Qdrant MCP Server

Overview

Qdrant MCP Server turns Qdrant into a semantic memory layer for an agent: two tools, one to store a piece of information and one to retrieve it later by meaning, backed by real vector search instead of exact string matching.

How it works

1. qdrant-store saves a piece of text plus optional metadata into a named collection, creating the collection automatically if it doesn't exist yet.
2. qdrant-find searches that collection with a natural-language query and returns the most semantically relevant stored entries, not just keyword matches.
3. By default it embeds text locally with FastEmbed (sentence-transformers/all-MiniLM-L6-v2), so it works against a local or self-hosted Qdrant instance without calling an external embedding API.
4. The tool descriptions for store and find are fully customizable via environment variables, which is how the same server gets repurposed as a semantic code-search tool for Cursor or Windsurf instead of a generic memory store.
5. QDRANT_READ_ONLY=true disables the store tool entirely, useful when you want an agent to search an existing knowledge base without being able to add to it.

Examples

"Store this design decision in Qdrant so I can find it again later: we
chose event sourcing for the orders service because of the audit
requirement."

Installation

You need uv (recommended) or Python with pip, and a running Qdrant instance (local via Docker, or Qdrant Cloud).

1. Make sure you have a Qdrant instance reachable, for example:

   docker run -p 6333:6333 qdrant/qdrant
   

2. Add the server to your MCP client's config, setting QDRANT_URL and COLLECTION_NAME as shown in the sidebar.
3. Restart your client.
4. Ask the agent to store a note, then ask it to find that note again by meaning, to confirm search works.