ComfyMCP

Give Claude the ability to generate images with ComfyUI. Just ask for what you want in natural language.

You: "Generate an image of a robot painting a sunset"

Claude: I'll create that image for you.
        [builds 7-node workflow, executes it]
        Done! Generated robot_painting_00001.png in 2.3 seconds.

What You Can Ask

Once installed, Claude can handle requests like:

Image Generation

• "Generate an image of a cat astronaut floating in space"
• "Create a 1024x1024 fantasy landscape using SDXL"
• "Make a portrait with negative prompt 'blurry, low quality'"

Model & System Info

• "What checkpoint models do I have?"
• "Show me the available samplers"
• "What's my GPU memory usage?"

Workflow Control

• "Use 30 steps instead of 20 for better quality"
• "Generate 4 variations with different seeds"
• "What's the status of my last generation?"

Claude handles all the complexity—discovering nodes, building connections, validating the workflow, and monitoring execution.

How It Works

When you ask Claude to generate an image, it builds a complete ComfyUI workflow:

[1] CheckpointLoaderSimple ─────────────────────────────┐
     ├── MODEL ──────────────────────────────────────────┤
     ├── CLIP ───┬──→ [3] CLIPTextEncode (positive) ────┤
     │           └──→ [4] CLIPTextEncode (negative) ────┤
     └── VAE ────────────────────────────────────────────┤
                                                         ▼
[2] EmptyLatentImage ──────────────────────────→ [5] KSampler
                                                         │
                                                         ▼
                                                 [6] VAEDecode
                                                         │
                                                         ▼
                                                 [7] SaveImage

This happens automatically. Claude:

1. Discovers available nodes and their inputs/outputs
2. Builds the workflow with proper connections
3. Validates everything before execution
4. Queues the job and monitors completion
5. Reports the output filename

Installation

Prerequisites

• ComfyUI running (default: localhost:8188)
• uv package manager

# Install uv if needed
curl -LsSf https://astral.sh/uv/install.sh | sh

Claude Code (CLI)

claude mcp add comfyui \
  --transport stdio \
  --env COMFYUI_HOST=127.0.0.1 \
  --env COMFYUI_PORT=8188 \
  -- uvx --from git+https://github.com/hernantech/comfymcp comfymcp

Claude Desktop

Add to your config file:

• Linux: ~/.config/claude/claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "comfyui": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/hernantech/comfymcp", "comfymcp"],
      "env": {
        "COMFYUI_HOST": "127.0.0.1",
        "COMFYUI_PORT": "8188"
      }
    }
  }
}

Verify Installation

Ask Claude: "Check if ComfyUI is connected"

You should see confirmation that the server is online with GPU info.

Configuration

For remote ComfyUI servers, update the host:

claude mcp add comfyui \
  --env COMFYUI_HOST=192.168.1.100 \
  ...

Reference

Available MCP Tools

<details> <summary><strong>Workflow Execution</strong></summary>

</details>

<details> <summary><strong>Workflow Building</strong></summary>

</details>

<details> <summary><strong>Assets & Models</strong></summary>

</details>

<details> <summary><strong>System</strong></summary>

</details>

MCP Resources

Python API

For programmatic use outside of MCP:

from comfymcp.workflow import WorkflowBuilder

builder = WorkflowBuilder()

# Nodes return refs with named outputs
checkpoint = builder.add_node("CheckpointLoaderSimple",
    ckpt_name="sd_turbo.safetensors")

latent = builder.add_node("EmptyLatentImage",
    width=512, height=512, batch_size=1)

positive = builder.add_node("CLIPTextEncode",
    clip=checkpoint.CLIP,  # Named output connection
    text="a beautiful sunset")

negative = builder.add_node("CLIPTextEncode",
    clip=checkpoint.CLIP,
    text="ugly, blurry")

sampler = builder.add_node("KSampler",
    model=checkpoint.MODEL,
    positive=positive.CONDITIONING,
    negative=negative.CONDITIONING,
    latent_image=latent.LATENT,
    seed=42, steps=4, cfg=1.0,
    sampler_name="euler", scheduler="normal", denoise=1.0)

decode = builder.add_node("VAEDecode",
    samples=sampler.LATENT,
    vae=checkpoint.VAE)

builder.add_node("SaveImage",
    images=decode.IMAGE,
    filename_prefix="output")

workflow = builder.build()

Templates

from comfymcp.templates import Text2ImgTemplate, Img2ImgTemplate

# Text to image
txt2img = Text2ImgTemplate(
    checkpoint="sd_turbo.safetensors",
    positive_prompt="a majestic mountain",
    negative_prompt="ugly, blurry",
    width=512, height=512,
    steps=4, cfg=1.0
)
workflow = txt2img.build()

# Image to image
img2img = Img2ImgTemplate(
    checkpoint="sd_turbo.safetensors",
    image="input.png",
    positive_prompt="enhance details",
    denoise=0.6
)
workflow = img2img.build()

Direct Client Usage

from comfymcp.client import ComfyUIClient

async with ComfyUIClient(host="127.0.0.1", port=8188) as client:
    # Queue workflow
    result = await client.queue_prompt(workflow)

    # Check status
    history = await client.get_history(prompt_id=result.prompt_id)

    # List models
    checkpoints = await client.get_models("checkpoints")

Requirements

• Python 3.10+
• ComfyUI server running
• MCP-compatible client (Claude Code, Claude Desktop, Cursor, etc.)

License

MIT License - see LICENSE for details.