Milvus MCP server

OfficialZillizConfig last verified Jun 1, 2026

Zilliz's official Milvus MCP server: vector, full-text, and hybrid search plus collection management over a Milvus vector database.

The Milvus MCP server is the official integration from Zilliz, the team behind Milvus, that gives an agent direct access to a Milvus vector database. It covers the full retrieval surface: dense vector similarity search with configurable distance metrics and range filtering, BM25 full-text search, text-similarity search over embedded fields, hybrid search that fuses text and vector results with Reciprocal Rank Fusion, and filter-expression queries for precise lookups. Beyond search it can manage the database itself, listing and switching databases, creating collections with custom schemas and indexes, loading and releasing collections from memory, inspecting collection metadata, and inserting or deleting entities.

The recommended way to run it is to clone the repository and launch the server module with uv, pointing it at a Milvus instance via --milvus-uri (or the MILVUS_URI environment variable) with an optional MILVUS_TOKEN for authenticated clusters and MILVUS_DB to choose a database. It runs over stdio by default and also supports SSE and streamable HTTP for networked deployments. Because it speaks to any Milvus or Zilliz Cloud instance, it works equally well as a long-term semantic memory backend or as a search layer over an existing embedded corpus.

Quick install

Copy-paste configs are provided for all 8 supported clients. Pick your client below.

Add to ~/.claude.json

~/.claude.json
json
{
  "mcpServers": {
    "milvus": {
      "command": "python",
      "args": [
        "/path/to/mcp-server-milvus/src/mcp_server_milvus/server.py",
        "--milvus-uri",
        "http://localhost:19530"
      ],
      "env": {
        "MILVUS_URI": "<MILVUS_URI>",
        "MILVUS_TOKEN": "<MILVUS_TOKEN>"
      }
    }
  }
}
Or via CLI
bash
claude mcp add milvus -- python /path/to/mcp-server-milvus/src/mcp_server_milvus/server.py --milvus-uri http://localhost:19530

Available tools

ToolDescription
milvus_text_searchPerforms full-text (BM25) search over a collection with configurable result limits and output fields.
milvus_vector_searchRuns dense-vector similarity search with selectable distance metrics and optional range filtering.
milvus_hybrid_searchCombines BM25 text search and vector similarity search, fusing the results with Reciprocal Rank Fusion.
milvus_text_similarity_searchPerforms text-based similarity search against embedded text fields in a collection (Milvus 2.6.0+).
milvus_queryExecutes a filter-expression query to retrieve entities matching scalar conditions.
milvus_list_collectionsLists all collections available in the current database.
milvus_create_collectionCreates a new collection with a custom schema, indexes, and vector dimensions.
milvus_get_collection_infoRetrieves the schema and structural metadata of a specific collection.
milvus_load_collectionLoads a collection into memory (with configurable replica count) so it can be queried.
milvus_release_collectionReleases a collection from memory to free resources.
milvus_insert_dataInserts records into a collection as a list of dictionary objects.
milvus_delete_entitiesDeletes entities from a collection matching a filter expression.
milvus_list_databasesLists all databases available in the Milvus instance.
milvus_use_databaseSwitches the active database context for subsequent operations.

Required configuration

  • MILVUS_URIRequired

    URI of the Milvus instance (e.g. http://localhost:19530 or a Zilliz Cloud endpoint).

  • MILVUS_TOKENOptional

    Authentication token for a secured Milvus or Zilliz Cloud cluster. Optional.

  • MILVUS_DBOptional

    Name of the Milvus database to use. Optional; defaults to 'default'.

What you can do with it

Hybrid semantic + keyword retrieval

Use milvus_hybrid_search so the agent fuses BM25 keyword matching with dense vector similarity, improving recall over a large embedded corpus for RAG.

Agent-managed vector store

Let the agent create collections, insert embeddings, and load or release them on demand, then query by vector or filter, turning Milvus into a programmable long-term memory.

FAQ

Is it free?
Yes. The MCP server is open source and free, and Milvus itself is open source and self-hostable at no cost. Zilliz Cloud is a managed option if you prefer not to run your own cluster.
Does it support remote/OAuth?
It runs locally over stdio by default and also supports SSE and streamable HTTP for networked use. It connects to Milvus with MILVUS_URI and an optional MILVUS_TOKEN; there is no separate OAuth flow.
← Browse all vector-search servers