MCP und LLMs: Tools

Einführung in das Model Context Protocol (MCP)

James Chapman

AI Curriculum Manager, DataCamp

MCP-Tools zu LLM-Tools

Tool-↔-LLM-Integration ist stark LLM-abhängig

LLMs akzeptieren Tools auf verschiedene Arten und mit unterschiedlicher Syntax
Nutze Claude über die Anthropic Messages API
Fokus: Prozesse und Transformationen

Der Tool-Calling-Workflow

MCP-Server: timezone_server.py

from mcp.server.fastmcp import FastMCP
import requests

mcp = FastMCP("Timezone Converter")

@mcp.tool()
def convert_timezone(date_time: str, from_timezone: str, to_timezone: str) -> str:
    ...

if __name__ == "__main__":
    mcp.run(transport="stdio")

Client-Code erweitern

async def get_tools_from_mcp():
            # ... 
            return response.tools

async def call_mcp_tool(tool_name: str, arguments: dict) -> str:        
            # ...
            result = await session.call_tool(tool_name, arguments)
            return str(result.content[0].text)


async def call_anthropic_llm(user_query: str):

Setup: Tools fürs LLM formatieren

async def call_anthropic_llm(user_query: str):
    """Call Claude with MCP tools."""


    mcp_tools = await get_tools_from_mcp()


    anthropic_tools = []
    for tool in mcp_tools:

        anthropic_tool = {
            "name": tool.name,
            "description": tool.description or "",
            "input_schema": tool.inputSchema,  # MCP uses JSON Schema format
        }

        anthropic_tools.append(anthropic_tool)

1. Query und Tools ans LLM senden

from anthropic import AsyncAnthropic

async def call_anthropic_llm(user_query: str):
    # ...


    client = AsyncAnthropic(api_key="<ANTHROPIC_API_TOKEN>")


    response = await client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": user_query}],
        tools=anthropic_tools,
    )

2. Auf Tool-Call prüfen

async def call_anthropic_llm(user_query: str):
    # ...

    if response.stop_reason == "tool_use":

        tool_use = next(b for b in response.content if b.type == "tool_use")
        name = tool_use.name
        args = tool_use.input

        print(f"Model decided to call: {name}")
        print(f"Arguments: {args}\n")

3. Tool aufrufen | 4. Follow-up-Nachricht

async def call_anthropic_llm(user_query: str):
    # ...

    if response.stop_reason == "tool_use":
        # ...


        result = await call_mcp_tool(name, args)

        tool_result = {"type": "tool_result",
            "tool_use_id": tool_use.id, "content": result}
        followup = await client.messages.create(
            model="claude-sonnet-4-6", max_tokens=1024,
            messages=[
                {"role": "user", "content": user_query},
                {"role": "assistant", "content": response.content},
                {"role": "user", "content": [tool_result]},
            ],
            tools=anthropic_tools,
        )

5. Die finale Antwort

async def call_anthropic_llm(user_query: str):
    # ...

    if response.stop_reason == "tool_use":
        # ...

        final = next((b.text for b in followup.content if b.type == "text"), None)
        if final:
            print(f"\nAssistant: {final}")
        else:
            print("No follow-up message from model.")


    else:
        text = next((b.text for b in response.content if b.type == "text"), "")
        print(f"\nAssistant: {text}")
        return str(text)

Die Funktion testen

if __name__ == "__main__":
    asyncio.run(call_anthropic_llm("It is 9:50 AM in the UK in January. What time 
    is it in Lisbon, Portugal?"))

Assistant: It's 9:50 AM in Lisbon as well.

Der Tool-Calling-Workflow

Lass uns üben!

Einführung in das Model Context Protocol (MCP)