
ComfyUI MCP Server (artokun): Setup for Claude Code, Cursor and Other Apps
Not the official ComfyUI MCP server. This one is built by the community. ComfyUI makes its own: Comfy MCP →
Lets Claude, Cursor and other assistants run ComfyUI for you: generate images, video and audio, edit workflows and manage models.
This community server drives a ComfyUI install on your own machine or a remote one. Your assistant can pick a model, build and run a workflow, look at the image and try again. It can also install custom nodes, download models and free GPU memory.
It runs through npx and needs Node 22 or newer, plus a running ComfyUI. It is not made by Comfy-Org. A few tools reach further than image generation, such as installing and removing node packs, starting and stopping ComfyUI, and renting RunPod GPUs, so read the safety notes.
Good for
- ComfyUI users who want to describe an image and let the assistant build the workflow
- Artists running parameter sweeps and comparing results
- Power users who manage models, custom nodes and VRAM from chat
Works in
12 should work
Tap an app to see how to set it up and any known problems.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Supports local MCP servers.
Client does not support this transport.
Client does not support this transport.
Install ComfyUI MCP (artokun)
Claude Code
- 1
Copy this prompt.
Add the ComfyUI MCP (artokun) server to Claude Code. Run this command: claude mcp add comfyui-artokun -- npx -y comfyui-mcp Then check that it loaded and tell me if I need to sign in or restart.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js 22+, A running ComfyUI (desktop or from source).
Read-only mode: No read-only mode is documented.
Start ComfyUI first (Node.js 22 or newer is required). For a remote ComfyUI, add a COMFYUI_URL environment variable with its address. CIVITAI_API_TOKEN is optional and only used for model downloads.
Built from the package data we hold. Not yet tested by us.
Tools
Generates images, video, audio or 3D from a prompt
Builds and runs a ComfyUI workflow from one line. Actions cover plain text-to-image, ControlNet and IP-Adapter guidance, video, audio, 3D, upscaling and background removal. It fills gaps from your saved defaults and picks a checkpoint for you.
Try: “Generate an image of a sunset over mountains.”
Write ·
generate_imageViews, lists and analyses generated images
Returns a generated image so the assistant can see it, lists recent outputs and assets, gives an image's palette and contrast, and shows how an asset was made.
Try: “Show me the last three images you made.”
Network ·
get_imageRuns a workflow you supply or a template
Submits a workflow and returns right away. It can also re-run a past generation, run a shared workflow from a URL, or run a bundled template with your overrides.
Try: “Run this workflow with a different seed.”
Write ·
enqueue_workflowShows and manages the ComfyUI job queue
Lists running and pending jobs, checks one job, and can reorder, edit or cancel pending jobs.
Try: “What is in the queue right now?”
Write ·
queueShows past runs, errors and good settings
Returns execution history with errors and timings, explains why a run failed, and suggests sampler, steps and CFG settings that worked before on your machine.
Try: “Why did my last run fail?”
Network ·
get_historyBuilds a workflow from a template or by editing
Creates workflows from templates such as txt2img, img2img, inpaint, upscale, ControlNet and audio, and edits them node by node.
Try: “Make an inpaint workflow I can run.”
Write ·
create_workflowReads saved workflows from your library
Lists the saved workflows, returns one by filename, or summarises what it does.
Try: “List my saved workflows.”
Files ·
get_workflowSaves a workflow to your library
Stores a workflow, overwriting a file with the same name, and can record a lock of the models and node versions it used.
Try: “Save this workflow as portrait-v2.”
Write ·
save_workflowDraws a workflow as a diagram
Renders a workflow you pass in as a Mermaid flowchart, and converts between workflow formats.
Try: “Draw a diagram of this workflow.”
Read ·
visualize_workflowReads or sets your generation defaults
Shows the merged defaults with where each came from, updates them, and can save to the config file. It can also read and write ComfyUI's own interface settings.
Try: “Set my default steps to 30.”
Write ·
get_defaultsCopies a local image into ComfyUI
Copies an image from your computer into ComfyUI's input folder for img2img, inpainting or ControlNet.
Try: “Upload ~/Pictures/pose.png for ControlNet.”
Write ·
upload_imageLists the models installed in ComfyUI
Shows installed checkpoints, LoRAs, VAEs, upscalers and more, and where ComfyUI looks for them.
Try: “Which checkpoints do I have?”
Network ·
list_local_modelsDownloads model files and tracks them
Fetches model weights from a URL into the right ComfyUI folder, checks transfer status, cancels, and searches HuggingFace and Civitai.
Try: “Download the SDXL base checkpoint.”
Write ·
download_modelEdits metadata inside model files
Reads and curates the metadata embedded in .safetensors model files.
Try: “Show the trigger words stored in this LoRA.”
Write ·
model_metadataFrees GPU memory
Unloads cached models through ComfyUI and reports memory before and after.
Try: “Free up VRAM.”
Write ·
clear_vramShows GPU, VRAM and server logs
Returns GPU, VRAM, Python and OS details, and can return ComfyUI's logs with a keyword filter.
Try: “How much VRAM is free?”
Network ·
get_system_statsSearches the ComfyUI Registry for node packs
Finds custom node packs by keyword, or returns one pack's details.
Try: “Find a node pack for face restoration.”
Network ·
search_custom_nodesFinds templates and dependencies for workflows
Lists custom-node workflow templates, checks whether a workflow runs locally or spends paid API credits, works out which node packs a workflow needs and installs missing ones, and can write a skill file from a pack.
Try: “What node packs does this workflow need?”
Write ·
list_packsInstalls, repairs or removes node packs
Installs, repairs, enables, disables and removes custom node packs on your ComfyUI. Removal deletes the pack.
Try: “Install ComfyUI-Impact-Pack.”
Delete ·
install_custom_nodeSaves and restores custom-node snapshots
Saves or restores the set of installed custom nodes through ComfyUI-Manager, like comfy node save-snapshot.
Try: “Snapshot my current nodes before I update.”
Runs code ·
node_snapshotFinds which custom node breaks things
Disables nodes in halves, git-bisect style, to find the one causing a problem. It changes which nodes are enabled while it runs.
Try: “ComfyUI crashes on start. Find the bad node.”
Runs code ·
bisectWrites and publishes your own node pack
Creates, edits, tests and publishes your own custom-node pack inside the custom_nodes folder. File actions stay inside that folder. It works on the local machine only.
Try: “Create a simple node pack that crops images.”
Delete ·
node_packSets up ComfyUI from a manifest file
Applies a setup manifest: installs custom nodes, downloads models and installs pip packages. System packages are reported as skipped.
Try: “Apply setup.yaml.”
Runs code ·
apply_manifestRuns the official comfy-cli tool
Wraps the official comfy command: status, start, stop and restart the server, job control, node search and more. It needs comfy-cli 1.11.1 or newer.
Try: “Search loaded nodes for upscale.”
Runs code ·
comfy_cliStops or restarts ComfyUI
Stops the running ComfyUI process, starts it again with the same launch arguments, or starts it from saved launch info.
Try: “Restart ComfyUI.”
Runs code ·
restart_comfyuiInstalls and configures ComfyUI itself
Installs, updates and configures the local ComfyUI, its sidebar panel and this MCP server.
Try: “Install ComfyUI on this machine.”
Runs code ·
install_comfyuiInspects and manages ComfyUI workspaces
Inspects and manages workspaces for local ComfyUI installs.
Try: “List my ComfyUI workspaces.”
Write ·
workspaceRuns many workflows under one batch
Runs many workflows under a durable batch ID, so a parameter sweep or a long list can be tracked and resumed.
Try: “Run this workflow for five different prompts.”
Write ·
batchManages micro-apps in the ComfyUI panel
Lists and runs named workflows packaged as one-click apps in the panel.
Try: “Run my portrait app.”
Write ·
appsFinds and runs paid partner API nodes
Discovers hosted partner nodes such as Flux, Ideogram and Kling and runs them. They spend paid credits from your Comfy account, unlike a local GPU render.
Try: “Which hosted image nodes can I use?”
Write ·
list_api_nodesChecks whether comfy-kitchen can speed up a graph
Reports what comfy-kitchen can do on this GPU, finds graph parts it could speed up, and applies the faster path.
Try: “Can my GPU run this graph faster?”
Write ·
kitchenDeploys and controls RunPod GPU pods
Deploys, starts, stops, inspects and connects to RunPod cloud GPU pods and switches rendering between your machine and a pod. The tool says it spends money.
Try: “Deploy a RunPod pod and render there.”
Delete ·
runpodWatches a RunPod pod's status
Shows a pod's live status in the panel, stops watching, or diagnoses why a pod is not usable. One action can stop a pod.
Try: “Why is my pod not ready?”
Write ·
runpod_watchChecks the LoRA training setup
Checks and sets up the training machinery: Docker, GPU and the Python environment. It touches no datasets or jobs.
Try: “Is my machine ready to train a LoRA?”
Runs code ·
train_doctorPrepares images and captions for LoRA training
Stages and curates the images and captions a training run uses, keyed by dataset name.
Try: “Prepare my product photos as a dataset.”
Write ·
train_prepare_datasetStarts, checks and stops LoRA training jobs
Launches a training run, polls it, stops it, deletes it, and reads back its settings.
Try: “Train a LoRA on the product dataset.”
Delete ·
train_startUsed by the assistant behind the scenes
Does exact math
Evaluates a batch of math expressions exactly, without ComfyUI. It handles numbers only.
Try: “What is 1024 * 1.5 rounded to a multiple of 64?”
Read ·
calculateDisabled issue reporting stub
Kept for older prompts. It files nothing and contacts no service, and only returns the reporting status.
Try: “(Not used.)”
Read ·
report_issueLists all tools in a compact catalog
Part of a built-in facade that lists every tool, so clients with a stale tool list can still reach them.
Try: “List the available tools.”
Read ·
list_toolsDescribes one tool's inputs
Returns the details of one tool in the catalog.
Try: “Describe the queue tool.”
Read ·
describe_toolCalls any tool by name
Runs any catalog tool by name with its arguments, the same as calling it directly.
Try: “Call the queue tool with action list.”
Runs code ·
call_tool
Read from the source code
Permissions
high risk · 4 things to watch
High risk
The assistant can generate files, install and remove ComfyUI node packs, restart ComfyUI and start paid cloud GPU pods. Approve installs, deletes and anything that costs money by hand.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It can install and remove software
Tools can install and delete custom node packs, run pip installs from a manifest and run training jobs. Approve those calls by hand. github.com
It can spend money
The RunPod tool says it spends money, and partner API nodes spend paid Comfy credits. Check costs before approving. github.com
It can stop and restart ComfyUI
It starts, stops and restarts the ComfyUI process, and can run the comfy-cli tool. Save your work first. github.com
The remote option opens a public address
The tunnel option makes a public URL protected by a token. Keep that token private and use it only when you need it. github.com
Security advisories
No published security advisories as of Nov 1, 2026.
Technical details
| Turning on read-only | No read-only mode is documented. The README describes an operator setting that limits which tools appear, but not a read-only preset. |
|---|---|
| Talks to | Your ComfyUI server. HuggingFace, Civitai and the node registry for downloads and searches. RunPod, if you use its tools |
| Sign-in and keys | ComfyUI needs no sign-in. Optional tokens are read from environment variables in your client config. |
| Scopes |
|
Questions
Is this the official ComfyUI MCP server?
No. It is a community project by artokun. Comfy-Org ships its own agent tools separately, including a hosted Comfy Cloud connector.
What do I need before installing it?
A running ComfyUI (desktop or from source) and Node.js 22 or newer. Optionally add a Civitai token for model downloads.
Is the tool list complete?
It lists the 38 tools the published package registers plus a 3-tool catalog facade. Autoloaded workflows can add more tools, and many tools take an action setting that covers several jobs.
Can it use a remote ComfyUI?
Yes. Set COMFYUI_URL to another server's address, for RunPod, a VPS or a LAN machine. The README also covers a tunnel mode for Claude Desktop custom connectors.
Can it spend money or delete things?
Yes. The RunPod tool and partner API nodes cost money, and it can remove custom node packs and delete training jobs.
Is it free?
The package is MIT licensed and free. Local rendering uses your own GPU. Hosted partner nodes and RunPod charge you directly.
Does it work with a local model through Ollama?
The README describes a panel agent that can use Ollama or hosted models. That is a separate part of the project from the MCP server.
Information
updated Nov 2, 2026
- Publisher
- Community
- Package
comfyui-mcp- Registry
- npm
- Transport
- stdio
- Auth
- None
- Version
- 0.52.206
- License
- MIT
- Last commit
- Oct 26, 2026
- Listed in
- Official MCP registry
Other ComfyUI MCP servers
Sources
Docs, issue trackers and discussions this page is built from. Issues, quotes and findings above link to their own source.
Show all 2Hide
- github.com · github.com
- www.npmjs.com · npmjs.com