ComfyUI's Official MCP: Let Claude or Codex Build Your Node Workflows
ComfyUI now has an official MCP server, letting Claude or Codex build and run node workflows for you. Here's how it works and how to set it up.
What is ComfyUI’s official MCP?
ComfyUI has released an official Model Context Protocol (MCP) server, a bridge that lets an LLM like Claude or Codex directly operate ComfyUI on your behalf. Instead of manually wiring together nodes for samplers, schedulers, checkpoints, and video models, you describe what you want in plain language and the model builds the node graph for you. It works alongside an existing community-built ComfyUI MCP that’s been around longer and is more battle-tested, while the official version is newer and currently in beta.
TL;DR
- ComfyUI’s official MCP server lets Claude Desktop, Codex, or other computer-use-capable LLMs build, edit, and run ComfyUI node workflows through natural language instead of manual graph editing.
- Installation is close to trivial: you tell your LLM to install the ComfyUI MCP (both the official and community versions, ideally), and it handles the setup itself.
- ComfyUI remains one of the most powerful and customizable tools in local generative image and video work, but its node-graph interface has a steep learning curve that the MCP is designed to remove.
- The tool supports both local open-source models (Miniax, LTX, Krea, Wan, and others arrive quickly after release) and remote API models like Grok Imagine, meaning MCP-driven workflows aren’t limited to what your GPU can run.
- Running everything locally means no per-generation cost, though hardware matters: a modern Nvidia GPU in the 4090 to 5090 range is the practical sweet spot, with cloud GPU rental or ComfyUI’s own cloud offering as alternatives.
- Once Claude or Codex builds a workflow, you can edit simple parameters yourself (prompt text, video length, aspect ratio) directly in the ComfyUI interface instead of spending LLM tokens on small tweaks.
- Using cheaper, faster sub-models (like Sonnet or Haiku instead of a heavier flagship model) as the “hands” that execute MCP actions is a practical way to control token costs.
One coffee. One working app.
You bring the idea. Remy manages the project.
Why has ComfyUI been hard to use until now?
ComfyUI’s core interface is a node graph: individual blocks for each step of an image or video generation pipeline, wired together into a chain. That structure is exactly why it’s so flexible. You can combine a local image model with a video model, add reference images, mix in remote APIs, and adjust granular settings like samplers, schedulers, and step counts. But building that graph from scratch requires understanding what each node does and how they connect, which has historically meant a mix of artistic sense and engineering literacy.
That combination has made ComfyUI simultaneously one of the most important tools in open-source image and video generation and one of the most divisive. Power users treat it as an infinitely flexible instrument. Newcomers often bounce off the visual complexity of a fully built-out graph, which can look like a tangle of boxes and connecting lines with no obvious starting point.
How does the MCP change the ComfyUI workflow?
The MCP effectively outsources the “engineer half” of ComfyUI to the LLM. You describe the outcome, for example a workflow that takes a Krea-generated image and animates it with Miniax, and Claude or Codex constructs the actual node graph, connects the right components, and can execute the run. The person driving it doesn’t need to know which sampler or scheduler to pick, or how to route an image output into a video model’s input.
In practice this means you can:
- Ask for a specific pipeline (image model into video model, or reference-to-video) in one sentence.
- Have the LLM generate an image, feed it into the next stage, and run the full workflow without touching the node graph.
- Upload a reference image directly to the LLM and have it inserted into the correct place in the workflow.
- Minimize the ComfyUI window entirely and work only through the chat interface if you choose to.
None of this changes what ComfyUI itself is capable of. It changes who can access that capability without first becoming fluent in node graphs.
How do you set up ComfyUI’s MCP with Claude or Codex?
Getting ComfyUI running locally starts with downloading ComfyUI Desktop rather than installing from GitHub. The desktop build avoids the older manual process of cloning a repository and managing Python environments, and it’s now the recommended starting point for new users.
Once ComfyUI is installed, setting up the MCP is a matter of instructing your LLM directly: tell Claude Desktop or Codex to install the ComfyUI MCP, specifying that you want both the official and community versions. The LLM handles the installation itself. There’s no separate configuration file to hand-edit or repository to clone for the MCP layer.
Why install both versions? The official MCP is newer and still in beta, while the community-built version has been in use longer and is more stable in certain situations. Keeping both available means you can ask your LLM to switch to whichever one handles a given task more reliably.
Requirements to actually use this setup:
- Claude Desktop specifically (not the web version), or Codex, or another LLM platform that supports agentic computer use.
- Comfort allowing the LLM some level of control over your machine, since MCP-driven workflows involve the model taking real actions rather than just generating text.
- A capable local machine if running models locally. ComfyUI runs on most modern PCs or Macs, but an Nvidia GPU in the 4090 to 5090 class avoids most friction, particularly for video generation.
- Alternatively, a rented cloud GPU or ComfyUI’s own cloud offering, for those without a strong local GPU.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Is it safe to let an LLM control ComfyUI?
Giving an LLM “computer use” access naturally raises the question of what else it might do to your system. In practice, current agentic models like Claude and Codex are built with strong guardrails against destructive actions such as deleting files, formatting drives, or exfiltrating personal data, and they tend to refuse those requests outright even when explicitly asked.
That doesn’t mean zero risk. The sensible approach is the same one that applies to any agentic tool: don’t instruct it to delete, remove, or overwrite files, monitor what it’s doing especially early on, and treat it the way you’d treat any automated process with real permissions on your machine. Trust it to build a Krea-to-video workflow. Don’t hand it root-level tasks you wouldn’t want a stranger performing on your computer.
Does the MCP work with local models only, or remote APIs too?
One of the more significant details is that ComfyUI’s template library already spans both local, open-source models and remote API-based models. Inside ComfyUI’s template browser, you can toggle between local ComfyUI models and external or remote APIs, and options like Grok Imagine and other hosted models show up alongside local ones like Miniax, LTX, and Krea.
That matters for MCP workflows specifically because it means the LLM isn’t limited to whatever fits on your GPU. A single workflow can mix a local image model with a remote video API, or vice versa, depending on what’s installed and available. New open-source model releases also tend to get added to ComfyUI’s template library quickly after launch.
Is ComfyUI’s MCP worth trying?
For anyone who has wanted ComfyUI’s flexibility without investing time in learning node graphs, the MCP substantially lowers the barrier. You still need to install ComfyUI itself and get comfortable handing a model some agentic control, but the actual graph-building work, previously the biggest obstacle for newcomers, is largely automatable now.
It’s not necessarily a way to skip learning ComfyUI entirely. Basic literacy (knowing what a sampler or scheduler does, how to adjust frame count or aspect ratio) still helps, if only so you’re not spending LLM tokens on trivial edits you could make yourself in seconds. But for building out multi-stage workflows, chaining models together, or working with template graphs that would otherwise take real study to understand, the MCP removes the single biggest reason people historically avoided ComfyUI.
Frequently Asked Questions
What’s the difference between ComfyUI’s official MCP and the community MCP?
The official MCP is ComfyUI’s own recently released version and is currently in beta. The community-built MCP has existed longer and has been tested more extensively by users. Both can be installed side by side, and switching between them for different tasks is a reasonable approach while the official version matures.
Do I need Claude specifically, or can I use another LLM?
Claude Desktop is a common choice, but Codex works as well, and in principle any LLM platform that supports agentic computer use could work with the MCP. The desktop version of whichever tool you choose is generally required, rather than a browser-based version.
What hardware do I need to run ComfyUI with MCP locally?
ComfyUI runs on most modern PCs and Macs, but for smoother generation, particularly video, an Nvidia GPU in the 4090 to 5090 range is recommended. Those without that hardware locally can use a rented cloud GPU or ComfyUI’s own cloud service instead.
Does using the MCP cost money beyond normal ComfyUI use?
Running models locally through ComfyUI doesn’t carry a per-generation cost. The main cost consideration is LLM token usage, since every instruction to Claude or Codex consumes tokens. Making small manual edits (prompt text, video length) directly in ComfyUI rather than asking the LLM to do it can help conserve tokens.
Can the MCP work with models that aren’t installed locally?
Yes. ComfyUI’s template system includes both local models and remote API-based models, so an MCP-built workflow can incorporate hosted models alongside anything running on your own machine, depending on what’s available and installed.