MCP 與 LLM：提示與資源

Model Context Protocol（MCP）入門

James Chapman

AI Curriculum Manager, DataCamp

LLM 流程中的資源與提示

Resources：唯讀情境
Prompts：設定行為並為任務最佳化模型的指令範本

→ 續用 Anthropic Messages API

LLM 基本元件

提示—資源工作流程

提示與資源流程 1

提示—資源工作流程

提示與資源流程 2

提示—資源工作流程

提示與資源流程 3

提示—資源工作流程

提示與資源流程 4

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("Timezone Converter")

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

@mcp.resource("file://locations.txt")
def get_locations() -> str:
    # ...

@mcp.prompt(title="Timezone Conversion")
def convert_timezone_prompt(timezone_request: str) -> str:
    # ...

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

用戶端輔助函式

read_resource(resource_uri)：依 URI 取得資源內容
read_prompt(prompt_name, user_input)：取得提示範本並注入使用者請求

1. 取得資源與提示

async def get_context_from_mcp(user_query: str) -> tuple[str, str]:
    """Fetch resource content and prompt text from the MCP server."""
    params = StdioServerParameters(command=sys.executable, args=["timezone_server.py"])

    async with stdio_client(params) as (reader, writer):
        async with ClientSession(reader, writer) as session:
            await session.initialize()

            # Get the resource (supported locations)
            resource_result = await session.read_resource("file://locations.txt")
            resource_text = resource_result.contents[0].text

            # Get the prompt with the user's query
            prompt_result = await session.get_prompt("convert_timezone_prompt",
                                                     arguments={"timezone_request": user_query})
            prompt_text = prompt_result.messages[0].content.text

            return resource_text, prompt_text

2. 建立系統訊息並呼叫 LLM

async def call_llm_with_context(user_query: str):
    """Call the LLM with resource and prompt context from MCP."""

    resource_text, prompt_text = await get_context_from_mcp(user_query)

    # Combine prompt (task + rules + user request) with resource (supported locations)
    full_prompt = prompt_text + "\n\nSupported locations:\n" + resource_text

    client = AsyncAnthropic(api_key="<ANTHROPIC_API_TOKEN>")
    response = await client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": full_prompt}],
        tools=anthropic_tools,  # from get_tools_from_mcp(), formatted for Anthropic
    )

3. 處理回應

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

    if response.stop_reason == "tool_use":
        tool_use = next(b for b in response.content if b.type == "tool_use")
        result = await call_mcp_tool(
            tool_use.name, tool_use.input)
        # ... send the tool_result back in a follow-up
        #     request, then print the reply, as before

範例：含糊的請求

if __name__ == "__main__":
    asyncio.run(call_llm_with_context("What time is it in Canada?"))

Assistant: 加拿大有多個時區。你指的是哪個城市或地區？
例如多倫多、溫哥華，或哈利法克斯？

範例：明確的請求

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

Assistant: 里斯本現在也是 9:50 AM。

重點回顧：LLM 的資源與提示

提示與資源流程 4

一起來練習吧！

Model Context Protocol（MCP）入門