ComfyUI drops Comfy MCP to let AI agents run node workflows for you

Comfy MCP is a new bridge framework from ComfyUI that connects AI agents directly to the Comfy ecosystem. It turns agents like Claude, Cursor, Codex, and Hermes into automated creative technologists that can build and run production pipelines for you.

What it can do

ComfyUI is widely considered the best environment for running open-source image, video, and audio models locally. The catch has always been the interface: a massive, tangled web of nodes, wires, and custom extensions that intimidates most creators.

Comfy MCP eliminates the need to touch that graph. You just prompt your agent in natural language. For instance, you can tell it to take a hero product shot and generate 20 ad variants in multiple aspect ratios, or turn storyboard sketches into video clips.

The agent searches for the required models, sets up the correct nodes, pulls best-practice templates, and runs the entire pipeline end to end. It can also read saved workflow URLs, re-run them with new inputs, and hand the editable workflow back to your team.

Comfy handles the best practices under the hood and updates them automatically, meaning your agent will not get stuck using deprecated nodes or broken pipelines.

How to set it up

Comfy MCP is currently in public beta and free to test. Setup takes roughly five minutes.

You will need a Comfy account, which you can create either at cloud.comfy.org or within the Comfy desktop application. If you are on Claude Desktop, you just add it as a Custom Connector. For other agent clients like Cursor, you can simply point your agent to the official documentation and let it configure the connection for you.

Alongside the MCP, Comfy rolled out a beta version of the Comfy CLI, and launched an open-source repository named comfy-skills on GitHub for community members to share modular creative recipes.

Why it matters

This is a massive unlock for non-technical creators who wanted the raw power of ComfyUI models without having to earn a degree in node wiring. You get the quality of open-source frontier models entirely through chat.

For production teams, it solves the workflow handoff bottleneck. Power users can build complex pipelines once, and teammates can prompt an agent to execute those pipelines repeatedly without touching the underlying graph. Keep in mind it is still in beta, so unexpected node conflicts and errors will still pop up, but turning natural language into deterministic node pipelines is exactly where creative AI needed to go.