The Hook
On August 18, the ComfyUI team open-sourced Comfy MCP on Local. In plain terms: you can now point Claude, Cursor, or Codex at the ComfyUI running on your own machine, and the agent builds, edits, and runs your image and video workflows by itself.
No cloud credits. No copy-pasting JSON graphs. You type a sentence, and the AI drives your GPU. For anyone with a decent card sitting in their tower, this is the missing bridge between “chat with an assistant” and “actually make the picture.”
The Story
ComfyUI is the node-based engine a huge slice of the open-source AI-art world runs on. It is powerful and it is fiendish: every workflow is a spaghetti graph of nodes you wire by hand. The learning curve is the whole joke of the tool.
Comfy MCP is a small server that speaks the Model Context Protocol — the same open standard the Lab keeps bumping into, now living inside Blender, SketchUp, Revit and the rest. MCP is just a shared language that lets an AI agent call real tools. Here, those tools are your ComfyUI install.
A cloud version shipped back in June. This new drop is the local one, and that is the important word. The agent connects to the ComfyUI on your desk, sees the actual custom nodes and model files you have installed, checks your VRAM before it downloads anything, then runs the graph and drops the output in your folder.
Under the hood it exposes around 40 MCP tools. A few that matter:
- Discovery —
search_templates,search_nodes,search_models: the agent finds what you already have and what you are missing. - Execution —
run_workflow,generate_image,vary_workflow: it runs graphs, tweaks them, and re-rolls variations. - Lifecycle —
launch_comfyui,restart_comfyui,update_comfyui: it can start and update Comfy itself. - Jobs & memory —
job_status,wait_for_job,free_memory: it watches renders and clears VRAM so a local LLM and Comfy can share one card.
That last point is the quiet genius. The tools know your machine is finite. The agent reads your hardware, frees memory, and coordinates instead of blindly hammering the GPU.
Why You Should Care
If you make 3D or motion work, ComfyUI is where a lot of the real pipeline magic happens — character consistency, ControlNet passes, video models, the TRELLIS and Qwen graphs we keep covering. The catch was always that you had to be the node wizard.
Comfy MCP moves the wizardry to the agent. “Confirm my local ComfyUI is running, then run the workflow at ~/workflows/txt2img.json and show me the image” is a real, working prompt. The agent verifies the server, executes the graph, and fetches the result. You stay in the creative loop; it handles the plumbing.
It is also honest about hardware. The docs say it straight: on a Mac, use the cloud connection, because today’s open-weight models are too big to run at a workable speed on the Apple GPU. On a PC, 24 GB+ of VRAM handles most tasks including video; 8–24 GB is fine for images; under 8 GB, lean on the cloud. No hype, just numbers.
Try It / Follow Them
Setup is genuinely short. You need Python 3.10+ and comfy-cli:
pip install comfy-mcp "comfy-cli>=1.14.0"
comfy install # create a ComfyUI workspace
comfy launch # start it upThen register the server with your agent. For Claude Code it is one line:
claude mcp add comfy-mcpClaude Desktop and Cursor each take a small JSON block pointing at the comfy-mcp command. The full walkthrough, including the Codex path, lives in the docs.
- Announcement: Open Sourcing Comfy MCP on Local (Comfy blog)
- Docs: docs.comfy.org/agent-tools/mcp
- Code: github.com/Comfy-Org/Comfy-mcp (AGPL-3.0 or commercial)
IK3D Lab Take
We have watched MCP quietly become the wiring of the whole creative stack — Blender, Revit, SketchUp, and now the most important open-source generation engine of them all. That is the real story here: your tools are learning to be called, not just clicked.
For a maker with an RTX card gathering dust between renders, this is close to free power. The agent already knows how to write a ComfyUI graph better than most of us; now it can run one on our own metal and hand back the file. The days of memorizing node spaghetti are numbered.
Honest caveat: it is version one, the model-management is only as smart as your library is tidy, and Mac users are still pushed to the cloud for the heavy lifting. But the direction is unmistakable. Wire it up, point Claude at a workflow, and watch your GPU do the talking.
