Modal ComfyUI Runner

⚠️ Work in Progress - This repository is under active development and not ready for production use.

A serverless ComfyUI deployment system that runs workflows on Modal and exposes them via MCP (Model Context Protocol) servers for easy integration with AI assistants.

Includes:

• Serverless ComfyUI: Deploy workflows with automatic scaling and GPU optimization
• Memory Snapshots: Fast cold starts using Modal's memory snapshot feature
• Workspace Management: Organize workflows, models, and custom nodes by workspace
• MCP Integration: Expose workflows as tools for AI assistants
• Parameter Injection: Dynamic workflow customization via YAML configuration

Architecture

Core Components

• modal_comfy.py: Main deployment script creating Modal apps with ComfyUI
• comfyclient.py: Client for testing deployed workflows
• deploy_constants.py: GPU, timeout, and resource configuration
• mcp/: MCP server implementation for AI assistant integration

Quick Start

1. Deploy a Workspace

# Set workspace and deploy to Modal
WORKSPACE=slow_new modal deploy modal_comfy.py

2. Test Deployment

# Test a specific workflow
python comfyclient.py \
  --modal-workspace edenartlab \
  --workspace_name slow-new-stage \
  --workflow txt2img \
  --test-json workspaces/slow_new/workflows/txt2img/test.json

3. Interactive Mode

# Serve ComfyUI as interactive endpoint accessible through browser:
WORKSPACE=slow_new modal serve modal_comfy.py

MCP Server Integration

The MCP server exposes ComfyUI workflows as tools for AI assistants.

Local Development

# Run MCP server locally
cd mcp/
python mcp_server.py

# Test with client
python mcp_client.py --local

Deploy MCP Server

# Deploy to Modal
modal deploy mcp/mcp_server.py

# Test remote deployment
python mcp_client.py --remote https://your-modal-url/mcp

Using with Claude Desktop

Add to your Claude Desktop config:

{
  "mcpServers": {
    "comfyui": {
      "command": "python",
      "args": ["/path/to/mcp_server.py"],
      "env": {
        "MODAL_WORKSPACE": "your-workspace",
        "WORKSPACE_NAME": "your-workspace-name"
      }
    }
  }
}

TODO:

• fix custom_node installations (comfy-cli vs git clone ... )
• correctly parse all workflow outputs and stream them to client
• implement all custom arg injection types (folder, array, ...)