CommencerCommencez gratuitement

Delegate the summary with sampling

Your server's summarize tool needs a language model to condense some text — but you don't want the server holding an API key. With sampling, the server asks the connected client to run the prompt instead, and the client returns the generated text. You also want the summary kept to a single sentence, so you'll steer the model with a system prompt.

FastMCP, Context, SamplingMessage, and TextContent are imported and the server mcp is created. run_tool(text_to_summarize) runs your tool through a real MCP client–server session; the connected client fulfills each sampling request by calling a real language model and returning its reply.

Cet exercice fait partie du cours

<cours>Model Context Protocol: Advanced Topics</cours>
Voir le cours

Instructions de l’exercice

  • Call ctx.session.create_message() to ask the client to run the prompt.
  • Steer the model by passing a system_prompt that tells it to reply with a single-sentence summary.
  • Return the generated text from the result with result.content.text.

Exercice interactif pratique

Essayez cet exercice en complétant ce code d’exemple.

@mcp.tool()
async def summarize(text_to_summarize: str, ctx: Context):
    prompt = f"Please summarize the following text:\n{text_to_summarize}"

    # Ask the client's model to summarize, steered by a system prompt
    result = await ctx.session.____(
        messages=[
            SamplingMessage(role="user", content=TextContent(type="text", text=prompt))
        ],
        system_prompt=____,
        max_tokens=1000,
    )

    # Return the generated text
    return result.____.text


# Run the tool through a real MCP client-server session
summary = run_tool("lots of text")
print(summary)
Modifier et exécuter le code