LanceDB

This MCP server lets your LLMs access and reason over your on-disk LanceDB index and document catalog, enabling agentic RAG and hybrid search directly against your data. It keeps data local and reduces token usage by letting the LLM request what it needs when it needs it.

How to use

You use this MCP server by connecting an MCP-compliant client (for example Claude Desktop) to the LanceDB MCP endpoint. From your client, you can ask questions about the entire dataset or drill into specific documents. The server exposes tools to search the catalog and retrieve relevant document chunks, allowing you to explore summaries, metadata, and vectorized content through natural language prompts.

How to install

Prerequisites you need before starting:

• Node.js 18+

• npx

• MCP Client (example: Claude Desktop App)

• Summarization and embedding models installed (customizable in config) — by default Ollama models are used, for example you can pull embeddings with Ollama using commands shown below.

ollama pull snowflake-arctic-embed2

ollama pull llama3.1:8b

Step by step setup

1. Create a local directory to store the index.

2. Add the MCP server configuration to your client. The quick-start example uses an npx-based command.

Quick start configuration (stdio, local server)

{
  "mcpServers": {
    "lancedb": {
      "command": "npx",
      "args": [
        "lance-mcp",
        "PATH_TO_LOCAL_INDEX_DIR"
      ]
    }
  }
}

Local development mode configuration (stdio)

{
  "mcpServers": {
    "lancedb": {
      "command": "node",
      "args": [
        "PATH_TO_LANCE_MCP/dist/index.js",
        "PATH_TO_LOCAL_INDEX_DIR"
      ]
    }
  }
}

Seed data and index setup

The seed script creates two LanceDB tables: one for the catalog of document summaries and metadata, and another for vectorized document chunks.

To seed the index, run the following command, replacing the paths with your local locations.

npm run seed -- --dbpath <PATH_TO_LOCAL_INDEX_DIR> --filesdir <PATH_TO_DOCS>

Catalog and Chunks structure

Catalog stores document summaries and metadata. Chunks store vectorized document chunks and their metadata.

Notes on usage and build

If you need to recreate the index, you can rerun the seed script with the --overwrite option. Build steps for the MCP project include running a build command before using the stdio server.

To inspect the MCP state during development, you can run the inspector tool with the following command.

npx @modelcontextprotocol/inspector dist/index.js PATH_TO_LOCAL_INDEX_DIR

Sample prompts to try

What documents do we have in the catalog?

Why is the US healthcare system so broken?

Available tools

catalog_search

Search the catalog for relevant documents and retrieve summaries and metadata based on a query.

chunks_search

Find relevant vectorized document chunks related to a specific catalog document.

all_chunks_search

Retrieve relevant chunks from all known documents to provide broader context.