MCP server · Memory + knowledge

OfficialSource ValidatedRead-writeApache-2.0Updated August 2026

Qdrant MCP

Give agents a local or hosted Qdrant memory layer for storing notes and retrieving them by semantic search.

Quick verdict

A two-tool store/find design is easy to reason about, works locally with no cloud credentials, and adds durable semantic recall to any MCP client.

  • MaintainerQdrant (official)
  • Security scopeRead-write
  • Transportsstdio, sse, streamable-http
  • Install channelpypi · mcp-server-qdrant
  • EvidenceSource Validated
  • LicenseApache-2.0 · Open source

Qdrant's official semantic-memory server: store agent notes or code context, then retrieve related items by meaning rather than exact match. Two tools do the work — qdrant-store writes and qdrant-find reads — and it runs against an embedded local database, a self-hosted Qdrant instance or Qdrant Cloud. Setting QDRANT_READ_ONLY removes the store tool entirely.

MCP security best practices

Install Qdrant MCP

uvx mcp-server-qdrant

Config entry (mcpServers)
{
  "mcpServers": {
    "qdrant-mcp": {
      "command": "uvx",
      "args": ["mcp-server-qdrant"],
      "env": {
        "QDRANT_LOCAL_PATH": ":memory:",
        "COLLECTION_NAME": "opensourcesai-memory"
      }
    }
  }
}

Runs with uvx (uv). Use either the embedded QDRANT_LOCAL_PATH mode or a remote QDRANT_URL — not both. The server creates the configured collection automatically if it is missing.

Credentials

NameRequiredWhat it is for
QDRANT_LOCAL_PATHRequiredEmbedded Qdrant location for the published configuration. Not a secret — it is configuration, declared here because the gate requires every environment key to be named. `:memory:` gives an ephemeral store.
COLLECTION_NAMERequiredCollection the store and find tools operate on. Configuration rather than a secret.
QDRANT_API_KEYOptionalAPI key for an authenticated remote Qdrant service or Qdrant Cloud. Unused by the embedded configuration published here.

Verified configurations

Not run by us yet. This profile is Source Validated: the install channel, config entry, credentials and license were checked against the source and upstream metadata, but no configuration has been installed, discovered and executed in a recorded environment. When one is, it appears here with the eight scope dimensions below.

Best for

Agents that need a small, explicit semantic-memory surface rather than access to a whole database API.

Why use it

A two-tool store/find design is easy to reason about, works locally with no cloud credentials, and adds durable semantic recall to any MCP client.

Tradeoffs

The default exposes writes — operators who only need retrieval should set QDRANT_READ_ONLY. Stored content is embedded and persisted into the chosen collection, so sensitive notes need the same protection as the database behind them. The configuration published here starts ephemeral (`:memory:`) on purpose; swapping in a real path makes it durable, and remote deployments add network, credential and service-cost concerns.

Upstream state

pypi package published · repository active. Lifecycle is derived from a committed upstream snapshot at build time, never authored by hand.

Editorial review

Reviewed byOpenSourcesAI EditorialLast updatedAugust 2026SourcesOfficial Qdrant MCP repository — tools, environment variables, transports, read-only mode; Qdrant MCP licence; PyPI: mcp-server-qdrant

Grade B — Source Validated. Re-review is due by 2027-02-15. Config entries, credential names and package status can change upstream; verify against the source repository before granting an agent access.

Sources

Related guides

All MCP servers →