Dify as MCP Server

What is Dify as MCP Server?

Dify as MCP Server is a Dify plugin that exposes Dify workflows as a Model Context Protocol (MCP) server for AI clients like Claude Desktop and Cursor. It allows AI assistants to discover and invoke Dify workflows as tools without modifying the workflows themselves.

How to use Dify as MCP Server?

Install the plugin in Dify, configure the plugin settings by providing the application ID of the workflow to expose. Then in your MCP-compatible client (such as Claude Desktop or Cursor), add a new MCP server with the URL https://<your-dify-instance>/difyapp_as_mcp_server, save, and enable the server.

Key features of Dify as MCP Server

• Exposes Dify workflows as MCP tools to AI clients
• Supports JSON-RPC over SSE and Streamable HTTP
• Automatic tool discovery and registration from workflows
• Compatible with Claude Desktop, Cursor, and other MCP clients
• Server-side session management with heartbeat (≈5 minutes)
• Implements the latest MCP Streamable HTTP specification

Use cases of Dify as MCP Server

• Enable Claude to call Dify workflows as built-in tools
• Integrate custom AI assistants with any Dify automation pipeline
• Provide tool access to Cursor agents during development
• Deploy a stateless MCP server in a Dify plugin environment

FAQ from Dify as MCP Server

What AI clients are supported?

Claude Desktop, Cursor, and any MCP-compatible client that supports SSE or Streamable HTTP.

How do I configure the plugin?

Set the application ID of the Dify workflow you want to expose. Ensure the workflow is published and the Dify API key has sufficient permissions.

Does it require a persistent connection?

No. The plugin supports a stateless server mode where each request is independent and associated via session ID, suitable for Dify’s stateless API environment.

How do I troubleshoot tool visibility?

Verify the server URL is correct and accessible, the app ID is configured correctly, the workflow is published, and the Dify API key has appropriate permissions.

What transport does the server use?

It uses a GET endpoint for SSE connections and HTML pages, and a POST endpoint for JSON-RPC requests, following the MCP Streamable HTTP spec.