MCP Server Qdrant

What qdrant/mcp-server-qdrant MCP server does

The qdrant/mcp-server-qdrant MCP server provides an MCP interface to Qdrant, a vector search engine. Its documented purpose is to maintain and retrieve memories using a semantic memory layer over a Qdrant database. Information is converted into vectors with an embedding model and stored in a named collection, where it can later be searched with a text query.

The server can work with a Qdrant instance reachable through a URL or with a local Qdrant database path. It can also create the configured collection automatically when that collection does not already exist.

How it works

A client sends information to the storage tool, optionally including JSON metadata and a collection name. The server embeds the information and writes it to Qdrant. A separate search tool accepts a query and collection name, then returns matching stored information as separate messages.

The default embedding provider is FastEmbed, and the default model is sentence-transformers/all-MiniLM-L6-v2. FastEmbed is currently the only supported embedding provider. Search returns up to 10 results by default, with the limit controlled through configuration.

The qdrant/mcp-server-qdrant MCP server supports standard input/output transport by default. It also supports Server-Sent Events and streamable HTTP, which can be selected with the --transport option for clients connecting remotely.

Setup and configuration

Run the server with uvx mcp-server-qdrant, or build and run the provided Dockerfile. A Qdrant URL and, when required by that deployment, an API key can be supplied through environment variables. Local mode instead uses QDRANT_LOCAL_PATH; the URL and local path must not be configured together.

COLLECTION_NAME sets the default collection. If no default is configured, storage and search requests must provide a collection name. EMBEDDING_MODEL changes the FastEmbed model, while QDRANT_READ_ONLY=true disables the storage tool.

For SSE or streamable HTTP, use --transport sse or the corresponding transport value and configure the server port with FASTMCP_SERVER_PORT. Container deployments should set FASTMCP_SERVER_HOST=0.0.0.0 so the process listens on the container's network interfaces.

Tools and capabilities

The documented MCP operations cover:

• Storing information in Qdrant with optional JSON metadata.
• Searching Qdrant with a text query.
• Selecting or using a collection name.
• Reporting information and metadata through the exposed MCP interface.
• Limiting the number of returned search results.
• Disabling writes for read-only deployments.

The README names the primary operations qdrant-store and qdrant-find; the listing exposes the related inputs as information, metadata, collection_name, and query.

Limitations and notes

Only FastEmbed is supported as the embedding provider at present. The server requires either a Qdrant URL or a local Qdrant database path, and those two connection modes cannot be used simultaneously. API-key authentication is relevant to Qdrant deployments that require it, but local Qdrant mode can be configured without a Qdrant URL or API key.

The default transport is intended for local MCP clients, while SSE and streamable HTTP are available for remote clients. The documented desktop-client setup specifically covers Claude Desktop.