Course B — MCP: AI apps as easy as 1, 2, 3
Unit 10: Resources, prompts, and the LLM
Self-paced · about 60 minutes · Course B, chapter 2 of 3
Learning objectives
- Serve a resource, and say why read-only context is not a tool.
- Serve a prompt, and keep its name and its title straight.
- Write the five steps of a tool call, including the one that gets left out.
- Answer an unanswerable request with a question instead of a number.
Lessons
| Lesson | What you leave with |
|---|---|
| 1. Resources in MCP servers | @mcp.resource(uri), list_resources, read_resource, and one thing per line |
| 2. Prompts in MCP servers | @mcp.prompt(title=...), and the name the client asks for |
| 3. MCP and LLMs: tools | The five-step workflow, and the trace that shows step four missing |
| 4. MCP and LLMs: prompts and resources | The ambiguous request, and clarification as an answer |
Checking your answers
Every exercise in notebook.ipynb has the same shape: context, numbered
instructions, a starter that runs as shipped and gives a wrong or incomplete
answer (marked <------ EDIT THIS LINE), the expected output, and a
check(...) cell that prints ✅ or ❌ with the fix named. The last cell,
review("w10"), is the scorecard.
The checks judge behaviour. The resource is compared as the client received it, not as your source reads. The prompt check refuses a name equal to its title, which is the deck's own trap. The loop check reads the trace and then the answer, because a five-step trace with the wrong answer is the case worth catching. The last check refuses any clock time at all: a converted time for "what time is it in Canada?" is a guess with a decimal point on it.
Working with a coding assistant: give it the exercise's context and
instructions, let it edit the marked lines, then run the check cell yourself.
You read the verdict, not the assistant. solutions/notebook.ipynb is the
reference; open it after the check, not before.
Offline, and honest about it
The deck calls Claude through the Anthropic Messages API. That needs a key, and a model's choice is not reproducible, so this unit does two things instead.
The seam is real: FakeLLM in src/bootcamp_agent/llm.py takes the same
complete(system, user) call every provider adapter in this repo takes. Set
BOOTCAMP_PROVIDER (see SETUP.md) and get_client(...) returns Anthropic,
OpenAI, OpenRouter or a local Ollama instead. The loop does not change.
The decision is a router you can read: two places named means convert, a country with several zones means ask. That is the one thing a model would do that a fake cannot, so it is written out as rules rather than mocked. When you swap a real model in, the router is what it replaces.
The server, the client, the protocol and the conversion are the real thing. The
conversion reads unit-09-mcp-first-server/fixtures/timezones.json, twelve
zones with their January offset, which is why every example here is a January
date. Seven of those zones move an hour in summer and the table does not know it.
What unit 9 already covered
The FastMCP to MCPServer rename, and why this repository teaches the second
name: see unit 9's README and the last slide of its deck. This unit uses
MCPServer throughout and does not repeat the explanation.
One thing did change between the SDK versions in this chapter's material: the
deck lists the locations resource as file://locations.txt/, with a trailing
slash. Version 2 lists it as file://locations.txt. Read the URI off
list_resources rather than typing it from the slide.
Slides
slides.md is a Marp deck. Read it as Markdown, or render it:
npx @marp-team/marp-cli slides.md -o slides.pdf
Previous: Your first MCP server · Next: Databases, APIs, and third-party servers