MCP Agent Package

A Python package for interacting with LLMs via the MCP protocol.

Directory Structure

mcp_agent_package/
├── orderli_recursive_mcp/            # Main package directory
│   ├── __init__.py       # Package exports
│   ├── client.py         # Main client functionality
│   ├── config.py         # Configuration settings
│   ├── utils.py          # Utility functions
│   └── README.md         # Package documentation
├── main.py               # Example script to test the package
├── check_server.py       # Utility to check server connection
├── setup.py              # Package installation config
└── .gitignore            # Git ignore file

Installation

From within this directory, run:

pip install -e .

Usage

Basic Import and Usage

import asyncio
from orderli_recursive_mcp import call_llm

# Define a response handler to process streamed responses
async def response_handler(text_chunk: str, event_type: str) -> None:
    """Handle different types of response events."""
    if event_type == "text":
        # Main LLM response text
        print(text_chunk, end="")
    elif event_type == "tool_call":
        # Tool being called
        print(f"\n[TOOL CALL] {text_chunk}")
    elif event_type == "tool_result":
        # Result from tool execution
        print(f"[RESULT] {text_chunk}")
    elif event_type == "info":
        # Informational messages
        print(f"\n[INFO] {text_chunk}")
    elif event_type in ("error", "warning"):
        # Error and warning messages 
        print(f"\n[{event_type.upper()}] {text_chunk}")

async def main():
    # User prompt/question
    prompt = "What is the meaning of life?"
    
    # Call the LLM with the prompt
    session = await call_llm(
        prompt=prompt,
        response_callback=response_handler,
        # Optional parameters:
        # server_url="http://localhost:8001/sse",  # Custom MCP server URL
        # api_key="your_anthropic_api_key",        # Custom API key
        # disable_console_output=True,             # Disable internal console output
        # debug=False                              # Debug the MCP server connection
    )
    
    # You can continue the conversation with the same session
    # await session.chat("Tell me more about that.")

if __name__ == "__main__":
    asyncio.run(main())

Advanced Usage with Custom Tool Calls

To view an example, run:

python test.py

To check if the MCP server is running properly:

python check_server.py

Requirements

• Anthropic API key (can be set as environment variable ANTHROPIC_API_KEY or passed directly)
• Running MCP server using SSE
• Python 3.8 or higher

Configuration

The package uses the following environment variables:

• ANTHROPIC_API_KEY: Your Anthropic API key (required)
• ANTHROPIC_MODEL: Anthropic model to use (default: "claude-3-5-sonnet-latest")
• MCP_SERVER_URL: URL of the MCP server (default: http://localhost:8001/sse)
• MCP_MAX_RECURSIONS: Maximum number of tool use recursions (default: 15)
• MCP_MAX_RETRIES: Maximum number of connection retries (default: 3)
• MCP_RETRY_DELAY: Delay between connection retries in seconds (default: 2.0)
• MCP_TOOL_TIMEOUT: Tool execution timeout in seconds (default: 30.0)

Troubleshooting

If you encounter issues with tool calls or server connections:

1. Check Server Connection

Run the server connection check utility:

python check_server.py

This will verify that:

• The MCP server is reachable
• The server has the necessary tools available
• Basic tool execution works correctly

2. Test Individual Tools

If specific tools are failing, you can test them directly using the test tool utility:

python test_tool.py <tool_name> --args '<json_args>'

3. Common Issues with Database Tools

Database tools like list_schemas will fail if:

• The database connection is not properly configured on the server
• Database credentials are incorrect
• The database service is not running
• There are permission issues for the database user

4. Debugging Tool Call Failures

When a tool call fails, check the following:

1. Server Logs: Look at the MCP server logs for detailed error messages
2. Tool Arguments: Make sure you're providing all required arguments with the correct names
3. Connection Issues: Verify that the connection to the MCP server is stable
4. API Key: Confirm your Anthropic API key is valid and has sufficient quota