Feat/mcp server

[quannadev]

This PR introduces a standalone MCP (Model Context Protocol) Server for NexusRAG, allowing AI coding assistants like Cursor or Claude Desktop to seamlessly integrate with and query the NexusRAG backend.

Key Features & Changes

• 🚀 New mcp-server Project: Initialized a dedicated Node.js/TypeScript project inside the mcp-server/ directory using the official @modelcontextprotocol/sdk.
• 🛠️ Integrated Tools: Added tailored MCP tools with highly descriptive instructions to help AI agents use them effectively:
  • get_workspace_list: Lists all available workspaces/knowledge bases.
  • get_document_by_id: Fetches a document's metadata and parsing state by ID.
  • query: Performs semantic/hybrid search against the vector database.
  • get_chunks: Retrieves raw text chunks extracted from a specific document.
• 🌐 Transport Protocol: Implemented the modern StreamableHTTPServerTransport standard (running on Express over port 8000 at the /mcp endpoint) replacing the deprecated SSEServerTransport/stdio.
• 🐳 Dockerization:
  • Created a multi-stage Dockerfile for optimal image size.
  • Integrated mcp-server into the root docker-compose.yml to automatically run alongside backend services.
  • Added a standalone docker-compose.yml inside mcp-server for flexible, isolated deployment.
• 📝 Documentation Update: Added a new section to README.md explaining how users can connect Cursor/Claude Desktop to the MCP Server.

How to Test

1. Run docker compose up -d --build from the root directory.
2. Ensure you see NexusRAG MCP server running on Streamable HTTP at http://localhost:8000/mcp in the logs.
3. Open Cursor or Claude Desktop settings.
4. Add a new MCP connection of type SSE (or HTTP) pointing to http://localhost:8000/mcp.
5. Verify that the agent can execute tools like get_workspace_list successfully.