Zum Inhalt springen

MCP Server / comfy pilot

comfy pilot

MCP server + embedded terminal that lets Claude Code see and edit your ComfyUI workflows

168von @ConstantineB6MITGitHub →

Transport

stdiosse

Tools (13)

get_workflow

Get the current workflow from the browser

summarize_workflow

Human-readable workflow summary

get_node_types

Search available node types with filtering

get_node_info

Get detailed info about a specific node type

get_status

Queue status, system stats, and execution history

run

Run workflow (optionally up to a specific node) or interrupt

edit_graph

Batch create, delete, move, connect, and configure nodes

view_image

View images from Preview Image / Save Image nodes

search_custom_nodes

Search ComfyUI Manager registry for custom nodes

install_custom_node

Install a custom node from the registry

uninstall_custom_node

Uninstall a custom node

update_custom_node

Update a custom node to latest version

download_model

Download models from Hugging Face, CivitAI, or direct URLs

Dokumentation

Comfy Pilot

Talk to your ComfyUI workflows. Comfy Pilot gives Claude Code direct access to see, edit, and run your workflows — with an embedded terminal right inside ComfyUI.

Why?

Building ComfyUI workflows means manually searching for nodes, dragging connections, and tweaking values one at a time. With Comfy Pilot, you just describe what you want:

  • "Build me an SDXL workflow with ControlNet" — Claude creates all the nodes, connects them, and sets the parameters
  • "Look at the output and increase the detail" — Claude sees your generated image and adjusts the workflow
  • "Download the FLUX schnell model and set up a workflow for it" — Claude downloads the model and builds a workflow from scratch

No copy-pasting node names. No hunting through menus. Just say what you want.

Installation

CLI (Recommended):

comfy node install comfy-pilot

ComfyUI Manager:

  1. Open ComfyUI
  2. Click ManagerInstall Custom Nodes
  3. Search for "Comfy Pilot"
  4. Click Install
  5. Restart ComfyUI

Git Clone:

cd ~/Documents/ComfyUI/custom_nodes && git clone https://github.com/ConstantineB6/comfy-pilot.git

Claude Code CLI will be installed automatically if not found.

Requirements

  • ComfyUI
  • Python 3.8+

Features

  • MCP Server - Gives Claude Code direct access to view, edit, and run your ComfyUI workflows
  • Embedded Terminal - Full xterm.js terminal running Claude Code right inside ComfyUI
  • Image Viewing - Claude can see outputs from Preview Image and Save Image nodes
  • Graph Editing - Create, delete, move, and connect nodes programmatically

Demo

https://github.com/user-attachments/assets/325b1194-2334-48a1-94c3-86effd1fef02

Usage

  1. Restart ComfyUI after installation
  2. The floating Claude Code terminal appears in the top-right corner
  3. The MCP server is automatically configured for Claude Code
  4. Ask Claude to help with your workflow:
    • "What nodes are in my current workflow?"
    • "Add a KSampler node connected to my checkpoint loader"
    • "Look at the preview image and tell me what you see"
    • "Run the workflow up to node 5"

MCP Tools

The MCP server provides these tools to Claude Code:

| Tool | Description | |------|-------------| | get_workflow | Get the current workflow from the browser | | summarize_workflow | Human-readable workflow summary | | get_node_types | Search available node types with filtering | | get_node_info | Get detailed info about a specific node type | | get_status | Queue status, system stats, and execution history | | run | Run workflow (optionally up to a specific node) or interrupt | | edit_graph | Batch create, delete, move, connect, and configure nodes | | view_image | View images from Preview Image / Save Image nodes | | search_custom_nodes | Search ComfyUI Manager registry for custom nodes | | install_custom_node | Install a custom node from the registry | | uninstall_custom_node | Uninstall a custom node | | update_custom_node | Update a custom node to latest version | | download_model | Download models from Hugging Face, CivitAI, or direct URLs |

Example: Creating Nodes

Create a KSampler and connect it to my checkpoint loader

Claude will use edit_graph to:

  1. Create the KSampler node
  2. Connect the MODEL output from CheckpointLoader to KSampler's model input
  3. Position it appropriately in the graph

Example: Viewing Images

Look at the preview image and describe what you see

Claude will use view_image to fetch and analyze the image output.

Example: Downloading Models

Download the FLUX.1 schnell model for me

Claude will use download_model to download from Hugging Face to your ComfyUI models folder. Supports:

  • Hugging Face (including gated models with token auth)
  • CivitAI
  • Direct download URLs

Terminal Controls

  • Drag title bar to move
  • Drag bottom-right corner to resize
  • Minimize
  • × Close
  • Reconnect session

Architecture

┌─────────────────────────────────────────────────────┐
│  Browser (ComfyUI)                                  │
│  ┌─────────────────┐  ┌──────────────────────────┐  │
│  │  xterm.js       │  │  Workflow State          │  │
│  │  Terminal       │  │  (synced to backend)     │  │
│  └────────┬────────┘  └────────────┬─────────────┘  │
│           │ WebSocket              │ REST API       │
└───────────┼────────────────────────┼────────────────┘
            │                        │
            ▼                        ▼
┌─────────────────────────────────────────────────────┐
│  ComfyUI Server                                     │
│  ┌─────────────────┐  ┌──────────────────────────┐  │
│  │  PTY Process    │  │  Plugin Endpoints        │  │
│  │  (claude CLI)   │  │  /claude-code/*          │  │
│  └─────────────────┘  └──────────────────────────┘  │
└─────────────────────────────────────────────────────┘
            │                        │
            │                        ▼
            │           ┌──────────────────────────┐
            └──────────▶│  MCP Server              │
                        │  (stdio transport)       │
                        └──────────────────────────┘

Files

  • __init__.py - Plugin backend: WebSocket terminal, REST endpoints
  • js/claude-code.js - Frontend: xterm.js terminal, workflow sync
  • mcp_server.py - MCP server for Claude Code integration
  • CLAUDE.md - Instructions for Claude when working with ComfyUI

Troubleshooting

"Command 'claude' not found"

Install Claude Code CLI:

macOS / Linux / WSL:

curl -fsSL https://claude.ai/install.sh | bash

Windows (PowerShell):

irm https://claude.ai/install.ps1 | iex

Windows (CMD):

curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

MCP server not connecting

The plugin auto-configures MCP on startup. Check ComfyUI console for errors, or manually add to ~/.claude.json:

{
  "mcpServers": {
    "comfyui": {
      "command": "python3",
      "args": ["/path/to/comfy-pilot/mcp_server.py"]
    }
  }
}

Terminal disconnected

Click the ↻ button to reconnect, or check ComfyUI console for errors.

License

MIT

comfy pilot | hub.ai-engineering.at