Is Qdrant MCP Server worth it?

A tidy, local-first memory for agents that already sit next to Qdrant, deliberately too small to be a general interface to the database. We rate it 3.7 out of 5.

How much does Qdrant MCP Server cost?

Free and open source (Apache 2.0). Qdrant is free to self-host; Qdrant Cloud has a free single-node tier (0.5 vCPU, 1 GB RAM, 4 GB disk), usage-based Standard and a Premium tier with a minimum spend.

What are the downsides of Qdrant MCP Server?

Only store and find; no update, delete or collection management; fastEmbed models only; pins fastmcp 2.7.0 with known advisories; no release since December 2025; no authentication on the HTTP endpoint itself.

How much does Qdrant MCP Server cost for a team or in production?

The server is free. Self-hosted Qdrant costs only the machine; Qdrant Cloud's free tier is one 0.5 vCPU, 1 GB RAM node, Standard is usage-based and Premium, with SSO and private VPC links, has a minimum spend.

Is Qdrant MCP Server safe for company data?

Reads and writes the configured collection with the Qdrant API key you provide; embeddings are computed locally. Stored text returns verbatim to the model, so memories saved from untrusted sources can carry injected instructions. The HTTP endpoint has no auth of its own.

What are the best Qdrant MCP Server alternatives?

The closest alternatives we have reviewed are Memory (reference server), Redis MCP Server and MongoDB MCP Server.

Qdrant MCP Server MCP server review (2026)

What it is

mcp-server-qdrant is the official MCP server from Qdrant, the open-source vector search engine. The README describes it as a semantic memory layer: the model saves text, and later finds it again by meaning rather than exact words. The repository has about 1,550 GitHub stars and an Apache 2.0 licence.

It has two tools. qdrant-store saves a piece of information with optional JSON metadata, and qdrant-find runs a semantic search returning up to 10 results by default. Text is embedded locally with FastEmbed, by default sentence-transformers/all-MiniLM-L6-v2, and FastEmbed is the only provider supported. It points at a Qdrant server or cloud cluster (QDRANT_URL plus an API key) or a local on-disk database, and creates the collection if it does not exist. QDRANT_READ_ONLY=true removes the store tool, and since 0.8.0 the find tool can expose payload filters configured in code.

Run it with uvx mcp-server-qdrant; stdio is the default, with SSE and Streamable HTTP bound to 127.0.0.1:8000 unless changed. The latest PyPI release is 0.8.1 from December 2025; the last commit, adding Kiro support, was in August 2026.

Install

uvx mcp-server-qdrant

Our review

The design is narrow on purpose. Two tools, one default collection and local embeddings make it easy to reason about, and rewriting the tool descriptions through environment variables turns it into a code-snippet store for Cursor, Windsurf or Claude Code. Nothing goes to an embedding API: FastEmbed runs MiniLM on your machine, and the only network traffic is to the Qdrant instance you chose.

The limits show quickly. It cannot list, update or delete points or manage collections, and only FastEmbed models are supported, so it will not search a collection built by another pipeline with OpenAI or Cohere vectors. Payload filtering does exist since 0.8.0, but it is configured by subclassing the server in Python rather than through the documented environment variables, and an open issue reports the 'except' condition never works.

On security, the HTTP transports bind to 127.0.0.1 by default, but the README's Docker example sets 0.0.0.0 and the MCP endpoint has no authentication of its own, so anything that can reach the port can read and write memories. Stored memories are also a prompt-injection channel: qdrant-find returns saved text verbatim, and an open issue notes it does not escape structural tags, so anything an agent once stored from an untrusted page comes back as context. Read-only mode is the right default for retrieval-only agents.

Maintenance is slow. The pyproject pins fastmcp to exactly 2.7.0, and an August 2026 issue lists 14 OSV findings against that version, including a critical SSRF in FastMCP's OpenAPI provider, with no fix released. Compared with the reference memory server's local knowledge-graph file, this gives genuine semantic search; compared with hosted memory services, control at the cost of features and patch cadence.

Where it earns its place

A developer using Claude Code across several repositories
Keep a searchable store of reusable code snippets and decisions
Run with COLLECTION_NAME=code-snippets and custom tool descriptions; the agent stores snippets with metadata and finds them by meaning later. Free with a local on-disk database; MiniLM vectors are 384 dimensions, about 1.5 KB each.
A platform team already running Qdrant for RAG
Give internal agents read-only access to a curated knowledge collection
Point the server at the cluster with QDRANT_READ_ONLY=true so only qdrant-find is exposed. Works only if the collection was embedded with the same FastEmbed model; otherwise it needs re-indexing.
A small AI start-up prototyping agent memory
Test long-term memory before paying for a managed service
Use Qdrant Cloud's free 1 GB node; at roughly 1.5 KB per MiniLM vector, 100,000 memories is about 150 MB of vectors before index overhead. Moving to Standard is usage-based when the prototype grows.

Before you commit

Setup: uvx mcp-server-qdrant with QDRANT_URL or QDRANT_LOCAL_PATH and COLLECTION_NAME set; the README documents Claude Desktop, Cursor, Windsurf, VS Code and Claude Code. The first run downloads the FastEmbed model, then a first call takes a minute or two.
Integrations: Two tools: qdrant-store (write) and qdrant-find (read, up to QDRANT_SEARCH_LIMIT results, default 10). Read-only mode drops the store tool; filters on payload fields need code-level configuration.
Data and privacy: Reads and writes the configured collection with the Qdrant API key you provide; embeddings are computed locally. Stored text returns verbatim to the model, so memories saved from untrusted sources can carry injected instructions. The HTTP endpoint has no auth of its own.
Cost at scale: The server is free. Self-hosted Qdrant costs only the machine; Qdrant Cloud's free tier is one 0.5 vCPU, 1 GB RAM node, Standard is usage-based and Premium, with SSO and private VPC links, has a minimum spend.
Limits: No update, delete or collection tools, FastEmbed only, fastmcp pinned at 2.7.0 with open advisories, and no score threshold or deduplication on results according to an open issue.

What's good

Official server from Qdrant
Local embeddings, no external API needed
Read-only mode and customisable tool descriptions
Works with a local file, self-hosted server or Qdrant Cloud

What's not

Only store and find; no update, delete or collection management
FastEmbed models only
Pins fastmcp 2.7.0 with known advisories; no release since December 2025
No authentication on the HTTP endpoint itself

Best for

Teams already running Qdrant who want a simple, self-hosted semantic memory or snippet store for their agents.

Alternatives

Memory (reference server)

Model Context ProtocolDeveloper

3.5/5

Official MCP reference server that gives a model persistent memory as a local knowledge graph of entities and relations.

Redis MCP Server

RedisData & databases

3.8/5

Redis's official MCP server: 53 tools for strings, hashes, lists, streams, JSON and vector search on any Redis database.

MongoDB MCP Server

MongoDBData & databases

4.4/5

MongoDB’s official server: 53 tools for queries, schema, indexes and Atlas admin, with read-only mode in every example.

Qdrant MCP Server: common questions

Is Qdrant MCP Server worth it?: A tidy, local-first memory for agents that already sit next to Qdrant, deliberately too small to be a general interface to the database. We rate it 3.7 out of 5.
How much does Qdrant MCP Server cost?: Free and open source (Apache 2.0). Qdrant is free to self-host; Qdrant Cloud has a free single-node tier (0.5 vCPU, 1 GB RAM, 4 GB disk), usage-based Standard and a Premium tier with a minimum spend.
What are the downsides of Qdrant MCP Server?: Only store and find; no update, delete or collection management; fastEmbed models only; pins fastmcp 2.7.0 with known advisories; no release since December 2025; no authentication on the HTTP endpoint itself.
Who is Qdrant MCP Server best for?: Teams already running Qdrant who want a simple, self-hosted semantic memory or snippet store for their agents.
How much does Qdrant MCP Server cost for a team or in production?: The server is free. Self-hosted Qdrant costs only the machine; Qdrant Cloud's free tier is one 0.5 vCPU, 1 GB RAM node, Standard is usage-based and Premium, with SSO and private VPC links, has a minimum spend.
Is Qdrant MCP Server safe for company data?: Reads and writes the configured collection with the Qdrant API key you provide; embeddings are computed locally. Stored text returns verbatim to the model, so memories saved from untrusted sources can carry injected instructions. The HTTP endpoint has no auth of its own.
What are the best Qdrant MCP Server alternatives?: The closest alternatives we have reviewed are Memory (reference server), Redis MCP Server and MongoDB MCP Server.

Answers checked Oct 23, 2026.