claude-qdrant

This MCP server provides a multi-tenant RAG system built on top of Qdrant with LM Studio integration. It isolates data per client by maintaining separate collections for each client (e.g., {client}_catalog and {client}_chunks), enabling privacy-focused multi-client workflows. The server exposes MCP tools for catalog and document search, including collection-wide info, and supports semantic search using 1024-dim embeddings via the LM Studio BGE-M3 model, with document summaries generated by the Qwen3-8B model. You can seed documents, perform various search operations, and seed or seed-related metadata in an isolated, client-specific manner. The available MCP tools are: collection_info to get status across collections and clients; catalog_search to search document summaries for a specific client; chunks_search to search within document chunks with optional source filtering; and all_chunks_search to search across all clients and collections. To integrate with Claude Desktop, you can register the MCP server in claude_desktop_config.json and provide the required environment variables. The server can be started with npm start, or in development mode via npm run watch, depending on your setup. The configuration relies on local LM Studio instances for embeddings and summaries, and a local Qdrant instance for vector storage and retrieval.

Prerequisites:

• Node.js 18+ installed on your machine
• NPM (comes with Node.js) or PNPM/Yarn as you prefer
• A running Qdrant instance (local Docker or managed) accessible at QDRANT_URL
• LM Studio running locally with BGE-M3 embeddings and Qwen3-8B model

step-by-step:

1. Clone the MCP repository or your project: git clone https://github.com/marlian/claude-qdrant-mcp.git cd claude-qdrant-mcp

2. Install dependencies: npm install

3. Build the project (if you need a build step): npm run build

4. Prepare environment variables:

   • Copy example env if provided and edit accordingly
   • Ensure QDRANT_URL is set to your Qdrant instance URL
   • Ensure LM Studio is reachable (BGE-M3 embeddings and Qwen3-8B)
   • Define CLIENT_COLLECTIONS as a comma-separated list of client names

5. Start the MCP server: npm start

6. Optional: run the interactive setup and tests if available: npm run setup npm run test-connection

Notes:

• If you run in development, you may use npm run watch to enable hot-rebuilds.
• If you deploy, ensure the dist/index.js path in mcp_config matches your build output.

Tips and considerations:

• Environment variables: QDRANT_URL, QDRANT_API_KEY, and CLIENT_COLLECTIONS are commonly used. Ensure LM Studio is configured and accessible at the specified URL for embeddings and summaries.
• For multi-client isolation, carefully manage CLIENT_COLLECTIONS and naming conventions like {client}_catalog and {client}_chunks.
• When integrating with Claude Desktop, register the MCP server with the node command and provide absolute path to dist/index.js along with env vars in claude_desktop_config.json.
• If you encounter connectivity or model loading issues, verify that LM Studio and Qdrant are running, ports are open, and no firewall blocks the traffic.
• Cryptographic keys or API keys should be stored securely; avoid embedding sensitive keys directly in code or config files in public repos.
• If you need to adjust performance, tuning parameters like concurrency, chunk size, and batch size can be configured via environment variables in the .env file (examples include CONCURRENCY, CHUNK_SIZE, BATCH_SIZE, CHUNK_OVERLAP).