cmf-mcp

name	cmf-mcp
description	Use when setting up the CMF MCP server so an AI coding assistant can query CMF metadata using natural language, or when the CMF MCP server is already running and the user wants to use it from their coding agent. Covers Docker deployment, per-agent configuration (Claude Code, GitHub Copilot, Cursor, Codex), and example prompts.
version	1.0.0

First check: is the CMF MCP server already running? Check for an existing .mcp.json in the project. If yes, skip to Using CMF MCP from your coding agent.

Prerequisites (new deployment)

• Docker + Docker Compose
• A running CMF Server (see ../cmf-server/SKILL.md)
• An MCP-compatible AI assistant

Deploy the MCP server

cd mcp/
cp example.env .env

Edit .env:

CMF_BASE_URL=http://<your-cmf-server-host>:80
# Optional second/third/fourth server:
# CMF2_BASE_URL=http://staging-server:80
# CMF3_BASE_URL=http://another-server:80

docker compose up -d
curl http://localhost:8000/health   # → {"status": "ok"}

Connect to your coding agent

Claude Code — .mcp.json (project root)

{
  "mcpServers": {
    "cmf": { "url": "http://localhost:8000/mcp", "transport": "http" }
  }
}

Restart Claude Code or run /mcp to reload.

GitHub Copilot — VS Code settings

{
  "github.copilot.chat.mcpServers": {
    "cmf": { "url": "http://localhost:8000/mcp", "transport": "http" }
  }
}

Add to .vscode/settings.json or VS Code user settings. Reload the window.

Cursor — ~/.cursor/mcp.json

{
  "mcpServers": {
    "cmf": { "url": "http://localhost:8000/mcp", "transport": "http" }
  }
}

Codex — ~/.codex/config.json

{
  "mcpServers": {
    "cmf": { "url": "http://localhost:8000/mcp", "transport": "http" }
  }
}

---

Using CMF MCP from your coding agent

Once connected, query CMF metadata in natural language:

What pipelines are tracked in CMF?
Show all executions for the Train stage in my_pipeline.
What datasets produced models/rf.pkl?
Show the model card for model ID 7.
What were the evaluation metrics for the latest run?

Check available tools — in Claude Code run /mcp; in Cursor/Copilot type @cmf in agent chat.

End-to-end workflow:

1. Instrument pipeline → cmf metadata push -p my_pipeline → ask the agent natural language questions

See references/mcp-tools.md for the full tool reference.

Troubleshooting

• Connection refused on port 8000 — run docker compose up -d in mcp/
• CMF Server unreachable — verify with curl http://<cmf-server>:80/apiv1.0/pipelines
• Tools not appearing — restart the agent after updating .mcp.json
• Empty results — push metadata first: cmf metadata push -p <name>

---

Docs: MCP Server overview · Quick Start · Configuration · Tools Reference