dify-mcp-client

What is dify-mcp-client?

dify-mcp-client is a Dify plugin that acts as an MCP (Model Context Protocol) Host, connecting MCP servers into Dify workflows as a ReAct Agent strategy. It is built for developers using Dify who want to integrate external MCP-compatible tools, resources, and prompts.

How to use dify-mcp-client?

Install the plugin from GitHub or from a .difypkg release file via Dify's plugin manager. After installation, add an Agent node in any Dify Chatflow and select "mcpReAct" as the agent strategy. Configure the config_json field with server URLs (SSE or Streamable HTTP) to connect MCP servers.

Key features of dify-mcp-client

• Converts MCP tools, resources, and prompts into Dify tools.
• Supports SSE, Streamable HTTP, and multiple MCP servers.
• Integrates UI-TARS-SDK for GUI automation capabilities.
• Allows user-configurable maximum loop count for GUI actions.
• Provides Chatflow test examples and DSL (.yml) imports.

Use cases of dify-mcp-client

• Connect an arxiv MCP server to let an LLM fetch papers autonomously.
• Build a Chatflow that uses MCP tools to interact with local or remote services.
• Automate GUI actions by combining UI-TARS vision with MCP tool calling in a Dify workflow.

FAQ from dify-mcp-client

What transport protocols does dify-mcp-client support?

It supports SSE (Server-Sent Events) and Streamable HTTP transports for MCP servers.

Can I use TypeScript stdio MCP servers with this plugin?

Yes, but only if you use the Docker image with pre-installed Node.js inside the plugin-daemon, or deploy from source code (not via .difypkg). Alternatively, convert stdio servers to Streamable HTTP using mcp-proxy.

Does dify-mcp-client include a human-in-the-loop mechanism?

No, it does not implement human-in-the-loop by default. The README recommends connecting only reliable MCP servers and suggests setting max iterations to 1 or adding a permission-asking phrase in the INSTRUCTION field.

Where is the data stored and processed?

Data is processed within the Dify plugin runtime; configuration (server URLs) is provided via the config_json field in the Agent node. No external database is mentioned.

What are the runtime requirements for UI-TARS integration?

UI-TARS integration is supported only in Dify Plugin's local debug deployment (not via .difypkg). It requires a Hugging Face Inference Endpoint URL and API key, and runs UI-TARS-1.5-7B model. Known limitations include single-monitor support and macOS Retina display resolution issues.