dify -client

This MCP client acts as an Agent Strategy Plugin that enables a ReAct-style agent to connect to MCP servers and use tools, resources, and prompts exposed by those servers. The client converts the server-provided Tool, Resource, and Prompt definitions into Dify Tools, allowing the selected LLM to inspect tool metadata (name, description, argument type) and invoke tools within the ReAct loop (Reason → Act → Observe). It integrates UI-TARS SDK for GUI automation capabilities, enabling on-demand GUI automation and lifecycle controls for long-running tasks. To use it, install and deploy the plugin, configure the MCP servers you want to connect to, and select the mcpReAct strategy in Dify so the agent node can communicate with your MCP servers and utilize the available tools during task execution.

Prerequisites\n- Docker installed on your machine/server (recommended for the official deployment path) or a Node.js environment if you choose to run locally.\n- Access to the repository where the MCP client lives (https://github.com/3dify-project/dify-mcp-client/).\n- Optional: If using UI-TARS integration, ensure OS-native API access is available in the deployment environment (UI-TARS is supported only in Dify plugin's local debug deployment).\n\nInstallation steps (Docker deployment)\n1) Build the image (if building locally):\nbash\ndocker build -t dify-mcp-client:latest .\n\n2) Run the container in detached mode with your MCP server endpoints configured (example):\nbash\ndocker run -d --name dify-mcp-client \\n -e MCP_SERVER_1_URL=http://host.docker.internal:8080/sse \\n -e MCP_SERVER_2_URL=http://host.docker.internal:8008/mcp \\n dify-mcp-client:latest\n\n3) Alternatively, use the pre-built image in docker-compose or directly pull and run the image as per your environment requirements.\n4) If you prefer running locally with Node.js/TypeScript stdio server (not the recommended path for Docker deployments): follow the repository's development guide to install dependencies and start the server with the appropriate entry point (e.g., node path/to/server.js).\n\nPrerequisites (Node.js approach)\n- Node.js v14+ (or as required by the project)\n- npm or yarn for package management\n\nInstallation steps (Node.js/local development)\n1) Clone the repository and install dependencies:\nbash\ngit clone https://github.com/3dify-project/dify-mcp-client.git\ncd dify-mcp-client\nnpm install\n\n2) Configure the MCP servers in your environment (see mcp_config below) and start the server:\nbash\nnpm start\n\n3) Verify the client connects to your MCP servers and exposes the ReAct-based workflow with UI-TARS integration (if configured).

Tips and caveats:\n- The MCP client uses UI-TARS integration for GUI automation, which requires OS-native API access. In some deployment environments (like Linux-based Docker containers), UI-TARS may require additional configuration or may not be fully functional outside of the intended local debug deployment.\n- If you encounter plugin verification issues during installation, you can disable signature verification in the Dify environment by adjusting FORCE_VERIFYING_SIGNATURE=false in /docker/.env and restarting the services.\n- For stdio-to-Streamable HTTP compatibility, consider converting stdio MCP servers to streamable HTTP via the recommended methods described in the project notes.\n- Ensure each MCP server URL/port is unique when configuring multiple servers.\n- The plugin supports a UI-TARS-powered flow for GUI automation, but this feature is limited to the environments that provide the required native OS capabilities.\n- The repository examples show how to configure mcpServers in the plugin’s config_json; adapt the server names and URLs to match your deployment.