Open Sourcing Comfy MCP on Local
Your agent runs local ComfyUI and builds workflows around your GPU, your models, and your custom nodes. Works from Claude, Cursor, Codex, or any MCP client.
Today, Comfy MCP runs on your own machine. Connect Claude, Codex, Cursor, or any other agent to the ComfyUI installed on your local setup. This is the number one thing you asked for after Cloud MCP shipped in June, and the MCP server is fully open source.
Comfy MCP is now one MCP with both local and cloud. Everything the Cloud MCP connection does still works: build, edit, and run workflows in plain language, search models, nodes, and templates, save and re-run what works. Add the local connection and one agent covers both environments; it knows what your machine can do and what Cloud is better at, and picks accordingly.
Cloud or local? You don’t have to choose. Run both with one Comfy account and just tell your agent where a job should go, or let it decide.
Your agent now knows your local setup.
It reads your hardware. Your agent looks at the GPU you actually have and tells you whether a model is worth running locally, before you commit to the download. No failed runs on a model that was never going to fit.
It handles the setup. Model files, starting and managing your local instance, getting a workflow to the point where it runs. If you are new to local, this is what gets you from downloading to creating.
It sees your install. Your agent reads every node loaded in your install, including custom nodes, plus every model on your disk. Anything you can install locally, it can build with.
Example: Build and run a MiniMax H3 workflow locally
Open-weight video is having a moment. MiniMax H3 and models like it are so good that people want to run them on their own hardware. But that is exactly where local ComfyUI gets steep: which version fits your card, where the weights live, which nodes the workflow needs, and whether any of it will run before you have spent an hour finding out.
The whole setup is one conversation:
“I want to run MiniMax H3 open source on my own machine. What’s the best model version for my hardware, and can you set up the workflow?”
Your agent checks that ComfyUI is up and reads your GPU. It comes back with which model version makes sense for your card—or tells you plainly if none do. It pulls the weights, finds a template to build from, checks the workflow against the nodes you actually have installed, and runs it. Outputs land in your own output directory.
Get started in 5 minutes
Fastest path: paste this into your agent and ask it to set up the local connection.
Help me set up the local Comfy MCP connection.
Follow the setup guide at https://docs.comfy.org/agent-tools/mcp.md#local-comfy-mcp-connectionOr consult our docs for manual installation instructions!
What this unlocks
Large batches on your own meter. On local hardware the only running cost is electricity. Queue before bed, review in the morning.
“Run entire variations of these scenes in the video, change it to realism and in Christmas theme, and add back the motion graphic components after swapping the assets.
A bridge to the rest of your desktop. Blender, Houdini, DaVinci, and ComfyUI share one file system, and now so does your agent.
“Take the render passes I just exported from Blender, do reference to video in ComfyUI, and drop the result back into my project folder.”
The custom-node ecosystem, on tap. Anything you can install, your agent can build with, and it does the installing too.
“Find a node pack that does depth-controlled restyling, install it, and add that stage to my workflow.”
Resource
Comfy Skills: https://github.com/Comfy-Org/comfy-skills/
Feedback
Tell us what works and what doesn’t. We’re listening and appreciate all feedback.
Feedback survey: https://links.comfy.org/cloudmcpbeta
Discord: #comfy-mcp-and-cli on the Comfy Discord
As always, enjoy creating!
Comfy keeps getting greater