MCP数据库查询服务

MCP Dummy DB Integration POC

This project demonstrates how to use the Model Context Protocol (MCP) to act as a connector between an LLM/Client and a database.

Overview

The system consists of three main components:

1. Database: A dummy SQLite database (dummy.db) containing employees, projects, and issues.
2. MCP Server (mcp_server.py): Exposes tools to query the database securely. It uses the mcp Python SDK to define tools like query_database and list_tables.
3. MCP Client (client.py): Simulates an Agent/LLM that connects to the server via standard input/output (stdio) and calls tools to retrieve data.

Prerequisites

• Python 3.10+
• mcp package

Setup

1. Install Dependencies:

   复制

   pip install -r requirements.txt
   

2. Initialize Database: Run the setup script to create dummy.db and populate it with sample data.

   复制

   python setup_db.py
   

Running the Demo

Run the client script to simulate the data retrieval flow:

python client.py

Expected Output

The client will:

1. Connect to the MCP server.
2. List available tools.
3. List tables in the database.
4. Execute a SELECT query to fetch employees in the 'AI' department.
5. Demonstrate error handling for invalid or forbidden queries.

Agentic Workflow Demo

The project includes an agent_workflow.py script that demonstrates a simulated Agentic flow:

python agent_workflow.py

How it works:

1. User Intent: The script accepts natural language queries (e.g., "Show me all employees in the AI department").
2. Mock LLM: A simple logic block simulates an LLM converting the intent into SQL.
3. Tool Selection: The agent decides whether to use query_database or list_tables.
4. Execution: The agent calls the MCP server to fetch data.
5. Response: The agent formats and displays the results.

Architecture Flow

Client (Request) -> MCP Protocol (JSON-RPC) -> MCP Server -> SQLite DB

• Security: The server restricts queries to SELECT only, preventing modification or deletion of data.
• Abstraction: The client doesn't need to know the database location or credentials, only the tool definition.