Milvus

Milvus is Zilliz's official MCP server that plugs Claude or Cursor into a production-grade vector database. If you're building an agent that needs to remember thousands of documents or handle semantic search at scale, this is the real thing, not a toy. The server exposes collection management, vector insertion, and hybrid search (dense vectors plus metadata filters) through tool calls. I'd reach for this when prototyping a RAG pipeline that might actually ship, because Milvus handles billions of vectors in production environments. The workflow is straightforward: create a collection with your embedding dimension, insert vectors with metadata, then query by similarity or filter by fields. The server supports both Milvus Lite (local, no dependencies) and full Milvus clusters via URI and token. Hybrid search is where it shines: you can filter by date ranges, categories, or custom fields before running the vector search, which beats naive RAG every time. Quirks: you need to know your embedding dimension upfront, and the server assumes you're bringing your own embeddings (it doesn't generate them). The install command uses `uv run` directly on the source file, which is fine for local testing but means you'll want to wrap it in a proper service for production. The repo is sparse on examples, so expect to read Milvus docs to understand index types and search parameters. If you're already running Milvus elsewhere, this is a no-brainer. If you're not, Milvus Lite makes it easy to start local and scale later. Who shouldn't bother: anyone happy with a simpler vector store like Chroma or just using Claude's native context. Milvus is overkill if you're indexing a few dozen documents. But if you're building a multi-tenant agent, need sub-100ms search on millions of vectors, or want to separate your vector store from your LLM provider, this is the correct choice. The official vendor status means it'll track Milvus features as they ship.