comfy-pilot

Comfy Pilot exposes an MCP server that lets Claude Code interact directly with your ComfyUI workflows. Through the MCP tools, Claude can inspect your current workflow, summarize it, query node types and information, manage the graph (create, delete, move, connect nodes), run the workflow (optionally up to a specific node), and even view images produced by nodes like Preview Image or Save Image. It also supports downloading models and managing custom nodes from the ComfyUI registry. To use it, ensure ComfyUI is installed and the Comfy Pilot plugin is set up. The MCP server is automatically configured for Claude Code on startup, and you can access the floating Claude Code terminal to issue commands like getting the current workflow or instructing Claude to add or connect nodes.

Prerequisites:

• Python 3.8+ (as required by ComfyUI and Comfy Pilot)
• ComfyUI installed and functional

Installation steps:

1. Ensure Python is available in your PATH and verify with: python3 --version
2. Clone the Comfy Pilot repository or obtain the MCP server files: git clone https://github.com/ConstantineB6/comfy-pilot.git
3. Install required Python dependencies (if a requirements.txt exists in the repo): cd comfy-pilot python -m pip install -r requirements.txt
4. Start the MCP server: python /path/to/comfy-pilot/mcp_server.py Note: Replace /path/to/comfy-pilot with the actual path where the repo is located.
5. Verify the server is running and accessible from the Claude Code MCP interface. If needed, configure your mcp.json or claude.json equivalents to point to the server using the provided example in Troubleshooting under MCP server not connecting.
6. (Optional) Install Comfy Pilot as a node within ComfyUI if you haven’t already, following the README installation steps for CLI, Manager, or Git Clone methods.

Tips and notes:

• The MCP server is designed to give Claude Code direct access to view, edit, and run ComfyUI workflows, including an embedded terminal interface inside ComfyUI.
• If the server fails to connect, ensure the Python path and mcp_server.py path are correct, and check for any firewall or network restrictions.
• The README mentions a sample configuration for manual MCP setup in ~/.claude.json; you can adapt this if your environment requires explicit MCP server registration.
• Ensure any required Claude Code CLI is installed if you rely on automated installation prompts.
• When editing graphs or running workflows, Claude may need appropriate permissions to modify nodes and download models; ensure paths and credentials (for gated models) are correctly configured.
• If you encounter terminal disconnections, you can use the reconnect option in the MCP UI to reestablish the session.