Agent Upstart

Cool things people are building with agents. Five links a day, one line each.

Skills · last verified 2026-10-23

qdrant/mcp-server-qdrant — install for Claude Desktop

github.com/qdrant/mcp-server-qdrant · MCP server · semantic memory on Qdrant

What it does

Qdrant's official MCP server, acting as a semantic memory layer on top of the Qdrant vector search engine. It exposes two tools: qdrant-store, which stores a piece of information with optional metadata in a collection, and qdrant-find, which retrieves relevant stored information for a query.

Configuration is by environment variables: a Qdrant server URL or a local database path, a collection name, an API key for hosted Qdrant, and the embedding model (the README default is sentence-transformers/all-MiniLM-L6-v2 via fastembed). A read-only mode exists that disables the store tool.

Install

Run with uvx (verified against the repo README's Installation section):

QDRANT_URL="http://localhost:6333" \
COLLECTION_NAME="my-collection" \
EMBEDDING_MODEL="sentence-transformers/all-MiniLM-L6-v2" \
uvx mcp-server-qdrant

Install via Smithery for Claude (same README):

npx @smithery/cli install mcp-server-qdrant --client claude

The README also documents a manual Claude Desktop JSON config using uvx with QDRANT_URL, QDRANT_API_KEY and COLLECTION_NAME. The URL and key values in that example are placeholders, so the full block is not reproduced here.

Verified against github.com/qdrant/mcp-server-qdrant (README, Installation), 2026-10-23.

Trust score: NO_SCORE

No score is given for this page. This is a Python server codebase, not a set of instruction files, and the server source was not fully read in this pass, so a scored review would not be honest. What was checked is listed under Verification below. Note the setup itself: the server connects to a database that stores whatever the agent is told to remember, so the collection and its access belong in the same trust decision as the server.

Verification

Tier: AUTHOR_DOCS_VERIFIED. Verified 2026-10-23 from the author's own repository (fresh clone this run). No install was executed and no account was connected in this pass.

What was checked: Author README (Overview, Components tools, environment variables, Installation, uvx, Docker, Smithery and manual configuration sections) was read from a fresh clone of the repo. The server source code was not fully read.

Limits: Install commands checked against the author README; not executed in this pass, and no Qdrant instance was connected. No full code read was done, so no trust score is given.

Sources: github.com/qdrant/mcp-server-qdrant (README, Installation)

Also in this series

Back to: Skills index