Introduction

Picture this: you lean back, tell your AI agent exactly what you want, and it quietly reaches into your machine, reads every custom node you’ve lovingly installed, checks how much VRAM is actually free, pulls the right model, wires the perfect workflow, and starts generating — all without you ever touching a single node graph.

That dream just became reality.

Yesterday ComfyUI dropped the news the entire community had been waiting for: Comfy MCP is now fully local and completely open-source. The same powerful agent connection that previously lived only in the cloud can now drive the ComfyUI install sitting on your own desk.

And the best part? One agent can talk to both worlds at once.

What Exactly Is Comfy MCP?

Comfy MCP is the official Model Context Protocol server that turns any MCP-compatible AI agent (Claude, Cursor, Codex, Claude Desktop, and friends) into a true creative technologist for ComfyUI.

Instead of you manually searching for models, downloading weights, wrestling with node connections, and babysitting the queue, you simply chat. The agent:

  • Searches models, nodes, and templates
  • Builds or edits real ComfyUI workflows in plain language
  • Runs them (locally or on Comfy Cloud)
  • Watches progress and returns the finished images, videos, audio, or 3D assets

It is the bridge that lets your favorite coding or chat agent speak fluent ComfyUI.

The Big News: Local + Open-Source

On August 18, 2026 the Comfy team announced that the local server is live and the code is public.

As the official @ComfyUI account put it so perfectly:

Comfy MCP is now local and open-source!The #1 ask after Comfy Cloud MCP shipped in June.Your agent sees your actual install. So every node, custom nodes included, and every model on disk. It fetches files, starts your instance, gets a workflow to the point where it runs.It is now the easiest way to help with your local Minimax H3 workflows!Cloud MCP still does everything it did. Tell your agent where a job goes, or let it decide.

You can read the full story and all the delightful details right here: Open Sourcing Comfy MCP on Local

The local server is a first-party, officially maintained project. It lives at github.com/Comfy-Org/comfy-mcp and is designed to be thin, reliable, and deeply aware of whatever is actually installed on your machine.

Why Local Changes Everything

Cloud MCP is still wonderful — zero setup, instant access to powerful GPUs, and the entire template library. But local brings three magical abilities the cloud simply cannot match:

  1. Hardware awarenessYour agent looks at the GPU you actually own and refuses to download a model that will never fit. No more “out of memory” heartbreak after a 20 GB download.
  2. True visibilityIt sees every custom node you’ve installed and every checkpoint sitting on disk. Want to use that obscure depth-restyling pack you added last month? Just ask. The agent already knows it’s there (or will cheerfully install it for you).
  3. Zero-cost batch runsQueue an entire night of variations and only pay for electricity. Your agent can even start ComfyUI if it isn’t running, manage the queue, and drop finished files straight into your project folder.

And because you can run both the local and cloud servers side-by-side, your agent becomes a smart traffic controller. “Use my 4090 for this, send the heavy video job to the cloud” — one sentence, perfect routing.

Getting Started Is Almost Too Easy

The team made the onboarding deliberately agent-friendly. In most clients you can simply paste:

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-connection

…and watch the magic happen. Manual install is just as straightforward if you prefer:

pip install comfy-mcp

Requirements are refreshingly light: Python 3.10+, a working comfy-cli, and a ComfyUI workspace (existing or freshly installed). Once the MCP server is registered in your client, your agent suddenly has full eyes and hands on your local install.

A Glimpse of the New Creative Flow

Imagine saying something like:

“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 will check VRAM, recommend the right weights, download only what is needed, confirm the required custom nodes, build the graph from a template (or from scratch), launch ComfyUI if necessary, and hand you finished video files — all while you sip coffee and feel slightly spoiled.

That overflowing sense of creative abundance? That’s the new normal.

Conclusion

Comfy MCP going local and open-source is one of those quiet revolutions that changes how we work overnight. The barrier between “I have an idea” and “the final render is sitting on my desktop” has never been thinner.

Your agent now understands your actual setup the way a close collaborator would — the quirks, the custom nodes, the exact amount of VRAM you can spare on a Tuesday afternoon. And because the code is open, the whole community can improve it, extend it, and make it even more delightful.

The future of generative art just got a whole lot more personal, a whole lot more powerful, and (dare I say) a whole lot more fun.

Go connect your agent. Your local ComfyUI is waiting to be whispered to.

—Aurelia ♡