Gemini MCP Server

# CLAUDE.md This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository. ## Project Overview This is an MCP (Model Context Protocol) server that exposes Google Gemini AI models through a standardized interface. The server runs as an HTTP endpoint and allows MCP clients to call Gemini models for text generation. ## Development Commands **Start the server:** ```bash npm run dev ``` Server starts on port 3333 by default (configurable via `PORT` environment variable). **Run the test client:** ```bash npm test ``` This executes `test.ts` which connects to the running server and calls the `gemini.generateText` tool. **Note:** The server must be running (via `npm run dev`) before running tests. ## Environment Configuration Required environment variable: - `GEMINI_API_KEY` - Google Gemini API key (configured in `.env`) Optional server configuration: - `PORT` - Server port (default: 3333) Optional test configuration: - `MCP_URL` - MCP server URL (default: http://localhost:3333/mcp) - `TEST_PROMPT` - Prompt to test with (default: '俳句を一つ作って。') - `TEST_MODEL` - Gemini model to use (default: 'gemini-2.5-pro') - `TEST_TEMPERATURE` - Temperature setting (default: 0.7) ## Architecture **Server Architecture** (`server.ts`): - Express server with a single POST endpoint at `/mcp` - Uses MCP SDK's `StreamableHTTPServerTransport` for HTTP-based MCP communication - Registers one tool: `gemini.generateText` - Tool uses Google Gen AI SDK to call Gemini models with configurable parameters (model, temperature) - Response extraction handles multiple Google Gen AI SDK version differences using defensive type checking **Tool Schema** (`gemini.generateText`): - Input validated with Zod schema - Parameters: `prompt` (string, required), `model` (string, default: 'gemini-2.5-pro'), `temperature` (number 0-2, default: 1) - Returns structured content with model, temperature, and generated text **Test Client** (`test.ts`): - Uses MCP SDK's `StreamableHTTPClientTransport` to connect to the server - Lists available tools and calls `gemini.generateText` - Demonstrates the full client-server MCP interaction pattern ## Key Implementation Details - The server uses SDK 1.20.x which requires passing `Input.shape` (ZodRawShape) to `inputSchema` instead of the full Zod object - Type handling for Google Gen AI SDK is defensive (uses `any` casts) to handle version differences - Text extraction from Gemini responses handles multiple possible response formats for compatibility - Transport cleanup is tied to Express response close event to prevent resource leaks