How to add Milvus MCP Server to Windsurf

Connect LLMs to the Milvus vector database for vector, text, and hybrid search plus collection management. Paste the config into ~/.codeium/windsurf/mcp_config.json and restart Windsurf.

Last updated June 14, 2026 · 233★ · stdio · apikey · official

Windsurf config for Milvus MCP Server

{
  "mcpServers": {
    "milvus-mcp-server": {
      "command": "/PATH/TO/uv",
      "args": [
        "--directory",
        "/path/to/mcp-server-milvus/src/mcp_server_milvus",
        "run",
        "server.py",
        "--milvus-uri",
        "http://localhost:19530"
      ]
    }
  }
}

Setup steps

1Open Windsurf → Cascade → the hammer/MCP icon → Configure (or edit ~/.codeium/windsurf/mcp_config.json).
2Paste the Milvus MCP Server config below.
3Fill in placeholder secrets, then save.
4Click Refresh in the MCP panel.
5Milvus MCP Server's tools become available to Cascade.

Before you start

Python 3.10 or higher
A running Milvus instance (local or remote)
uv installed (recommended for running the server)

What Milvus MCP Server can do in Windsurf

milvus_text_search

Search for documents using full text search. Parameters: collection_name, query_text, limit (default 5), output_fields, drop_ratio (proportion of low-frequency terms to ignore, 0.0-1.0, default 0.2).

milvus_vector_search

Perform vector similarity search on a collection. Parameters: collection_name, vector, vector_field (default 'vector'), limit (default 5), output_fields, filter_expr, metric_type (COSINE, L2, IP; default COSINE), radius, range_filter.

milvus_hybrid_search

Perform hybrid (text + vector) search on a collection. Parameters: collection_name, query_text, text_field, vector, vector_field, limit (default 5), output_fields, filter_expr, sparse_radius, sparse_range_filter, dense_radius, dense_range_filter.

milvus_text_similarity_search

Perform text similarity search on a collection (Milvus 2.6.0+ only; requires an embedding function set on the Milvus server). Parameters: collection_name, query_text, anns_field, limit (default 5), output_fields, metric_type (default COSINE), filter_expr, radius, range_filter.

milvus_query

Query a collection using filter expressions. Parameters: collection_name, filter_expr (e.g. 'age > 20'), output_fields, limit (default 10).

milvus_list_collections

List all collections in the database.

milvus_create_collection

Create a new collection with quick setup or a customized schema. Parameters include collection_name, auto_id (default True), dimension (default 768), primary_field_name (default 'id'), vector_field_name (default 'vector'), metric_type (default COSINE), field_schema, index_params, and other_kwargs.

milvus_load_collection

Load a collection into memory for search and query. Parameters: collection_name, replica_number (default 1).

Security

Authentication to Milvus is optional and configured via the MILVUS_TOKEN environment variable. If your Milvus instance requires auth, set MILVUS_TOKEN; ensure the configured account has only the permissions needed for the operations you intend to perform. Connection URIs and tokens are passed via CLI args or a local .env file in src/mcp_server_milvus/.

Milvus MCP Server + Windsurf FAQ

Where is the Windsurf config file?

Windsurf reads MCP servers from ~/.codeium/windsurf/mcp_config.json. Paste the Milvus MCP Server config there under the "mcpServers" key and restart the client.

Is Milvus MCP Server safe to use with Windsurf?

Which transports does the server support?

Three: stdio (default), SSE (HTTP Server-Sent Events), and Streamable HTTP (recommended for production, with an optional stateless mode). The transport is selected with CLI flags such as --sse or --streamable-http.

How do I authenticate to Milvus?

Authentication is optional. Set the MILVUS_TOKEN environment variable if your Milvus instance requires it. The Milvus URI is set with --milvus-uri or the MILVUS_URI variable, and the database with MILVUS_DB (defaults to 'default').

Do I need to install the package?

No. The recommended approach is to clone the repo and run the server directly with uv (e.g. uv run src/mcp_server_milvus/server.py --milvus-uri http://localhost:19530); both Claude Desktop and Cursor are configured to launch it this way.

View repo Full Milvus MCP Server page