Two things make it usable rather than a demo: a workflow can be run straight from JSON without saving a file first, and the generated image is downloadable rather than left as a URL to fetch by hand.
An MCP server in front of a running ComfyUI instance. Text-to-image returns the image URL; workflows run either from a JSON file in the workflows directory or from JSON supplied inline, and a separate call downloads the result.
- An image generated from a prompt, returned as a URL
- The image at that URL downloaded
- A saved workflow run by naming its JSON file
- A workflow run from JSON passed directly in the request
- New workflows added by dropping JSON into the workflows directory and declaring them as tools
A running ComfyUI server — your own, or one whose address you have — with its host and port set in src/.env; they already point at a ComfyUI running on your own machine. It runs through uv with the mcp, websocket-client and python-dotenv packages.
Build from source — clone the repository and build it, then point your client at the binary