原文地址: https://www.thatjeffsmith.com/archive/2025/12/can-i-build-an-agent-still-use-sqlcls-mcp-server-yes/ # 能否自定义AI代理并同时使用SQLcl的MCP服务器？完全可以！  这是一篇快速解答近期常见问题的技术短文："我们能否不依赖Copilot/Claude/Windsurf等现成工具，而是构建自己的AI代理？" 答案很明确：当然可以！即使是我这样的菜鸟（读者朋友们肯定比我强）也能通过Python编写AI代理，并通过Oracle SQLcl的MCP服务器连接到Oracle AI数据库。  ## SQLcl MCP服务器日志 这些日志会便捷地存储在你的数据库中 仅用少量Python代码，我就实现了一个可通过自然语言提问与Oracle数据库交互的聊天提示工具。 ```bash thatjeffsmith$ /Users/thatjeffsmith/langchain/.venv/bin/python main.py ============================================================ NL2SQL Chat Agent - Oracle Database Edition Powered by LangChain + Claude + MCP ============================================================ Initializing agent... Agent ready! Type 'exit' or 'quit' to end the session. You: Connect to Project Raptor 🦖 database and find the highest paid employee by salary Agent: ---------- MCP SERVER STARTUP ---------- MCP Server started successfully on Mon Dec 01 14:55:19 EST 2025 Press Ctrl+C to stop the server ---------------------------------------- Dec 01, 2025 2:55:20 PM ## Summary I successfully connected to the **Project Raptor 🦖** database and found the highest paid employee: **Highest Paid Employee:** - **Name:** Steven King - **Employee ID:** 100 - **Email:** SKING - **Job Title:** AD_PRES (President) - **Salary:** $24,000 - **Department:** Executive (Department ID: 90) Steven King is the President of the company and works in the Executive department, earning the highest salary of $24,000 among all employees in the database. You: kthxbye ``` ## Java开发者也能实现 如果你想用Java实现，可以参考Oracle开发者大使Anders的完整代码示例（他的代码比我专业得多！）  ## 关于使用的AI模型 Anders和我的实现都不是运行本地大语言模型。我使用Claude控制台API密钥访问其Sonnet 4模型，而Anders使用的是OpenAI的gpt-40-mini模型。 值得注意的是，如果你构建自己的AI代理，完全可以运行本地LLM——这样能确保数据不会离开你的环境。MCP服务器并不关心你使用什么模型或代理。 ## 系统提示模板 ```bash self.system_prompt = """You are an expert SQL assistant with access to an Oracle database. Your role is to: 1. Understand natural language questions about data 2. Generate accurate SQL queries to answer those questions 3. Execute the SQL using the available tools 4. Summarize the results in a clear, understandable way When a user asks a question: - First, use get_schema_info if you need to understand the database structure - Then generate the appropriate SQL query - Execute it using execute_sql - Finally, provide a clear summary of the results Always ensure your SQL queries are safe and follow Oracle SQL syntax. """ ``` ## MCP服务器配置 ```json { "command": "/opt/sqlcl/25.3/sqlcl/bin/sql", "args": ["-mcp"], "type": "stdio", "env": {} } ``` ## 测试MCP服务器 简单的Python测试脚本： ```python from src.agent import NL2SQLAgent import json # Load config with open('config/mcp_config.json') as f: config = json.load(f) # Test the agent agent = NL2SQLAgent(config) print('Agent initialized!') print() # Test a simple query response = agent.chat('List all available database connections') print('Response:', response) agent.close() ``` 测试结果： ```bash Agent initialized! ---------- MCP SERVER STARTUP ---------- MCP Server started successfully on Tue Dec 02 15:33:08 EST 2025 Press Ctrl+C to stop the server ---------------------------------------- Response: Here are all the available Oracle database connections: 1. **Project Raptor 🦖** 2. **Free26ai** 3. **🅾️ Autonomous Reporting** 4. **system** 5. **Project Payroll 💲** 6. **HR123** 7. **ADMIN** 8. **demo** 9. **windy** You can connect to any of these databases by specifying the connection name. ```