MCP (Multi-Component Platform) Integration

The MCP module enables LLMs to connect to external services via tool calling functionality, allowing seamless integration with various backend services.

Key Features

• Server Management: Connect to multiple MCP servers simultaneously
• Tool Discovery: Automatic tool listing from connected servers
• Unified Interface: Call tools without knowing which server provides them
• Transport Flexibility: Supports multiple connection types:
  • WebSocket (ws)
  • Server-Sent Events (sse)
  • Standard I/O (stdio)
• Provider Integration: Works with:
  • OpenAI
  • Gemini
  • Deepseek
  • Anthropic

Configuration

Configure MCP servers in your LLM config:

llm:
  mcp_config_path: "./mcp_config.json"  # Path to MCP configuration
  max_tool_calls_per_response: 3       # Maximum tool calls per LLM response

Example MCP config (mcp_config.json):

{
  "mcpServers": {
    "search-server": {
      "transport_type": "ws",
      "url": "ws://localhost:8080"
    },
    "data-server": {
      "transport_type": "stdio",
      "command": "python",
      "args": ["data_server.py"]
    },
    "brave-search": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-brave-search"
      ],
      "env": {
        "BRAVE_API_KEY": "BSApeUEgTfLYRikQwW8NAd4Qv8wK2v2"
      }
    }
  }
}

Usage

Basic workflow:

1. Initialize LLM with MCP config
2. The provider will automatically:
   • Connect to configured servers
   • Discover available tools
   • Handle tool calling when needed

Asynchronous Example

from aicore.config import Config
from aicore.llm import Llm
import asyncio

async def main():
    config = Config.from_yaml()
    llm = Llm.from_config(config.llm)
    
    response = await llm.acomplete(
        "Search for Elon Musk news",
        system_prompt="Use available tools when needed",
        agent_id="mcp-agent"
    )
    print(response)

asyncio.run(main())

Tool Calling Workflow

1. Initialization:

   • LLM loads MCP configuration
   • Connects to all configured servers
   • Discovers available tools

2. Execution:

   • User makes a request that may require tool usage
   • LLM determines if tools should be called
   • If tools are needed:
     • Calls appropriate tools via MCP
     • Processes tool responses
     • May make additional tool calls if needed
   • Returns final response combining tool outputs

Error Handling

The MCP module provides robust error handling:

• Connection Errors: Automatic retries with exponential backoff
• Tool Errors: Detailed error messages including:
  • Which tool failed
  • Error details from the server

Best Practices

1. Tool Naming: Use clear, descriptive names for tools
2. Error Handling: Implement comprehensive error handling in your tools
3. Rate Limiting: Consider rate limits when making multiple tool calls
4. Testing: Test tools thoroughly before production use
5. Monitoring: Use the observability features to track tool usage

Advanced Features

Custom Transports

You can implement custom transport types by extending the FastMCPTransport class.

Tool Schema Validation

All tools are validated against their schema before being called, ensuring proper input formatting.

Parallel Tool Execution

The MCP client can handle parallel tool execution when multiple tools are called simultaneously.

For more examples, see the MCP Examples documentation.