Qdrant MCP setup: collection, search and read-only

Choose an existing test collection containing a harmless reference text. Keep its identifier and expected content available outside the search tool. A plausible search result alone does not establish that the intended source was retrieved.

Define the collection and search model

The official Qdrant MCP server provides qdrant-find and qdrant-store. QDRANT_READ_ONLY=true disables the store tool. COLLECTION_NAME selects a default collection. Use an embedding model compatible with that collection; the value below is a documented example, not a universal choice.

VS Code configuration for a remote database

Install uv, adjust host, collection and model, then save the example as .vscode/mcp.json. VS Code prompts for the API key. The database connection uses HTTPS while the MCP process runs locally over stdio.

{
  "servers": {
    "qdrant": {
      "type": "stdio",
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ],
      "env": {
        "QDRANT_URL": "https://YOUR_QDRANT_HOST:6333",
        "QDRANT_API_KEY": "${input:qdrant-key}",
        "COLLECTION_NAME": "YOUR_EXISTING_TEST_COLLECTION",
        "EMBEDDING_MODEL": "sentence-transformers/all-MiniLM-L6-v2",
        "QDRANT_READ_ONLY": "true",
        "QDRANT_SEARCH_LIMIT": "5"
      }
    }
  },
  "inputs": [
    {
      "id": "qdrant-key",
      "type": "promptString",
      "description": "Qdrant API key",
      "password": true
    }
  ]
}

Use a key scoped to the required data and operations. Disabling a tool does not revoke the key’s permissions. Inspect the read-only guide before widening access.

Test a known reference

1. Inspect the offered tools after startup. In read-only mode, confirm that qdrant-store is absent.

2. Use qdrant-find for a distinctive phrase from your reference text. Keep the test limited to the prepared collection.

3. Compare returned text and metadata with your original reference. Record missing or incorrect identifiers separately from ranking differences.

4. Repeat with an unrelated query. Inspect why any result was returned; semantic similarity is not an exact-match guarantee.

For local storage the project documents QDRANT_LOCAL_PATH as an alternative to QDRANT_URL; do not set both. The Qdrant registry search links declared server entries. Use the database planner to choose the first operation.

Evaluate retrieval separately from connectivity

Prepare three small search cases: a known passage, a paraphrase and an unrelated term. For each one, record the reference you expect and the results actually returned. Compare identifiers and content, not just the position of the first hit. A reachable database can still produce unsuitable results if the collection or its preparation does not match the test. Avoid changing the query, model and dataset together. Keep the original test cases and repeat them after a change. This gives you a concrete comparison for a particular retrieval task without treating a connection check as evidence of search quality.

Why is qdrant-store absent?

The example deliberately disables it through QDRANT_READ_ONLY=true.

Does every existing collection work with the example model?

No. Verify compatibility with the model and vector configuration used for your stored data.

Does a matching result prove completeness?

No. Compare against a known reference and record which expected items were retrieved.

Can I set a URL and a local path together?

The project documents them as alternatives. Choose the storage mode you intend to test.

Vendor documentation checked on 2 October 2026. No authenticated account test.

1. Qdrant: MCP server
   Show retrieval command

   curl -s https://github.com/qdrant/mcp-server-qdrant
2. Qdrant: Security
   Show retrieval command

   curl -s https://qdrant.tech/documentation/operations/security/
3. VS Code: MCP configuration
   Show retrieval command

   curl -s https://code.visualstudio.com/docs/agent-customization/mcp-servers

All posts