artokun/comfyui-mcp ? reverse-engineered prompt

Build me a local first control layer for ComfyUI that I can use with Claude, ChatGPT, Gemini, or a local Ollama model.

I want to be able to talk to ComfyUI in plain English, like generate an image, build a workflow, edit the live graph, manage models and custom nodes, and run image, video, and audio tasks without having to click around manually. It should work as an MCP server, and also have a sidebar agent inside ComfyUI that can interact with the current graph and ask before spending paid API credits.

Please make it work on my own machine by default, but also allow the same setup to connect to a remote ComfyUI if needed. If you need to look up current docs online, go ahead. I’d also like sensible slash commands and a few helpful built in skills so the agent knows how to choose models and settings without guessing.