NeonDB MCP数据库助手

NeonDB MCP Server with LangChain Agent

A Model Context Protocol (MCP) server that enables AI agents to query and interact with a NeonDB PostgreSQL database. This project demonstrates how to build an intelligent database assistant using MCP, LangChain, and Google's Gemini AI.

🌟 Features

• MCP Server: Exposes database operations as tools via the Model Context Protocol
• AI-Powered Queries: Natural language database interaction using Google Gemini
• Read-Only Safety: Built-in security to prevent destructive operations
• Conversational Memory: Multi-turn conversations with context retention
• Interactive CLI: Terminal-based chat interface for database queries

🏗️ Architecture

┌─────────────────┐
│  User (CLI)     │
└────────┬────────┘
         │
         v
┌─────────────────┐
│  LangChain      │
│  React Agent    │ ← Powered by Google Gemini
└────────┬────────┘
         │
         v
┌─────────────────┐
│  MCP Adapter    │
└────────┬────────┘
         │
         v
┌─────────────────┐
│  MCP Server     │ ← server.py
│  (FastMCP)      │
└────────┬────────┘
         │
         v
┌─────────────────┐
│  NeonDB         │
│  (PostgreSQL)   │
└─────────────────┘

📋 Prerequisites

• Python 3.8+
• A NeonDB account (sign up here)
• Google AI API key (get one here)

🚀 Installation

1. Clone the repository:

   复制

   git clone <your-repo-url>
   cd database_mcp
   

2. Create a virtual environment:

   复制

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate
   

3. Install dependencies:

   复制

   pip install -r requirements.txt
   

4. Set up environment variables:

   Create a .env file in the project root:

   复制

   NEON_DATABASE_URL='your_neondb_connection_string'
   GOOGLE_API_KEY='your_google_ai_api_key'
   

5. Seed the database:

   复制

   python seed_db.py
   

🎯 Usage

One-Shot Query (Test Mode)

Run a single query to test the setup:

python test.py

This will ask the agent to analyze your database schema and find the largest table.

Interactive Chat Mode

Start a conversational session with persistent memory:

python client.py

Example conversation:

You: What tables do I have?
Agent: You have two tables: 'users' and 'tasks'...

You: How many users are there?
Agent: There are 3 users in the database...

You: Show me all pending tasks
Agent: Here are the pending tasks: ...

Type quit, exit, or q to end the session.

📁 Project Structure

• server.py: MCP server with database tools (get_database_schema, run_read_only_query)
• client.py: Interactive CLI client with conversation memory
• test.py: One-shot testing script
• seed_db.py: Database initialization script (creates users and tasks tables)
• requirements.txt: Python dependencies

🛠️ Available MCP Tools

The MCP server exposes two tools to the AI agent:

1. get_database_schema()

Returns the complete schema of all tables in the public schema, including column names and data types.

2. run_read_only_query(sql_query: str)

Executes a SELECT query against the database. Automatically blocks non-SELECT statements for security.

🔒 Security Features

• Read-only queries: Only SELECT statements are permitted
• SQL injection protection: Uses parameterized queries via asyncpg
• Connection pooling: Efficient connection management
• Environment variables: Sensitive credentials stored in .env (not committed to Git)

🧪 Sample Database Schema

The seed_db.py script creates:

users table:

• user_id (SERIAL PRIMARY KEY)
• name (VARCHAR)
• email (VARCHAR, UNIQUE)
• created_at (TIMESTAMP)

tasks table:

• task_id (SERIAL PRIMARY KEY)
• user_id (FOREIGN KEY → users)
• title (VARCHAR)
• status (VARCHAR)
• due_date (DATE)

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Model Context Protocol (MCP)
• FastMCP
• LangChain
• NeonDB
• Google Gemini

📧 Support

For questions or issues, please open an issue on GitHub.

⚠️ Important: Never commit your .env file to version control. The .gitignore file is configured to prevent this.