BlazeSQL MCP Server

BlazeSQL MCP Server

This project implements a Model Context Protocol (MCP) server using the @modelcontextprotocol/sdk that acts as a proxy to the BlazeSQL Natural Language Query API. It allows MCP-compatible clients (like Cursor, Claude 3 with tool use, the MCP Inspector, etc.) to interact with BlazeSQL using natural language.

Features

• Built using the modern McpServer helper class from the MCP SDK.
• Exposes the BlazeSQL Natural Language Query API as an MCP tool named blazesql_query.
• Uses zod for robust validation of tool input parameters.
• Handles API key authentication securely via environment variables.
• Communicates with clients using the standard MCP stdio transport.

Workflow Diagram

This diagram shows the sequence of interactions when a client uses the blazesql_query tool (Note: The internal server logic now uses McpServer which simplifies tool registration compared to the low-level handlers shown in the diagram):

Prerequisites

• Node.js (LTS version recommended)
• Yarn (Classic or Berry)
• A BlazeSQL account with an API Key (Team Advanced subscription required for the API).
• At least one database connection configured in your BlazeSQL account.
• BlazeSQL Natural Language Query API Documentation: https://help.blazesql.com/en/article/natural-language-query-api-1fgx4au/

Setup

1. Clone the Repository:
   git clone <repository-url> cd blaze-sql-mcp-server
2. Install Dependencies:
   yarn install
   This will install all necessary dependencies, including the @modelcontextprotocol/sdk, dotenv, and zod.
3. Configure Environment Variables:
   • Copy the example environment file:
     cp .env.sample .env
   • Edit the .env file:
     # .env BLAZE_API_KEY=YOUR_BLAZESQL_API_KEY_HERE
     Replace YOUR_BLAZESQL_API_KEY_HERE with your actual API key obtained from your BlazeSQL account settings.

Running the Server

1. Build the Server: Compile the TypeScript code to JavaScript:
   yarn build
2. Run the Server: Execute the compiled code:
   node build/index.js
   The server will start and log messages to stderr (you might see "API Key loaded successfully..." etc.). It is now listening for an MCP client connection via standard input/output (stdio).

Connecting an MCP Client

This server uses the stdio transport mechanism.

Using the MCP Inspector (Recommended for Testing)

1. Make sure the server is not already running separately.
2. Run the Inspector, telling it to launch your server:
   npx @modelcontextprotocol/inspector node build/index.js
3. The Inspector UI will launch, automatically connecting to your server.
4. Navigate to the "Tools" tab to interact with the blazesql_query tool.

Using Integrated Clients (Cursor, Claude 3, etc.)

1. Start the server in a terminal:
   node build/index.js
2. Configure the client: In your MCP client's settings, you need to add a custom server configuration.
   • Transport: Select stdio.
   • Command: Specify the exact command used to run the server. You need to provide the absolute path to node and the absolute path to the build/index.js file.
     • Example (macOS/Linux - adjust paths as needed): /usr/local/bin/node /Users/your_username/path/to/blaze-sql-mcp-server/build/index.js
     • You can find the path to node using which node in your terminal.
     • You can find the path to the project using pwd inside the project directory.
   • Save the configuration.
3. The client should now be able to connect to your locally running server and list/use its tools.

Using the blazesql_query Tool

Once connected, the client can call the blazesql_query tool.

• Tool Name: blazesql_query
• Arguments:
  • db_id (string, required): The ID of the target database connection in your BlazeSQL account. You can find this ID in the BlazeSQL web application when managing your database connections.
  • natural_language_request (string, required): The query you want to execute, written in plain English (e.g., "show me the total number of users"). (Input is validated using zod)
• Example Call (using mcp test syntax for illustration):
  call-tool blazesql_query --db_id "db_your_actual_db_id" --natural_language_request "What were the total sales last month?"
• Output: If successful, the tool returns a single text content block containing:
  • The natural language response from the BlazeSQL agent.
  • The generated SQL query within a Markdown code fence (sql ... ).
  • The data results formatted as JSON within a Markdown code fence (json ... ).

  Example structure within the text block:

  **Agent Response:** The total sales last month were $12345.67. **Generated SQL:** ```sql SELECT sum(sales_amount) FROM sales WHERE sale_date >= date('now', '-1 month');

  Data Result (JSON):

  [ { "sum(sales_amount)": 12345.67 } ]
  If unsuccessful, it returns a `text` content block containing the error message from the BlazeSQL API and marks the response as an error (`isError: true`).