Skip to content
NTPU AI Learn

AI 201 · Module 1

Why MCP?

AI is already capable. Why does it still need a standard way to reach tools and trusted data?

Start Section 1
  • Duration60 min
  • LevelIntermediate
  • PrerequisiteAI 101 or equivalent basics recommended
0 / 5 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Explain why a model on its own cannot reliably answer questions about current, internal or system-state information
  • Compare four ways of answering a campus question: model only, web search, RAG and an MCP tool
  • Describe the mental model User → AI → MCP → Trusted tool / data, and where the human stays responsible
  • State in one sentence what MCP does and does not do

The Problem

10 min

A Simple Campus Question

Start with a question any student might type into an AI assistant. It sounds easy, but it hides a real problem.

Campus scenario

Which AI courses are offered in 115-1?

To answer well, the AI needs data that is current (this semester, not last year), structured (course names, times, instructors) and trusted (the official course system, not a guess).

  • Current: the answer changes every semester
  • Structured: a list with fields, not a paragraph of prose
  • Trusted: it should come from the authoritative source
Check yourself

Which three qualities does the campus question "Which AI courses are offered in 115-1?" require of its data?

Which three qualities does the campus question "Which AI courses are offered in 115-1?" require of its data?

Key takeaway: A good answer to a factual campus question needs current, structured and trusted data.

12 min

What a Model Cannot Do on Its Own

A language model is good at language. It learned from data collected up to some point in the past, and it only sees what you put into the conversation. That creates five limits.

  • Its knowledge can be out of date
  • It does not naturally know private or campus-internal data
  • It should not guess the current state of a system
  • It should not be handed sensitive information without control
  • A prompt alone cannot safely carry out an action in an external system

Warning

When a model does not know, it may still produce a fluent, confident answer. For facts about courses, people or services, fluency is not evidence.

This is the same lesson as AI 101: a person defines the task and reviews the result. What is new here is that the AI needs a controlled way to reach the right source in the first place.

Check yourself

A model is asked for this semester's course list and answers fluently without any source. What is the safest reading?

A model is asked for this semester's course list and answers fluently without any source. What is the safest reading?

Key takeaway: A model alone can be out of date, blind to internal data and confidently wrong.

Four Ways to Answer

14 min

Model, Web Search, RAG or MCP Tool?

Four approaches can each produce an answer to "Which AI courses are offered in 115-1?". They differ in where the facts come from. Compare them before you decide.

Four ways to answer the 115-1 course question (illustrative comparison)
ApproachWhere facts come fromStrengthWatch out for
Model onlyTraining data from the pastFast, no setupMay be out of date; no source
Web searchPublic web pages found at question timeRecent and broadPages may be unofficial, stale or unstructured
RAGDocuments someone indexed in advanceCan include internal documentsOnly as fresh as the last indexing
MCP toolA tool that queries the authoritative systemCurrent, structured, with provenanceNeeds a tool that exists and is authorized

Info

No approach is always best. For current, structured and trusted data, an MCP tool that queries the authoritative system is usually the strongest fit. Module 4 compares MCP, APIs and RAG in depth.

Try it yourself: for the 115-1 question, which of the four would you trust for the final list, and why?

Check yourself

Which approach is best suited to fetch current, structured and trusted data such as this semester's official course list?

Which approach is best suited to fetch current, structured and trusted data such as this semester's official course list?

Key takeaway: Choose the approach by where the facts come from: for current, structured, trusted data, an authoritative tool.

The Mental Model

12 min

The Mental Model

MCP (Model Context Protocol) is a standard way for an AI application to discover, understand and use external tools. Hold this simple chain in mind for the rest of the course.

User → AI → MCP → Trusted tool / data
  1. User: asks a question in plain language
  2. AI: works out what is needed and which capability could help
  3. MCP: the standard connection through which the AI finds and calls a tool
  4. Trusted tool / data: the authoritative source that returns the facts

Around that chain sits the human. Human defines intent → AI selects a capability → MCP invokes a trusted tool → Human verifies the result. Using a tool does not remove human responsibility.

Info

MCP is not an "API replacement". An API defines how a system exposes its capability. MCP is a standard interface that lets AI applications discover, understand and use tools, and those tools often sit on top of existing APIs.
Check yourself

Which statement about MCP is accurate?

Which statement about MCP is accurate?

Key takeaway: Human → AI → Tool → Human: MCP is the standard link between the AI and the tool, and a human still verifies.

Recap

12 min

Recap: What MCP Does and Does Not Do

Done

MCP does not make the model smarter. It gives AI a standard way to discover and use external capabilities.

What MCP does

  • Gives AI applications one standard way to find available tools
  • Describes what each tool needs and returns
  • Lets the AI call a trusted tool instead of guessing
  • Makes it possible to carry the source of a result along with it

What MCP does not do

  • It does not make the model reason better
  • It does not replace existing APIs
  • It does not guarantee the data is correct or complete
  • It does not remove the need for human review

Next, Module 2 opens the chain up and names each part: host, client, server and tool.

Check yourself

Which sentence best completes: "MCP does not make the model smarter, it..."?

Which sentence best completes: "MCP does not make the model smarter, it..."?

Key takeaway: MCP does not make the model smarter; it gives AI a standard way to discover and use external capabilities.

What next?