CoCo

LM mediator. Help LMs understand you.

Quickstart

1. Create virtual env and set up environment
   • Create a .env file with at least OPENAI_API_KEY and WANDB_API_KEY
   • pip install requirements.txt (or use any virtual environment manager like uv)
2. Start Redis DB - Redis Quickstart
   • docker run -d --name redis-stack -p 6379:6379 -p 8001:8001 redis/redis-stack:latest
   • docker exec -it redis-stack redis-cli
   • LRANGE request_queue 0 -1 and LRANGE response_queue 0 -1 to check the status of the queues on Redis
3. Backend
   • python backend.py (use python backend_fastmcp.py or python backend_fastapi.py)
4. Frontend
   • python frontend.py

Cursor Integration

CoCo supports two integration methods with Cursor:

Method 1: FastMCP Integration (Recommended)

This method uses Anthropic's MCP (Model Context Protocol) for direct integration.

1. Make sure to have Redis DB and gradio frontend running (frontend only necessary for CoCo's collaboration mode)

2. Add the CoCo MCP server to your Cursor configuration

   • Edit your ~/.cursor/mcp.json file to include:

   {
     "mcpServers": {
       "CoCo": {
         "command": "/path/to/CoCo/.venv/bin/python",
         "args": [
           "/path/to/CoCo/backend/backend_fastmcp.py"
         ]
       }
     }
   }
   • Replace /path/to/CoCo with your actual CoCo installation path

3. Using CoCo with FastMCP in Cursor

   • Can only be used in agent mode (chat interface)
   • No specific decorators are required, but it helps to specify "CoCo" and whether you want to use it in collaborative mode or delegate mode
   • In-line edits are currently not supported in Cursor's agent mode

Method 2: FastAPI + ngrok Integration

This method exposes CoCo as an OpenAI-compatible API that can be used with Cursor's OpenAI configuration.

1. Install additional requirements (fastapi, uvicorn, ngrok)

2. Make sure to have Redis DB and gradio frontend running (frontend only necessary for CoCo's collaboration mode)

3. Set up ngrok:

   • Run ngrok http http://localhost:8000
   • In Cursor's Model section override OpenAI base URL with <ngrok-address>/v1

4. Using CoCo with FastAPI in Cursor:

   • @coco_delegate can be used for in-line editing (command K) and works without the frontend
   • @coco_collab should be used in the chat (command L) and calls the mediator UI frontend

Using CoCo

CoCo has two modes:

• Delegation Mode: Hand off a task to the agent and work on something else in the meantime
• Collaboration Mode: Jointly co-create and teach the agent/mediator through an interactive process

The idea is to use collaboration to train the agent on your preferences, then delegate similar tasks once the agent has learned how to handle them.