Qdrant MCP Server

Official Qdrant MCP server that turns a vector database into a semantic memory layer for Claude, Cursor, Windsurf, and any MCP client.

Free· Free and open source (Apache-2.0). Qdrant itself can be self-hosted for free or used via Qdrant Cloud (free tier available, paid plans from ~$25/mo for managed clusters).MCP ServersFastEmbed (default: sentence-transformers/all-MiniLM-L6-v2); pairs with any MCP-capable LLM such as Claude 3.5/4, GPT-4o, or local models

Visit website →

Best for

Developers who already use Qdrant (or want to) and need to give Claude, Cursor, or Windsurf a persistent semantic memory or private knowledge-base lookup with minimal glue code.

Skip if

Teams that need a turnkey managed RAG service, hybrid/BM25 search, or tight coupling with a hosted embedding provider like OpenAI without forking the server.

The Qdrant MCP Server is the official Model Context Protocol implementation from the Qdrant team, exposing a Qdrant vector database to any MCP-compatible LLM client such as Claude Desktop, Cursor, Windsurf, VS Code, or a homegrown agent. It gives models two clean tools out of the box: qdrant-store to persist a piece of text (with optional metadata and a target collection) and qdrant-find to retrieve semantically similar records for a natural-language query. That minimalism is the point: the server does not try to be a full RAG framework, it just wires an LLM into Qdrant with sensible defaults so you can bolt on long-term memory or a semantic search skill in minutes.

Under the hood it uses FastEmbed for on-device embeddings (defaulting to sentence-transformers/all-MiniLM-L6-v2) so no separate embedding API key is required, and the tool descriptions themselves are configurable via environment variables, which lets teams re-purpose the same binary as a code-snippet memory, a documentation lookup, or a scoped agent knowledge base without touching Python. Installation is flexible: uvx for one-shot execution, Docker for containerized deploys, Smithery for one-click Claude Desktop wiring, or a manual JSON entry for any MCP client. QDRANT_READ_ONLY lets you ship it as a pure retriever, and QDRANT_LOCAL_PATH supports a fully local, disk-backed vector store for offline dev.

Typical workflows include giving Cursor a persistent "remember this pattern" memory across sessions, letting a Claude agent search internal documentation, or pointing multiple agents at a shared team knowledge collection in Qdrant Cloud. It is Python, Apache-2.0, actively maintained by the Qdrant team, and has become one of the most-starred database-backed MCP servers in the ecosystem.

Editor's take

This is exactly the shape an official MCP server should take: two well-named tools, sane defaults, bundled embeddings, and Docker/uvx installs that just work. If you already run Qdrant, wiring it into Cursor or Claude Desktop takes under five minutes. The lack of an OpenAI embeddings adapter and the single-collection scope are real limits, but for a memory layer it is hard to beat.

— The AI Tool Bible editorial team

Pros

✅ Official, actively maintained by the Qdrant team with 1.4k+ stars and frequent releases
✅ Two-tool surface (store/find) is small enough that models actually use it correctly
✅ Bundled FastEmbed means no separate OpenAI/Cohere embedding key is required to get started
✅ Configurable tool descriptions let you rebrand the same server as memory, code search, or docs lookup
✅ Works across the major MCP clients: Claude Desktop, Cursor, Windsurf, VS Code, and custom agents
✅ Apache-2.0 with local, Docker, and uvx install paths including a fully offline QDRANT_LOCAL_PATH mode
✅ Read-only mode makes it safe to expose a curated knowledge base without letting the model write to it

Cons

⚠️ Only FastEmbed is supported today, so you cannot plug in OpenAI, Cohere, or Voyage embeddings without forking
⚠️ Just two tools: no filtering, hybrid search, payload updates, or collection management surfaced to the model
⚠️ Single active collection per server process; multi-collection agents need multiple server instances or wrapping
⚠️ Assumes you already run and secure a Qdrant instance (self-hosted or Cloud) — not a turnkey managed product
⚠️ Chunking, ingestion pipelines, and re-ranking are entirely your problem; this is a thin bridge, not a RAG framework

Use cases

Persistent memory for Claude Desktop agentsSemantic code snippet search in Cursor and WindsurfPrivate documentation retrieval for internal LLM copilotsTeam knowledge base backed by Qdrant CloudLocal offline vector memory via QDRANT_LOCAL_PATHRead-only knowledge lookup skill for customer-support agentsCross-session context store for autonomous coding agents