Milvus MCP Server

This program is a Milvus vector database access service designed for large language models, following the Model Context Protocol (MCP) standard. It can serve as an external vector database interaction tool for large language models (such as Antigravity).

It supports Milvus 2.6 vector databases and includes a built-in OpenAI-compatible Embedding service (e.g., Tongyi Qianwen text-embedding-v4), allowing large models to perform semantic searches directly using natural language.

💡 Key Features

1. Vector Search: Supports semantic search for text input (automatic embedding) and raw vector search.

2. Built-in Embedding: Integrates an OpenAI-compatible endpoint; large models can pass text to be automatically converted into vectors, eliminating the need for external embedding services.

3. Collection Lifecycle Management: Create, view, and delete collections, with automatic indexing and loading upon creation.

4. Data Read/Write: Supports vector data insertion and conditional deletion.

5. Multi-connection Management: Define multiple named connections via a JSON configuration file; a single MCP Server instance can serve multiple Milvus instances.

6. Security Design: drop_collection and delete_entities are independent tools that can be toggled separately in the MCP client; drop_collection additionally requires a confirmDrop=true parameter as a double-safety measure.

7. Hybrid Search: Supports combined queries of vector similarity search + scalar filtering.

Related MCP server: Qdrant MCP Server

⚙️ Requirements

1. Node.js Environment: Requires Node.js version v18 or higher.

2. Milvus Instance: Requires an accessible Milvus 2.6 service instance.

3. Embedding API: Requires an OpenAI-compatible embedding service endpoint (e.g., Tongyi Qianwen Bailian platform).

4. Dependencies: Run npm install --omit=dev on the host machine to install dependencies.

🛠 Installation and Build

1. Build from Source

npm install
npm run build

Build artifacts will be stored in the ./dist directory.

2. Global Local Installation

npm link

After execution, you can use the global command milvus-mcp-server to quickly invoke the service from anywhere in the system.

📝 Configuration File

Use a JSON file to configure Milvus connections and the embedding service.

Configuration Example

{
  "connections": [
    {
      "name": "my-milvus",
      "address": "localhost:19530",
      "username": "root",
      "password": "Milvus",
      "database": "default",
      "description": "本地开发 Milvus 实例，存储文档和图片的向量索引"
    }
  ],
  "embedding": {
    "baseUrl": "https://dashscope.aliyuncs.com/compatible-mode/v1",
    "apiKey": "sk-your-api-key",
    "model": "text-embedding-v4",
    "dimensions": 1024
  }
}

Connection Field Descriptions

Embedding Field Descriptions

🚀 Startup and Usage

milvus-mcp-server --config <path-to-config.json>

Parameter List:

• -c, --config <path>: Configuration file path (required).

🧠 MCP Client Configuration Example

Taking a standard MCP Client (like Antigravity) as an example:

{
  "mcpServers": {
    "MilvusMCP": {
      "command": "node",
      "args": [
        "绝对路径/milvus_mcp_server/dist/index.js",
        "--config", "绝对路径/milvus-config.json"
      ]
    }
  }
}

🧰 Tools Available to the Model

1. list_milvus_connections

• Description: Lists all configured Milvus connection information (sanitized, excluding passwords).

• Parameters: None.

2. list_collections

• Description: Lists all collections under a specified connection.

• Parameters:

  • connectionName (String): Target connection name.

3. describe_collection

• Description: Retrieves detailed schema information for a collection, including field definitions, indexes, and row counts.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Collection name.

4. create_collection

• Description: Creates a new collection. Automatically creates an AUTOINDEX for the vector field and loads the collection, making it immediately searchable.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): New collection name.

  • description (String, optional): Collection description.

  • fields (Array): Array of field definitions, each containing name, dataType, isPrimaryKey?, autoId?, dimension?, maxLength?, description?.

5. drop_collection

• Description: Permanently deletes a collection and all its data. Irreversible operation.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Name of the collection to delete.

  • confirmDrop (Boolean): Must be set to true to execute the deletion. Safety double-confirmation mechanism.

• Toggle Control: This tool can be enabled or disabled individually in the MCP client.

6. insert_vectors

• Description: Inserts data into a collection. Each data item is a JSON object containing field values (including the vector field).

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Target collection name.

  • data (Array): Array of data.

7. delete_entities

• Description: Deletes entities based on a filter expression.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Target collection name.

  • filter (String): Milvus filter expression, e.g., 'id in [1, 2, 3]'.

• Toggle Control: This tool can be enabled or disabled individually in the MCP client.

8. vector_search

• Description: Vector similarity search. Supports passing text (automatically calls the embedding service to convert to a vector) or a raw vector.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Target collection name.

  • text (String, optional): Search text, choose one between text and vector.

  • vector (Number[], optional): Raw search vector, choose one between text and vector.

  • topK (Number, optional): Number of results to return, defaults to 10.

  • filter (String, optional): Scalar filter expression.

  • outputFields (String[], optional): List of fields to return.

9. query_entities

• Description: Queries entities based on scalar conditions, not involving vector search.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Target collection name.

  • filter (String): Milvus filter expression.

  • outputFields (String[], optional): List of fields to return.

  • limit (Number, optional): Maximum number of results to return, defaults to 100.

10. hybrid_search

• Description: Combined query of vector search + scalar filtering. The difference from vector_search is that filter is a required parameter.

• Parameters:

  • connectionName (String): Target connection name.

  • collectionName (String): Target collection name.

  • text (String, optional): Search text, choose one between text and vector.

  • vector (Number[], optional): Raw search vector, choose one between text and vector.

  • filter (String): Required, scalar filter expression.

  • topK (Number, optional): Number of results to return, defaults to 10.

  • outputFields (String[], optional): List of fields to return.