Skip to content
NTPU AI Learn

AI 201 · Module 2

MCP Architecture

Host, client, server and tool: who does what between your question and the 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

  • Name the parts of an MCP setup (host, client, server, tool) and say what each is responsible for
  • Trace a request from the user to authoritative data and back
  • Explain schema, arguments and result in plain words
  • Map the NTPU AI4X MCP example onto the general architecture

The Chain

10 min

The Chain in One Picture

Module 1 gave you a short chain: User, AI, MCP, trusted tool. Now we look at what sits inside it. Each step has one clear job, and keeping the jobs separate is what makes the setup easy to reason about and to secure.

The MCP architecture, from question to data
  1. User
  2. Host (the AI application)
  3. MCP client (inside the host)
  4. MCP server
  5. Tool
  6. Authoritative data / service

Info

Read the chain top to bottom for the request. The result travels back up the same path, and the human at the top verifies it.

Key takeaway: Every step in the chain has one job, and the human stays at both ends.

10 min

Host, Client, Server

  • Host: the AI application you use, such as a chat assistant or a code editor. It talks with you and decides when a tool could help
  • MCP client: the part of the host that speaks MCP. It connects to one server and asks what it offers
  • MCP server: a program that offers tools and data through MCP. It is the door to the outside capability
  • Tool: one specific capability the server exposes, such as "find research experts"

A host can hold several clients, one for each server it connects to. That is why one assistant can use campus tools and, say, a calendar tool in the same conversation.

Check yourself

Which part is the program that offers tools and data through MCP?

Which part is the program that offers tools and data through MCP?

Key takeaway: The host is the AI app, the client is its MCP connector, the server offers the tools.

Explore the Architecture

14 min

Explore the Architecture

Select each part of the diagram to read its definition, its responsibility, an example and a common misconception. The text list below the diagram says the same thing in order.

MCP architecture explorer

Select a node to see what it is, what it is responsible for, an example and a common misconception.

User

Definition
The person who has a goal and asks a question in plain language.
Responsibility
Defines the intent, decides what may be shared, and verifies the final result.
Example
A student asks which faculty work on agentic AI.
Common misconception
"Once the AI can use tools, I no longer need to check the answer." Tool use never removes the human's responsibility.
Text alternative: the same chain in order
  1. User: asks the question and verifies the result
  2. Host: the AI application that holds the model and the client
  3. MCP client: connects to the server and sends requests
  4. MCP server: exposes tools and runs them
  5. Tool: performs one defined capability
  6. Authoritative data / service: the source of the facts
Check yourself

A learner says, "MCP guarantees the answer is correct." Which correction is best?

A learner says, "MCP guarantees the answer is correct." Which correction is best?

Key takeaway: Each node has its own job, and the answer is only as good as the source and the review.

Schema, Arguments, Result

14 min

Tool, Schema, Arguments, Result

A tool is more than a name. To let an AI use it correctly, the server describes it in a fixed shape.

The four things that describe a tool call
TermPlain meaningExample
ToolA named capability with a descriptionfind_research_experts: find people by research topic and role
SchemaThe rules for what inputs the tool acceptstopic is required text; role is one of a few allowed values
ArgumentsThe actual values sent for one calltopic = "agentic AI", role = "faculty"
ResultWhat the tool returns, in a structured formA list of people with fields and a source

The schema works like a form with labeled boxes. The AI fills in the boxes (the arguments), and the server can reject a form that is filled in wrongly. This is what makes a tool call predictable instead of a free-text guess.

Check yourself

In the description of a tool, what does the schema define?

In the description of a tool, what does the schema define?

Key takeaway: Schema is the rulebook for inputs, arguments are the values for one call, result is the structured answer.

The NTPU Scenario

12 min

The NTPU AI4X MCP Scenario

Here is the general architecture mapped to a real campus example. NTPU AI4X MCP is the server, and find_research_experts is one of its tools.

NTPU scenario
  1. AI application (the host)
  2. MCP client
  3. NTPU AI4X MCP (the server)
  4. find_research_experts (the tool)
  5. AI4X authoritative data

Info

Learn teaches the ideas. NTPU AI4X MCP is the service that runs, and its own documentation is the place for connection details, so this course does not repeat them.
Check yourself

In the NTPU scenario, which item is the tool?

In the NTPU scenario, which item is the tool?

Key takeaway: The general pattern maps directly: AI app, client, NTPU AI4X MCP server, find_research_experts tool, AI4X data.

What next?