Gemini Context MCP Server

A powerful MCP (Model Context Protocol) server implementation that leverages Gemini's capabilities for context management and caching. This server maximizes the value of Gemini's 2M token context window while providing tools for efficient caching of large contexts.

🚀 Features

Context Management

Up to 2M token context window support - Leverage Gemini's extensive context capabilities
Session-based conversations - Maintain conversational state across multiple interactions
Smart context tracking - Add, retrieve, and search context with metadata
Semantic search - Find relevant context using semantic similarity
Automatic context cleanup - Sessions and context expire automatically

API Caching

Large prompt caching - Efficiently reuse large system prompts and instructions
Cost optimization - Reduce token usage costs for frequently used contexts
TTL management - Control cache expiration times
Automatic cleanup - Expired caches are removed automatically

🏁 Quick Start

Prerequisites

Node.js 18+ installed
Gemini API key (Get one here)

Installation

# Clone the repository
git clone https://github.com/ogoldberg/gemini-context-mcp-server
cd gemini-context-mcp

# Install dependencies
npm install

# Copy environment variables example
cp .env.example .env

# Add your Gemini API key to .env file
# GEMINI_API_KEY=your_api_key_here

Basic Usage

# Build the server
npm run build

# Start the server
node dist/mcp-server.js

MCP Client Integration

This MCP server can be integrated with various MCP-compatible clients:

Claude Desktop - Add as an MCP server in Claude settings
Cursor - Configure in Cursor's AI/MCP settings
VS Code - Use with MCP-compatible extensions

For detailed integration instructions with each client, see the MCP Client Configuration Guide in the MCP documentation.

Quick Client Setup

Use our simplified client installation commands:

# Install and configure for Claude Desktop
npm run install:claude

# Install and configure for Cursor
npm run install:cursor

# Install and configure for VS Code
npm run install:vscode

Each command sets up the appropriate configuration files and provides instructions for completing the integration.

💻 Usage Examples

For Beginners

Directly using the server:

Start the server:
```
node dist/mcp-server.js
```

Interact using the provided test scripts:

# Test basic context management
node test-gemini-context.js

# Test caching features
node test-gemini-api-cache.js

Using in your Node.js application:

import { GeminiContextServer } from './src/gemini-context-server.js';

async function main() {
  // Create server instance
  const server = new GeminiContextServer();
  
  // Generate a response in a session
  const sessionId = "user-123";
  const response = await server.processMessage(sessionId, "What is machine learning?");
  console.log("Response:", response);
  
  // Ask a follow-up in the same session (maintains context)
  const followUp = await server.processMessage(sessionId, "What are popular algorithms?");
  console.log("Follow-up:", followUp);
}

main();

For Power Users

Using custom configurations:

// Custom configuration
const config = {
  gemini: {
    apiKey: process.env.GEMINI_API_KEY,
    model: 'gemini-2.0-pro',
    temperature: 0.2,
    maxOutputTokens: 1024,
  },
  server: {
    sessionTimeoutMinutes: 30,
    maxTokensPerSession: 1000000
  }
};

const server = new GeminiContextServer(config);

Using the caching system for cost optimization:

// Create a cache for large system instructions
const cacheName = await server.createCache(
  'Technical Support System',
  'You are a technical support assistant for a software company...',
  7200 // 2 hour TTL
);

// Generate content using the cache
const response = await server.generateWithCache(
  cacheName,
  'How do I reset my password?'
);

// Clean up when done
await server.deleteCache(cacheName);

🔌 Using with MCP Tools (like Cursor)

This server implements the Model Context Protocol (MCP), making it compatible with tools like Cursor or other AI-enhanced development environments.

Available MCP Tools

Context Management Tools:
- generate_text - Generate text with context
- get_context - Get current context for a session
- clear_context - Clear session context
- add_context - Add specific context entries
- search_context - Find relevant context semantically
Caching Tools:
- mcp_gemini_context_create_cache - Create a cache for large contexts
- mcp_gemini_context_generate_with_cache - Generate with cached context
- mcp_gemini_context_list_caches - List all available caches
- mcp_gemini_context_update_cache_ttl - Update cache TTL
- mcp_gemini_context_delete_cache - Delete a cache

Connecting with Cursor

When used with Cursor, you can connect via the MCP configuration:

{
  "name": "gemini-context",
  "version": "1.0.0",
  "description": "Gemini context management and caching MCP server",
  "entrypoint": "dist/mcp-server.js",
  "capabilities": {
    "tools": true
  },
  "manifestPath": "mcp-manifest.json",
  "documentation": "README-MCP.md"
}

For detailed usage instructions for MCP tools, see README-MCP.md.

⚙️ Configuration Options

Environment Variables

Create a .env file with these options:

# Required
GEMINI_API_KEY=your_api_key_here
GEMINI_MODEL=gemini-2.0-flash

# Optional - Model Settings
GEMINI_TEMPERATURE=0.7
GEMINI_TOP_K=40
GEMINI_TOP_P=0.9
GEMINI_MAX_OUTPUT_TOKENS=2097152

# Optional - Server Settings
MAX_SESSIONS=50
SESSION_TIMEOUT_MINUTES=120
MAX_MESSAGE_LENGTH=1000000
MAX_TOKENS_PER_SESSION=2097152
DEBUG=false

🧪 Development

# Build TypeScript files
npm run build

# Run in development mode with auto-reload
npm run dev

# Run tests
npm test

📚 Further Reading

For MCP-specific usage, see README-MCP.md
Explore the manifest in mcp-manifest.json to understand available tools
Check example scripts in the repository for usage patterns

📋 Future Improvements

Database persistence for context and caches
Cache size management and eviction policies
Vector-based semantic search
Analytics and metrics tracking
Integration with vector stores
Batch operations for context management
Hybrid caching strategies
Automatic prompt optimization

📄 License

MIT