Which exposes hybrid and full-text search?

Milvus — its official server exposes vector, full-text, and hybrid search as distinct tools plus collection management. Qdrant's official server focuses on semantic store-and-retrieve rather than every query mode.

Are both open-source and self-hostable?

Yes. Both back onto open-source engines you can self-host, with managed options (Zilliz Cloud for Milvus, Qdrant Cloud for Qdrant). Both official servers run locally over stdio.

Milvus vs Qdrant

Milvus MCP and Qdrant MCP both connect an agent to an open-source vector database, but they frame the job differently. Zilliz's official Milvus server runs locally over stdio and exposes the database directly: vector search, full-text search, hybrid search that blends the two, and collection management against a Milvus instance — a full query-and-manage surface for a high-scale vector store. Qdrant's official server, also stdio, is deliberately narrower and opinionated: it presents Qdrant as a semantic memory layer, with tools to store information and retrieve it by meaning, embeddings handled for you. Both are official and both back onto open-source engines you can self-host (or run as managed Zilliz Cloud / Qdrant Cloud), so the decision is less about hosting and more about whether you want a broad database-style interface with hybrid and full-text search, or a focused store-and-recall memory abstraction the agent uses to remember and find context.

How they compare

Dimension	Milvus	Qdrant
Interface philosophy	Database-style: the agent gets vector, full-text, and hybrid search plus collection management over Milvus.	Memory-style: a focused store / find abstraction that treats Qdrant as a semantic memory the agent reads and writes.
Search capabilities	Vector, full-text, and hybrid (dense + sparse) search exposed as distinct tools for fine-grained retrieval.	Semantic retrieval by meaning is the core; the server keeps the surface simple rather than exposing every query mode.
Collection management	Includes managing collections, so the agent can do schema-ish operations against the Milvus instance.	Geared toward storing and recalling entries; collection plumbing is largely handled for you, not surfaced as a primary task.
Embeddings handling	Works against vectors you manage in Milvus, fitting an existing high-scale vector pipeline.	Designed so the agent stores and retrieves information without you wiring embeddings explicitly into each call.
Best-fit task	Driving a full-featured vector database — hybrid and full-text search, collection management — at scale from an agent.	Giving an agent a drop-in semantic memory: remember this, recall what is relevant, with minimal surface.

Verdict

Pick by whether you want a vector database or a vector memory. Reach for Milvus MCP when you run Milvus (or Zilliz Cloud) and want the agent to use a broad, database-style surface — vector, full-text, and hybrid search plus collection management — over a high-scale store. Reach for Qdrant MCP when you want a focused semantic memory layer the agent uses to store and recall information by meaning, with a deliberately minimal surface and embeddings handled for you. In short: Milvus for full database control and hybrid retrieval; Qdrant for a clean store-and-recall memory abstraction on top of an open-source engine.

FAQ

Which exposes hybrid and full-text search?: Milvus — its official server exposes vector, full-text, and hybrid search as distinct tools plus collection management. Qdrant's official server focuses on semantic store-and-retrieve rather than every query mode.
Are both open-source and self-hostable?: Yes. Both back onto open-source engines you can self-host, with managed options (Zilliz Cloud for Milvus, Qdrant Cloud for Qdrant). Both official servers run locally over stdio.