ComfyUI-MCP

Model Context Protocol (MCP) integration for ComfyUI, enabling workflows to connect to and interact with any MCP-compatible server.

Features

• 🔌 Connect to MCP Servers: Support for both local (stdio) and remote (SSE) MCP servers
• 🛠️ Execute Tools: Run any tool exposed by MCP servers with automatic type conversion
• 📁 Access Resources: Read resources from MCP servers with support for text, JSON, and binary data
• 🔍 Discovery: Browse available tools and resources with detailed schema information
• 🔄 Lifecycle Management: Automatic server start/stop with workflow execution
• 🔁 Retry Logic: Built-in exponential backoff for resilient connections
• 📊 Progress Tracking: Real-time progress updates during tool execution
• 🗃️ Result Caching: Leverages ComfyUI's caching system to avoid redundant calls

Installation

1. Navigate to your ComfyUI custom nodes directory:

   cd ComfyUI/custom_nodes/

2. Clone or copy the ComfyUI-MCP folder

3. Install dependencies:

   cd ComfyUI-MCP
   pip install -r requirements.txt

4. Restart ComfyUI

Nodes

MCP Server Connect

Connects to an MCP server and manages its lifecycle.

Inputs:

• server_id: Unique identifier for the connection
• transport: Connection type (stdio or sse)
• connection_string: Command for stdio or URL for sse
• action: connect, disconnect, or status
• command_args: Additional command arguments
• env_vars: Environment variables (KEY=VALUE format)
• capture_output: Capture server stdout/stderr

Outputs:

• server_id: Server identifier
• status: Connection status
• info: Detailed server information

MCP Tool Execute

Executes tools on connected MCP servers.

Inputs:

• server_id: Server to execute tool on
• tool_name: Name of the tool to run
• arguments: Tool arguments as JSON
• retry_count: Number of retry attempts
• retry_delay: Initial retry delay

Outputs:

• result: Tool execution result
• success: Whether execution succeeded
• error: Error message if failed

MCP Resource Read

Reads resources from MCP servers.

Inputs:

• server_id: Server to read from
• resource_uri: URI of the resource
• output_format: Format (text, json, base64)

Outputs:

• content: Resource content
• mime_type: Content MIME type
• success: Whether read succeeded
• error: Error message if failed

MCP Discover Tools/Resources

Lists available tools and resources from a server.

Inputs:

• server_id: Server to query
• discovery_type: What to discover (tools, resources, both)
• output_format: Output format (summary, detailed, names_only)

Outputs:

• discovery_info: Formatted discovery information
• tools_count: Number of available tools
• resources_count: Number of available resources
• status: Discovery status

Example Workflows

Local File System Server

1. Add "MCP Server Connect" node
2. Set connection_string to: npx -y @modelcontextprotocol/server-filesystem /path/to/directory
3. Connect to "MCP Tool Execute" node
4. Use tool "read_file" with arguments: {"path": "file.txt"}

GitHub Integration

1. Add "MCP Server Connect" node
2. Set connection_string to: npx -y @modelcontextprotocol/server-github
3. Set env_vars to: GITHUB_TOKEN=your_token
4. Use discovery node to see available tools

Remote MCP Server

1. Add "MCP Server Connect" node
2. Set transport to: sse
3. Set connection_string to server URL
4. Add authentication token if required

Authentication

For servers requiring authentication:

• Local servers: Use env_vars to pass tokens
• Remote servers: ComfyUI will use MCP_AUTH_TOKEN from environment

Debugging

Enable output capture to see server logs:

• Set capture_output to True in server node
• Check ComfyUI console for stdout/stderr
• Use status action to check connection state

Limitations

• WebSocket transport not yet implemented
• Dynamic tool discovery (UI updates) not supported
• Limited to JSON-serializable data types

Development

To extend or modify:

1. Follow ComfyUI node development guidelines
2. Maintain compatibility with MCP protocol spec
3. Test with various MCP server implementations

License

Same as ComfyUI