LanceDB

You deploy this MCP server to store and retrieve memories in LanceDB as a semantic memory layer. It lets you add memories with text content and retrieve semantically similar memories for context when you chat with an LLM or build AI-powered workflows, enabling richer interactions with stored data.

How to use

You run the MCP server locally and connect your MCP client to perform two main actions: add-memory to store new text memories with embeddings, and search-memories to fetch memories similar to a given query. Start the server through your MCP client’s configuration, then use the client’s memory endpoints to add content or retrieve relevant memories. Ensure your client is configured to use the provided stdio entry point, so requests flow through the same process that handles embeddings and similarity scoring.

How to install

Prerequisites you need before starting the server
- Python 3.x or a compatible runtime environment for your MCP tooling
- The MCP runtime tooling that executes stdio-based servers (here, the UVX runtime is used in examples)
- A LanceDB-compatible environment or directory layout for the vector database

Step-by-step setup
1) Ensure your environment has the required runtime installed. If you use UVX tooling, install it per your environment’s guidelines.
2) Prepare your LanceDB data directory. The server uses a path like a local LanceDB database directory to store vectors and memories.
3) Configure the MCP client to launch the server via the provided stdio entry point. You will typically point the client to invoke the runtime command with the server’s module name.

Example local startup configuration for the MCP client (stdio) follows this pattern. You will run a runtime command that starts the mcp-lance-db server, passing the server name as an argument.

{
  "lancedb": {
    "command": "uvx",
    "args": [
      "mcp-lance-db"
    ]
  }
}

Additional sections

Configuration details for the server include the database path and collection name used by LanceDB, the embedding provider, and the model for generating embeddings. These values determine how memories are stored and retrieved. The system uses a similarity threshold to bound retrieval quality and performance.

Security and access considerations: since this server is designed to operate locally with a LanceDB instance, ensure that access to the LanceDB data directory is restricted to trusted processes. If you expose endpoints or run in shared environments, apply appropriate access controls and authentication for clients.

Troubleshooting tips: start by confirming the stdio command is reachable and that the LanceDB directory exists. If the MCP client reports resource changes or memory indexing updates, verify that the embeddings are being computed and stored correctly. For debugging, use the MCP Inspector to observe messages between the client and server.

Available tools

add-memory

Adds a new memory to LanceDB by storing the provided content along with vector embeddings for retrieval.

search-memories

Retrieves memories that are semantically similar to a given query, with an optional limit for the number of results.