Plenty of people have tried to install ComfyUI once and given up. The node graph looks alien, the install fails because of a path issue, and the graphics card cannot keep up. Comfy MCP, released in public beta in July 2026, lets you skip that setup and drive ComfyUI by talking to Claude instead. This guide, based on official documentation as of August 2026, covers which connection to pick and what it costs.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What made people give up on ComfyUI
It helps to start with the barriers, because the value of Comfy MCP lies in how many of them it removes.
- Path sensitivity — installs fail when the folder path contains spaces or non-ASCII characters. ComfyUI is fussy about path handling.
- Graphics card requirements — per the official troubleshooting docs, Comfy Desktop on Windows supports only CUDA-capable NVIDIA GPUs. Other GPUs need the portable build or a manual install.
- Disk and permissions — the installer must be run as administrator and needs at least 15GB of free disk space.
- Update conflicts — users report that updating custom nodes and ComfyUI itself at the same time causes errors, and recommend doing them separately. Reports of spending a very long time just on installation are not rare.
- The node graph — the first reaction to opening it is often "what is this complicated graph tool?"
The items above come from official troubleshooting documentation and user reports. Individual environments vary.
What Comfy MCP actually is
MCP (Model Context Protocol) — an open standard that connects AI agents to outside tools and data. Comfy MCP uses it to connect the ComfyUI engine to an agent like Claude. Once connected, you can generate images, video, audio and 3D, search models, nodes and templates, and run real workflows from a conversation.
You never call the tools yourself. Ask in plain language — "generate an image of a cat astronaut" — and the agent picks the right tool. The server prefers matching an existing pre-built template over building a workflow from scratch, which tends to produce better results faster.
Cloud or local?
The official guidance is clear.
| If this is you | Recommended |
|---|---|
| New to ComfyUI | Cloud — the simplest setup |
| You use claude.ai or the Claude Desktop chat app | Cloud — the more compatible choice |
| You want to run open-source models on a Mac | Cloud — today's open-weight models are large and will not run at a workable speed on the Apple GPU |
| You already run ComfyUI locally | Local |
| You work mostly in Claude Code, Cursor or Codex | Local |
| You want your own models, LoRAs and custom nodes | Local — it reads your live install |
This is not an either-or choice. Running both is normal, and the docs recommend it for people who run ComfyUI locally. Note that the sign-ins are separate: signing in on one does not sign you in on the other, even on the same Comfy account.
Can your machine handle the local connection?
The official FAQ gives these VRAM guidelines.
| Setup | What it handles |
|---|---|
| Mac | Use cloud for generating |
| Dedicated GPU, 24GB or more | Most things, including video |
| 8–24GB | Fine for images; video will be slow or will not fit |
| Under 8GB | Use cloud |
What it costs
This is the part that confuses people most. The split is precise:
- Discovery is free — searching templates, models and nodes needs only a Comfy account, on both connections.
- Cloud generations require an active subscription. A credit or top-up balance alone does not grant access — you need an active subscription even if you have unused credits.
- New users get 5 free runs.
- Local runs are free because they happen on your hardware. The exception is partner models, which execute on partner infrastructure and spend credits.
Subscription tiers and credit policy can change. Check the official Comfy Cloud page for current terms before paying.
Connecting from Claude Code
The simplest path is the official plugin. Register the marketplace once, install the plugin, then authenticate.
/plugin marketplace add Comfy-Org/comfy-skills
/plugin install comfy-cloud@comfy-skillsThen run /mcp, select comfy-cloud and choose Authenticate. Your browser opens for sign-in, and tokens refresh automatically after that.
The plugin adds commands such as /comfy-cloud:generate-image, /comfy-cloud:generate-video, /comfy-cloud:generate-audio, /comfy-cloud:generate-3d, /comfy-cloud:upscale-image, /comfy-cloud:remove-background and /comfy-cloud:search-templates.
If you want the connection without the plugin:
claude mcp add --transport http comfy-cloud https://cloud.comfy.org/mcpAdd -s user to make it available across all projects. This path exposes the same workflows, but command names become /mcp__comfy-cloud__generate-image, which is why the docs recommend the plugin.
Connecting from Claude Desktop
In the sidebar, open Customize → Connectors, click +, and choose Add custom connector. Give it a name (for example, Comfy Cloud MCP) and set the remote MCP server URL to:
https://cloud.comfy.org/mcpA browser window opens. Pick your workspace, click Continue, and you are connected.
Slash commands do not work in Claude Desktop. They ship with the Claude Code plugin. In Desktop, use the prompt picker or simply ask in plain language — it is the same MCP server, so the tools work identically.
What the local connection needs
The local side uses the open-source comfy-mcp server. Requirements:
- Python 3.10 or newer
comfy-cli1.14.0 or newer on your PATH —pip install "comfy-cli>=1.14.0"- A ComfyUI workspace — create one with
comfy installif you do not have one - A running ComfyUI — start it with
comfy launch. The server does not launch it for you
pip install "comfy-cli>=1.14.0"
comfy install
pip install comfy-mcpRegistering it with Claude Code is one command:
claude mcp add comfy-mcp -e COMFY_BIN=/path/to/venv/bin/comfy -- comfy-mcpMCP clients launch the server with their own environment, which often does not include your shell's PATH. If comfy lives in a virtualenv or a non-standard location, set COMFY_BIN to its absolute path.
Limitations worth knowing
This is an early release, and the docs are explicit about the rough edges.
- Outputs require a shell download step. The cloud server never writes to your machine. When a generation finishes you get a temporary signed URL plus a ready-to-run command, which the agent executes in your shell. Run it verbatim — the signature lives in the query string and breaks if the URL is edited.
- Workflow metadata may be missing. Assets generated via
submit_workflowmight not reopen the originating workflow when opened in ComfyUI. - Complex workflows may need a retry. Multi-node builds depend on agent accuracy and may need refinement.
- Cursor needs an API key. It does not support MCP OAuth today, so the key goes in your MCP config. Claude Code, Claude Desktop, Codex and OpenClaw use browser OAuth.
- Uploads and downloads depend on your client's file access. The docs recommend Claude Code for Claude users because it has more capabilities.
FAQ
Does this work on claude.ai in the browser?
The cloud connection works with any client that supports remote HTTP. The local connection does not work in browser-based clients — it needs a client that can launch a local stdio server as a subprocess. claude.ai and ChatGPT accept remote connectors only.
Do I need to reconfigure to switch between cloud and local?
No. With both connected, just say where a job should run — "run this on Comfy Cloud", "do this locally". There is no mode to toggle.
How do updates work?
The cloud connection is hosted, so you are always current. For local, update and then restart your client — MCP servers load when a session starts, so a running session keeps serving the old version.
Related articles
- Connect Claude to Figma: Figma MCP (Dev Mode) Setup for Claude Code (2026)
- Claude Code MCP Setup: claude mcp add, Scopes, and Transports
- Claude Code settings.json Guide — Permissions, Env Vars & Tools
- Adding a Custom Connector to Claude Code — Complete MCP Setup Guide
- Claude Code MCP Connectors — How to Link GitHub, Sentry, and Databases