Qdrant MCP — Anwendungsfälle, Installieren & Live-Demo

Name: Qdrant MCP Server
Author: qdrant

Warum nutzen

Hauptfunktionen

Two-tool minimal surface — store text, find text; no schema gymnastics
Embedding model runs locally via fastembed by default (no extra API key)
Works with Qdrant Cloud or any self-hosted instance via QDRANT_URL
Metadata stored alongside each entry for filterable recall
Ideal building block for agent memory and lightweight RAG

Live-Demo

In der Praxis

qdrant.replay ▶ bereit

0/0

Installieren

Wählen Sie Ihren Client

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "qdrant": {
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ]
    }
  }
}

Öffne Claude Desktop → Settings → Developer → Edit Config. Nach dem Speichern neu starten.

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "qdrant": {
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ]
    }
  }
}

Cursor nutzt das gleiche mcpServers-Schema wie Claude Desktop. Projektkonfiguration schlägt die globale.

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "qdrant": {
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ]
    }
  }
}

Klicken Sie auf das MCP-Servers-Symbol in der Cline-Seitenleiste, dann "Edit Configuration".

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "qdrant": {
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ]
    }
  }
}

Gleiche Struktur wie Claude Desktop. Windsurf neu starten zum Übernehmen.

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "qdrant",
      "command": "uvx",
      "args": [
        "mcp-server-qdrant"
      ]
    }
  ]
}

Continue nutzt ein Array von Serverobjekten statt einer Map.

~/.config/zed/settings.json

{
  "context_servers": {
    "qdrant": {
      "command": {
        "path": "uvx",
        "args": [
          "mcp-server-qdrant"
        ]
      }
    }
  }
}

In context_servers hinzufügen. Zed lädt beim Speichern neu.

claude mcp add qdrant -- uvx mcp-server-qdrant

Einzeiler. Prüfen mit claude mcp list. Entfernen mit claude mcp remove.

Anwendungsfälle

Praxisnahe Nutzung: Qdrant

Give a Claude agent persistent memory across sessions

👤 Builders making personal assistants or internal copilots ⏱ ~15 min beginner

Wann einsetzen: You want Claude to remember user preferences, past decisions, or ongoing projects even after the chat ends.

Voraussetzungen

Running Qdrant (local Docker or cloud) — docker run -p 6333:6333 qdrant/qdrant OR a Qdrant Cloud cluster URL + API key
COLLECTION_NAME env var set — Any string, e.g. claude_memory

Ablauf

Teach it to store important facts

Whenever I tell you something important about a project (deadlines, stakeholders, decisions), store it with qdrant-store, metadata {project, category}.✓ Kopiert

→ Claude starts echoing 'stored' for durable facts
Verify recall works

What do you remember about project 'atlas'? Use qdrant-find with a query like 'project atlas decisions'.✓ Kopiert

→ Relevant prior messages returned with scores
Curate and forget

Search for anything about project 'atlas' that's more than 90 days old or marked obsolete, and delete those entries.✓ Kopiert

→ List of pruned items with confirmation

Ergebnis: An assistant that actually remembers what you told it last week — scoped per-project, prunable.

Fallstricke

Storing every message bloats the collection and degrades recall quality — Only store explicit facts/decisions, not chit-chat. Make the 'store or not' decision part of the system prompt.
Collection created with wrong vector size after switching embedding models — Qdrant rejects mismatched vectors — drop and recreate the collection when you change EMBEDDING_MODEL

Kombinieren mit: filesystem · notion

Build a lightweight RAG over a docs folder

👤 Devs who want RAG without a framework ⏱ ~30 min intermediate

Wann einsetzen: You have 50–5000 Markdown files and want Claude to answer questions against them, with citations.

Voraussetzungen

Docs on disk as Markdown — Any folder of .md files

Ablauf

Chunk and store the docs

Read every .md under /docs. Split into ~500-token chunks on heading boundaries. For each chunk, call qdrant-store with the text and metadata {source_path, heading}.✓ Kopiert

→ N chunks stored, one per section
Query with a user question

User asks: 'How do I rotate API keys?' Use qdrant-find to pull the top 5 most relevant chunks. Cite source_path in your answer.✓ Kopiert

→ Answer with inline [source_path] citations
Measure retrieval quality

For these 10 eval questions [list], which of the expected source paths appear in the top-5 retrieval? Report recall@5.✓ Kopiert

→ A retrieval-quality score you can improve iteratively

Ergebnis: A working RAG loop you can tune chunk size and k until quality is acceptable.

Fallstricke

Too-large chunks dilute the embedding and kill recall — Keep chunks under ~1000 tokens; split by headings first, then by token count as a fallback
Updating a doc doesn't remove old chunks — answers become stale — Use a deterministic point id (hash of source_path+heading) so upserts replace rather than duplicate

Kombinieren mit: filesystem · firecrawl

Deduplicate a messy list of tickets, leads, or FAQs semantically

👤 Ops teams with a CSV of near-duplicates ⏱ ~25 min intermediate

Wann einsetzen: Exact-match dedup misses things like 'reset password' vs 'how do I change my password' — you need semantic similarity.

Ablauf

Store each item with its row id as metadata

Read rows.csv. For each row, qdrant-store with information=<text> and metadata={row_id: <id>}.✓ Kopiert

→ N points stored
Cluster by similarity

For each row, query qdrant-find for its top-5 neighbors with score > 0.85. Output groups of row_ids that are mutually near.✓ Kopiert

→ Duplicate groups printed
Pick canonical + mark rest as duplicates

For each group, pick the longest/most-informative row as canonical. Output a CSV {row_id, canonical_id}.✓ Kopiert

→ Dedup map ready for the source system

Ergebnis: A dedup mapping CSV with confidence scores, reviewable by a human before applying.

Fallstricke

Similarity threshold is domain-specific — 0.85 may be too lenient or too strict — Hand-label 20 pairs first, then pick the threshold that best separates dup from non-dup

Kombinieren mit: postgres · filesystem

Searchable meeting-notes memory

👤 Managers / ICs drowning in Notion/Obsidian notes ⏱ ~20 min beginner

Wann einsetzen: You take weekly notes but can never find the one where a specific decision was made.

Voraussetzungen

Folder of meeting notes — Any text or markdown files

Ablauf

Index existing notes

Walk /meetings/**/*.md. For each, qdrant-store the body with metadata {date, attendees, project}.✓ Kopiert

→ All notes indexed with dates
Recall decisions

Find every note where we discussed 'pricing for enterprise tier'. Show me the date and a 2-line summary of each.✓ Kopiert

→ Ranked list of matching meetings
Keep it fresh

Add today's note <paste>, then tell me which past notes most likely contradict or update decisions in today's.✓ Kopiert

→ Contradiction check via semantic neighbors

Ergebnis: A semantic index over your notes you can keep updating weekly.

Fallstricke

Mixing personal + work notes in one collection leaks scope — Use separate collections or enforce a scope metadata filter on every find

Kombinieren mit: filesystem · notion

Kombinationen

Mit anderen MCPs für 10-fache Wirkung

qdrant + filesystem

Index a local docs folder then answer questions with citations

Index every .md under /docs into Qdrant, then answer: 'how does our auth flow work?' with citations to the original file paths.✓ Kopiert

qdrant + firecrawl

Crawl a site and build a searchable knowledge base

Crawl docs.mycompany.com with Firecrawl, store each page in Qdrant collection company_docs.✓ Kopiert

qdrant + postgres

Semantic search over unstructured columns in a relational DB

SELECT id, body FROM support_tickets created in the last 30 days, embed each body into Qdrant with metadata {ticket_id}, then let me search them by meaning.✓ Kopiert

Werkzeuge

Was dieses MCP bereitstellt

Werkzeug	Eingaben	Wann aufrufen	Kosten
qdrant-store	information: str, metadata?: object	Persist a fact, chunk, or note for later semantic recall	free (local embedding)
qdrant-find	query: str, limit?: int	Retrieve semantically similar entries to answer a question or deduplicate	free

Kosten & Limits

Was der Betrieb kostet

API-Kontingent: Self-hosted: unlimited. Qdrant Cloud: depends on cluster size.
Tokens pro Aufruf: Store: ~100 tokens overhead per call. Find: ~200 tokens + result payload.
Kosten in €: Free if self-hosted. Qdrant Cloud free tier: 1GB cluster. Paid from ~$25/mo.
Tipp: Start with local Docker for dev; upgrade to Cloud only when you need persistence and multi-device access.

Sicherheit

Rechte, Secrets, Reichweite

Credential-Speicherung: QDRANT_URL and optional QDRANT_API_KEY in env vars

Datenabfluss: If self-hosted: none. If Qdrant Cloud: all vectors and metadata sent to your cluster region.

Embeddings can be reversed enough to recover approximate text — treat the collection as sensitive as the source data.
Set COLLECTION_NAME per tenant if hosting a multi-user agent.

Fehlerbehebung

Häufige Fehler und Lösungen

Collection does not exist / Not found

The server creates the collection on first store only if COLLECTION_NAME is set. Verify the env var and restart the MCP.

Prüfen: curl $QDRANT_URL/collections

Vector dimension mismatch

You changed EMBEDDING_MODEL without dropping the old collection. Drop it and start fresh (or use a new COLLECTION_NAME).

Prüfen: curl $QDRANT_URL/collections/<name>

Connection refused on localhost:6333

Qdrant container isn't running. docker run -p 6333:6333 qdrant/qdrant and retry.

Prüfen: curl localhost:6333/healthz

Searches return irrelevant results

Chunks may be too big or the embedding model too weak. Try FastEmbed's bge-small-en-v1.5 and chunks ≤500 tokens.

Alternativen

Qdrant vs. andere

Alternative	Wann stattdessen	Kompromiss
Chroma MCP	You prefer an embedded vector DB with zero infra	Less production-grade than Qdrant for heavy loads
Pinecone MCP	You're already on Pinecone and want hosted-only	Paid from day one; more opinionated
Memory MCP	You want ultra-simple key-value memory, not semantic	No embeddings — exact recall only

Mehr

Ressourcen

📖 Offizielle README auf GitHub lesen

🐙 Offene Issues ansehen

🔍 Alle 400+ MCP-Server und Skills durchsuchen