Qdrant’s official MCP server takes the opposite approach to most database servers: instead of exposing the API surface, it exposes two tools. Store something with its metadata, and find things semantically similar to a query. Embedding happens inside the server, so the agent never handles a vector.
Bring your own
Claude Sonnet 4.5
Chat with 60+ AI models on the same workflow — switch to a different model mid-conversation and re-run the same prompt, or use Compare mode to put several side-by-side and balance quality vs. cost.
A Qdrant URL and API key for Cloud or a remote instance. Neither is needed if you run against a local on-disk path.
How models use it and what it is built for.
The server is built as a semantic memory layer rather than a database client. qdrant-store takes a piece of text plus optional structured metadata, embeds it with the configured model and writes it to a collection. qdrant-find takes a natural-language query, embeds it the same way and returns the closest matches with their metadata attached. That is the whole surface, and the narrowness is the point — a model picks correctly between two well-described tools far more reliably than between twenty. The detail that makes it genuinely flexible is that both tool descriptions are configurable by environment variable. Rewriting them turns the same server into a code-snippet memory, a decisions log or a support-ticket search, because the description is what the model reads when deciding whether to call it. It runs over stdio via uvx, or in SSE mode when you want to host it centrally, and can point at Qdrant Cloud, a self-hosted instance or a local on-disk path with no server at all.
Typical tools an AI model can call. Exact names vary by version.
Taken from the official Qdrant documentation — see qdrant/mcp-server-qdrant — official repository for the full reference.
Environment variables
QDRANT_URLURL of the Qdrant instance, e.g. https://xyz.cloud.qdrant.io:6333. Omit if using QDRANT_LOCAL_PATH.
QDRANT_API_KEYAPI key for Qdrant Cloud or a secured self-hosted instance.
COLLECTION_NAMErequiredCollection to read and write. Created automatically if it does not exist.
EMBEDDING_MODELFastEmbed model name. Defaults to sentence-transformers/all-MiniLM-L6-v2.
QDRANT_LOCAL_PATHLocal directory for an embedded instance. Mutually exclusive with QDRANT_URL.
TOOL_FIND_DESCRIPTIONThe description the model reads for the find tool. Rewriting this is how you specialise the server.
Client configuration
uvx against Qdrant Cloud
The collection is created on first write if it does not already exist.
{
"mcpServers": {
"qdrant": {
"command": "uvx",
"args": ["mcp-server-qdrant"],
"env": {
"QDRANT_URL": "https://xyz.cloud.qdrant.io:6333",
"QDRANT_API_KEY": "YOUR_API_KEY",
"COLLECTION_NAME": "agent-memory"
}
}
}
}Local on-disk, no server
Good for a single-machine setup or trying the server out. Not suitable for concurrent clients.
{
"mcpServers": {
"qdrant": {
"command": "uvx",
"args": ["mcp-server-qdrant"],
"env": {
"QDRANT_LOCAL_PATH": "/path/to/qdrant-data",
"COLLECTION_NAME": "agent-memory"
}
}
}
}Copy any of these into MCP Agent Studio after connecting.
Remember this debugging approach and why it worked, tagged with the service name.
Have we solved something like this connection-pool error before?
Store this architecture decision with its date and the alternatives we rejected.
Find the three most relevant past incidents to what is happening now.
This is not a single-model product: you get the same MCP connection with 60+ models (Claude, GPT, Gemini, DeepSeek, open-weight, and more), you can switch mid-conversation, and you can open Compare mode to run the same prompt against multiple models at once. The card above is a suggested starting point for this server — not the only choice.
Default pick for Qdrant
Claude Sonnet 4.5
Retrieval quality depends on the model writing a good query and judging which results are actually relevant. Sonnet 4.5 discards weak matches instead of citing them.
Listing tools proves the server is reachable, not that a model can work with it. Evals go further: they read every tool on the server, write a test suite from its real schemas, and run it — code decides pass/fail on the responses (schema conformance, error codes, pagination, result caps) while a scoring model grades plain-English tasks driven through the tools.
Get a pass/fail report per tool with the evidence behind each verdict — and replay the same suite after every schema change. Destructive tools are excluded from the run.
Run evalsOpen MCP Agent Studio with the connection pre-filled. Add your token, pick any of 60+ models, and start chatting — no install required.
Open Agent StudioCommon questions about connecting, scoping and using it safely.
It is Qdrant’s official MCP server, a semantic memory layer over a Qdrant collection. It exposes two tools — qdrant-store and qdrant-find — and handles embedding internally, so the agent works in plain text.
No. The server embeds on both write and query using FastEmbed with the model named in EMBEDDING_MODEL, defaulting to sentence-transformers/all-MiniLM-L6-v2. Changing the model after you have written data means re-indexing, since old and new vectors are not comparable.
Because tool selection degrades as the tool list grows. Two clearly described tools are chosen correctly far more often than a full database API. If you need richer querying, the tool descriptions are configurable and the collection is still a normal Qdrant collection you can hit directly.
The memory server stores explicit entities and relations in a readable graph — precise, auditable, and only recalls what you can name. Qdrant recalls by meaning across unstructured text, which scales better and is fuzzier. Teams often run both.
Yes. Set QDRANT_LOCAL_PATH and it uses an embedded on-disk instance with no service to run. That is fine for one client on one machine; use a real instance once more than one process needs the same collection.
Pinecone
Build and query Pinecone indexes without leaving the conversation.
Memory
A local knowledge graph that gives an agent memory across sessions.
Elasticsearch
Search indices, read mappings and run query DSL with an AI agent.
MongoDB
Query collections, run aggregations and inspect schemas with natural language.