Comfy Cloud MCP

Closed Beta — invite only. Comfy Cloud MCP is currently in a closed beta with per-user feature flag gating. Features, tools, and behavior may change as the project evolves. If you don’t have access yet, sign up for the waitlist.

Comfy Cloud MCP connects AI assistants — including Claude Desktop, Claude Code, Cursor, and Amp — to Comfy Cloud via the Model Context Protocol (MCP). It lets AI agents generate images, video, audio, and 3D content, search models and nodes, and run full ComfyUI workflows on cloud GPUs, with no local GPU required.

Quick Start

Before you start, make sure you have:

An active Comfy Cloud subscription (required to submit workflows)
Your email enabled for the closed beta (see note above)
A way to sign in: OAuth (recommended where supported) or a Comfy API key (comfyui-…)

OAuth (no API key to paste) — In Claude Code or Claude Desktop, add the remote server https://cloud.comfy.org/mcp and use the client’s Authenticate control to sign in with your Comfy account. Partner generation and workflows use the same tools once connected. API key — Still fully supported. Use the one-click installer or manual setup below if your client does not offer OAuth, or if you prefer a key.

OAuth rollout. The MCP server already accepts OAuth on production, but sign-in may not work end-to-end for everyone until the next Comfy Cloud production deployment finishes (the discovery endpoint must catch up). Until OAuth works for you, use an API key — existing API-key setups are unchanged.

If you use an API key and do not have one yet, see the guide to get one:

Getting an API Key

Learn how to create and manage your Comfy Platform API Key for accessing Cloud APIs, Partner Nodes, and more.

Install & Connect (API key)

The one-click installer detects your MCP client (Claude Code, Cursor, Amp), asks for your Comfy API key, and configures the remote MCP server — no Node.js or other dependencies required.

macOS / Linux
Windows (PowerShell)

curl -fsSL https://raw.githubusercontent.com/Comfy-Org/comfy-cloud-mcp/main/install.sh | bash

irm https://raw.githubusercontent.com/Comfy-Org/comfy-cloud-mcp/main/install.ps1 | iex

Manual setup with API key

Comfy Cloud MCP is hosted at https://cloud.comfy.org/mcp. Point your MCP client at that URL with your API key:

Claude Code
Claude Desktop
Cursor

claude mcp add comfyui-cloud \
  --transport http \
  https://cloud.comfy.org/mcp \
  -H "X-API-Key: your-api-key-here"

Add the following to your Claude Desktop config file:

OS	Config path
macOS	`~/Library/Application Support/Claude/claude_desktop_config.json`
Windows	`%APPDATA%\Claude\claude_desktop_config.json`
Linux	`~/.config/Claude/claude_desktop_config.json`

{
  "mcpServers": {
    "comfyui-cloud": {
      "url": "https://cloud.comfy.org/mcp",
      "headers": {
        "X-API-Key": "your-api-key-here"
      }
    }
  }
}

Add the following to your Cursor MCP config (.cursor/mcp.json in your project or global config):

{
  "mcpServers": {
    "comfyui-cloud": {
      "url": "https://cloud.comfy.org/mcp",
      "headers": {
        "X-API-Key": "your-api-key-here"
      }
    }
  }
}

See Cursor’s MCP documentation for more details.

After adding the server, restart your MCP client to pick up the new configuration.

Usage

Just ask. The agent picks the right tools on its own — “generate a cat astronaut in space”, “find SDXL checkpoints”, “upscale this”.

How generation is routed

The agent chooses between two paths:

Path	Tool	Best for
Partner APIs	`partner_generate`	Named providers (Flux, Grok, Nano Banana, Ideogram, GPT-image, Seedream, etc.)
Custom workflows	`submit_workflow`	Open-source models, LoRA/ControlNet, multi-step graphs, or anything not covered by `partner_generate`

For partner providers, partner_generate always tries a real Comfy Cloud workflow run first. The output is saved to your asset library with the workflow embedded so you can reopen it in the editor. The tool returns a prompt_id — poll with get_job_status, then download with get_output (same as any other workflow). For image-input requests or models not yet on the persistable set, the server falls back automatically to a direct partner-proxy response with instant download URLs. That fallback is internal; you cannot opt into or out of it. Those direct results are not saved to the asset library.

Available Tools

Workflow execution

Tool	Description
`submit_workflow`	Submit a ComfyUI API-format workflow for execution on Comfy Cloud
`get_job_status`	Poll execution status of a submitted workflow
`get_output`	Retrieve output images, videos, or audio from a completed workflow
`cancel_job`	Cancel a pending or running job
`get_queue`	Check how many jobs are running and pending

Partner API generation

Tool	Description
`partner_generate`	Generate with partner APIs (Flux, Nano Banana, Grok, GPT-image, Ideogram, Seedream, etc.). Runs through a Comfy Cloud workflow by default; returns `prompt_id` and saves to your asset library. Poll `get_job_status` / `get_output` like any workflow. Unsupported cases fall back to instant proxy URLs automatically.

Discovery

Tool	Description
`search_templates`	Search pre-built workflow templates from comfy.org by text, tag, media type, or model
`search_models`	Search model catalog by text, type, base model, or source
`search_nodes`	Search available nodes by text, category, or input/output types
`cql`	Run Comfy Query Language queries for advanced discovery

Inputs & Chaining

Tool	Description
`upload_file`	Upload an input image or file for use in workflows (e.g. LoadImage)
`use_previous_output`	Chain workflows by reusing the output of one as input to another

Saved Workflows

Tool	Description
`list_saved_workflows`	Browse your saved workflows from Comfy Cloud
`get_saved_workflow`	Inspect a saved workflow’s nodes, inputs, and configuration
`save_workflow`	Save a workflow to your workspace (save or API format; API format is converted automatically)
`run_saved_workflow`	Fetch a saved workflow, convert to API format, and submit it for execution

Feedback (Beta)

Tool	Description
`submit_feedback`	Submit feedback with a rating and comment
`report_session_summary`	Report a consent-gated session summary (no prompts, file paths, or PII)

The server checks for a matching pre-built template before building a workflow from scratch, leading to better results and faster generation.

Feedback

This is a closed beta — please tell us what’s working.

In-agent: ask the agent to call submit_feedback (rating + comment) or report_session_summary (consent-gated session summary; no prompts/file paths/PII).
Survey: links.comfy.org/cloudmcpbeta
Issues: file in this repo

Tool errors include a once-per-session pointer back to these channels so you don’t have to remember them.

How It Works

┌──────────────┐    HTTPS/MCP    ┌─────────────────────────────────────────────┐
│  AI Agent    │◄───────────────►│            Comfy Cloud                      │
│  (Claude,    │  API key or     │  cloud.comfy.org/mcp → Workflow execution   │
│   Cursor,    │  OAuth bearer   │                        on Cloud GPUs        │
│   Amp)       │                 │                                             │
└──────────────┘                 └─────────────────────────────────────────────┘

Your AI agent connects directly to the hosted MCP server at cloud.comfy.org/mcp. The server translates MCP tool calls into workflow executions on Comfy Cloud GPUs — no local server or GPU required. For partner models, the agent typically calls partner_generate first (workflow + library save). For everything else it uses discovery tools (search_templates, search_nodes, cql), builds ComfyUI API-format workflow JSON, and submits via submit_workflow — just describe what you want in natural language.

Example Prompts

After installing, try these prompts in your AI assistant:

Generate an image of a cat astronaut floating in space, cartoon style

Find me workflow templates for text-to-video generation

Search for SDXL checkpoint models and tell me which ones are available

Generate a fantasy landscape with mountains and a glowing river at sunset

What nodes are available for image upscaling?

The agent will search for a matching template, build the ComfyUI workflow, submit it to Comfy Cloud, and return the generated image directly in the conversation.

Output Handling

Images are displayed inline in the conversation (Claude Code) or in the artifact side panel (Claude Desktop)
Video and audio outputs are returned as downloadable links
Animated images (GIF, WebP) are saved but not previewed inline to preserve animation
Inline image previews are resized to 1024px for efficient display; full-resolution outputs are always available via the get_output tool

Known Limitations

This is an early release. The following limitations are known and being actively worked on.

Workflows

run_saved_workflow coverage. The tool fetches a saved workflow, converts save format → API format, and submits it. Complex input overrides or workflows with unsupported nodes may still need manual fixes or a rebuild via submit_workflow.
Workflow metadata varies by path. Outputs from partner_generate (workflow path) land in the asset library with the graph embedded for reopening in the editor. Outputs from submit_workflow or the direct partner_generate fallback may not include full workflow metadata in the asset library.
Workflow accuracy depends on the AI. The agent constructs ComfyUI workflows from natural language. Complex multi-node workflows or unusual node configurations may require iteration.
partner_generate scope. Text-to-image on the persistable model set runs as a saved workflow. Image-input generations, unmapped models, and non-image modalities use the direct proxy fallback (instant URL, not library-persisted) until workflow persist support expands.

File Handling

Upload size limits may apply depending on your MCP client.
Image previews are resized. Inline previews are limited to 1024px (JPEG). Full-resolution files are saved to disk.

Authentication

API key or OAuth. Manual setup uses a Comfy Cloud API key in the X-API-Key header. MCP clients that sign in with Comfy OAuth send an Authorization: Bearer token instead; both are supported on the hosted server.

Client-Specific

Claude Desktop — Generated images appear in the artifact side panel via HTML, not as native image artifacts.

Troubleshooting

MCP server not showing up

Restart your MCP client (close and reopen Claude Code, Claude Desktop, Cursor, or Amp). MCP servers are loaded at startup. Double-check that the server URL is exactly https://cloud.comfy.org/mcp in your config.

API key errors

Verify your API key is valid at platform.comfy.org/profile/api-keys. For manual config, pass the key in the X-API-Key header — do not send a comfyui- key as a Bearer token. OAuth-connected clients should use Authorization: Bearer with the token from the Comfy sign-in flow, not X-API-Key. Generate a new API key if needed and update your client config.

Connection errors

If the MCP client can’t reach the server, check that:

You have an active internet connection
Your firewall or proxy isn’t blocking cloud.comfy.org
You have an active Comfy Cloud subscription

Documentation Index

​Quick Start

​Sign in with OAuth or API key

Getting an API Key

​Install & Connect (API key)

​Manual setup with API key

​Usage

​How generation is routed

​Available Tools

​Workflow execution

​Partner API generation

​Discovery

​Inputs & Chaining

​Saved Workflows

​Feedback (Beta)

​Feedback

​How It Works

​Example Prompts

​Output Handling

​Known Limitations

​Troubleshooting

​MCP server not showing up

​API key errors

​Connection errors

Quick Start

Sign in with OAuth or API key

Install & Connect (API key)

Manual setup with API key

Usage

How generation is routed

Available Tools

Workflow execution

Partner API generation

Discovery

Inputs & Chaining

Saved Workflows

Feedback (Beta)

Feedback

How It Works

Example Prompts

Output Handling

Known Limitations

Troubleshooting

MCP server not showing up

API key errors

Connection errors