ComfyUI_MCP

An experimental Model Context Protocol (MCP) server that orchestrates ComfyUI workflows. It discovers workflow templates, applies high-level mutations, submits prompts to a running ComfyUI instance, and offers lightweight tooling for local experimentation.

Quick Start

1. Create a configuration file (see config.example.toml) and update the ComfyUI base URL plus any asset directories you keep on disk.

2. Install the package in editable mode:

   pip install -e .

3. Interact with the helper CLI:

   # List available workflow templates comfyui-mcp list --json # Describe a specific workflow comfyui-mcp describe basic_workflow --json # Inspect discovered checkpoints/LoRAs/VAEs comfyui-mcp assets --json

   You can override configuration values at runtime without editing the TOML file. For example:

   comfyui-mcp list --base-url http://localhost:8188 --workflows-path ~/custom-workflows \ --directory checkpoints=~/models/StableDiffusion --json

4. Launch the MCP server runtime using FastMCP:

   comfyui-mcp-serve --config path/to/config.toml --transport stdio

   The runtime exposes tools for listing, describing, customising, and executing workflows. Choose --transport streamable-http if you need an HTTP endpoint rather than stdio.

5. To integrate with MCP-compatible IDEs (e.g., Cursor), copy mcp.json into the location they expect and adjust the configuration path or transport as needed. The file declares the stdio command to launch comfyui-mcp-serve, making it easy for third-party clients to connect.

The server loads workflow templates from the workflows/ directory by default and scans configured asset directories (checkpoints, LoRAs, VAEs, text encoders, embeddings). Asset validation helps catch typos before submitting prompts to ComfyUI.