Homework Solutions

Lesson 06 · MCP Servers

Solution to Task 1 — Search resource template

Reuse pkb.tools.search with its default limit, then JSON-encode the result so FastMCP can serialize it. Drop this into pkb/resources.py and register it with the server the same way you registered get_note in the lab:

import json
from pkb.tools import SearchArgs, search as _search


@mcp.resource("pkb://search/{query}")
def search_resource(query: str) -> str:
    """Search results as a JSON-encoded list of titles."""
    return json.dumps(_search(SearchArgs(query=query)))

Common mistake

Returning the raw list[str] — FastMCP resources expect a string (or bytes). JSON-encode the list with json.dumps(...).

Solution to Task 2 — Multi-message prompt

A prompt that returns list[PromptMessage] is FastMCP’s way of seeding a multi-turn conversation. Each entry carries a role and a typed content payload:

from mcp.types import PromptMessage, TextContent


@mcp.prompt()
def summarize_prompt(topic: str) -> list[PromptMessage]:
    """Walk the LLM through summarizing notes on TOPIC."""
    return [
        PromptMessage(role="user", content=TextContent(type="text", text=f"Please summarize my notes on {topic}.")),
        PromptMessage(role="assistant", content=TextContent(type="text", text="Of course — let me search.")),
        PromptMessage(role="user", content=TextContent(type="text", text=f"Use pkb.search with query={topic}, then synthesize.")),
    ]

Common mistake

Returning a plain string from a multi-message prompt — FastMCP only sees one message. Use list[PromptMessage] for multi-turn.

Solution to Task 3 — Pick the primitive

Feature	Primitive	Why
Get the current weather for a city	Tool	Takes input (city) and returns a time-varying result; can’t be discovered via `resources/list`.
Contents of the user’s most recent note	Resource	Read-only, discoverable; the LLM should see it in `resources/list` automatically.
Email template for follow-up after a meeting	Prompt	Reusable template the user invokes; produces text for the LLM to expand.
Delete all notes older than 30 days	Tool	Mutates state; the act of calling it is the point.

Looking ahead

Module 3 (Plan C) introduces sampling — server-initiated calls back to the LLM — and the full multi-tool PKB server.