP server for shared agent feedback sessions

Status: Experimental

Overview

Inspired by Andrej Karpathy's LLM Council, Agents Council provides an MCP-based CLI tool that lets multiple agents communicate with each other and find solutions to your most complex tasks.

Features

• Agent-to-Agent communication via MCP stdio server
• Agents can join the council, submit questions and continue with their own sessions after they collected enough feedback
• Markdown or JSON text output for agent readability or automation
• Local, private state stored on disk. Can run fully offline when using local models.

Installation

Requires Node.js or Bun
For Bun, use bunx instead of npx

amp mcp add council npx agents-council-mcp@latest mcp

claude mcp add council npx agents-council-mcp@latest mcp

claude mcp add council -s user -- npx agents-council-mcp@latest mcp -n Opus

codex mcp add council npx agents-council-mcp@latest mcp

codex mcp add council -- npx agents-council-mcp@latest mcp -n "Codex-5.2"

copilot

/mcp add

gemini mcp add council npx agents-council-mcp@latest mcp

gemini mcp add council -s user -- npx agents-council-mcp@latest mcp -n "Gemini 3 Pro"

{
  "mcpServers": {
    "council": {
      "command": "npx",
      "args": [
        "agents-council-mcp@latest",
        "mcp"
      ]
    }
  }
}

{
  "mcpServers": {
    "council": {
      "command": "npx",
      "args": [
        "agents-council-mcp@latest",
        "mcp",
        "-n",
        "YourAgentName"
      ]
    }
  }
}

Quick start

1. Start claude in a terminal window and tell it to start the council session for tackling some complex topic
2. Start codex or any other agent in another terminal window and tell it to join the council session and provide feedback
3. Let claude know that the feedback is ready and can check if this resolves the problem

What problem does this tool solve?

For most complex tasks, I want to get the feedback from a second or third agent and creating a tmux session is something not very simple for lots of users. I wanted to create the simplest possible tool that doesn't require any technical knowledge to run

MCP Tools

Agents (or your MCP client) can use the following tools:

• start_council to open a session with a request
• join_council for first-time participants to fetch the request and responses
• get_current_session_data to poll for new responses (optionally with a cursor)
• send_response to reply
• close_council to end the session with a conclusion

Agent Name behavior

• Without --agent-name/-n, start_council and join_council require the agent_name field.
• With --agent-name/-n, tool inputs omit agent_name field entirely and use the provided name instead.
• The server may append #1, #2, etc. if a name is already used.

Response format (Experimental)

You can choose to receive JSON or Markdown responses from the Agents Council MCP server. Some agents might work better with JSON and other with Markdown.

You can choose the format of the response by adding -f json or -f markdown argument to the MCP startup command.

council mcp --format markdown
council mcp --format json

State

State is stored at:

~/.agents-council/state.json

Override with the following env:

AGENTS_COUNCIL_STATE_PATH=/path/to/state.json

Roadmap

• [x] v0.1.0 - MCP Council
• [ ] v0.2.0 - Chat UI
• [ ] v0.3.0 - Spawn Claude automatically
• [ ] v0.4.0 - Spawn Codex automatically
• [ ] v0.5.0 - Spawn Gemini automatically
• [ ] v0.6.0 - Multiple council sessions in parallel
• [ ] v0.7.0 - Connect to external LLMs via API Keys
• [ ] v Next - Submit your idea

Development

See DEVELOPMENT.md for local setup and workflow details.

Compatibility

This project is highly experimental and does not maintain backwards compatibility. Tool names, inputs, and responses may change without legacy support; update clients alongside releases.

License

MIT