Milvus MCP

Zilliz maintains an MCP server for Milvus that lets an AI client operate a vector database directly, which is useful both for building retrieval systems and for debugging the ones already running. Search coverage is the core: milvus_vector_search performs similarity search on a collection, milvus_text_search runs full text search, milvus_hybrid_search combines dense and sparse signals in one request, milvus_text_similarity_search handles text driven similarity, and milvus_query retrieves entities using filter expressions rather than a vector. Collection management sits alongside it. milvus_list_collections enumerates what exists in the database, milvus_create_collection defines a new collection with a custom schema, milvus_get_collection_info returns metadata for one collection, and milvus_load_collection and milvus_release_collection move a collection in and out of memory, which is the step people most often forget when a search returns nothing. Data tools milvus_insert_data and milvus_delete_entities write records and remove them by filter expression, so an agent can seed a test collection and clean it up. The server is a Python package run through the uv package manager or Docker, pointed at an instance through the Milvus URI with an optional token for authenticated deployments, and it supports stdio, SSE and streamable HTTP transports so it works both as a local subprocess and as a shared network service. It fits engineers building retrieval augmented applications on Milvus or Zilliz Cloud who want an assistant that can inspect and repair collections without a separate admin console.