Home▸Free interview questions▸MCP interview questions

free · no sign-up · no login

MCP interview questions, answered

20 questions with short answers, picked from the Q&A in the AI Engineer course. Tap a question to see the answer.

20 questionsBy Samir KumarUpdated

What is MCP (Model Context Protocol)?

MCP is an open standard, introduced by Anthropic, for connecting AI apps to tools and data through one common interface. An MCP server offers tools, resources and prompts. Any MCP client, such as an assistant or an agent, can find and use them, so you build an integration once instead of once per app.

No question matches that search. Try a shorter word.

20questions

MCP interview questions

The full set, with labs and mock interviews, is in the AI Engineer course.

See the AI Engineer course ▸

Tools, Function Calling & MCP

1

What is function calling?

The model doesn't run the tool itself. It asks for a tool with some inputs. Your code runs the tool and sends back the result. Then the model writes the answer.

2

What are structured outputs?

Structured outputs force the model to reply in a fixed JSON shape. Every field is there, with the right type, so your code can read it safely.

3

What is MCP?

MCP (Model Context Protocol) is an open standard for connecting AI apps to tools and data. Each system is wrapped once as a server, and any MCP app can use it.

4

What is the difference between JSON mode and schema-enforced outputs?

JSON mode only promises valid JSON, so fields can be missing or have the wrong type. Schema-enforced, or structured, outputs make the reply match your exact schema: every field, the right type, only allowed values. For anything your code reads, use the schema.

5

How do you validate tool arguments?

Check every argument in code before running the tool: types, required fields, allowed values, ranges, and that this user is allowed to do it. If something is wrong, return a clear error so the model can correct itself and retry.

6

What does the 'tool choice' setting control?

It decides whether the model may use tools: auto lets it decide, any forces it to call some tool, naming a tool forces that one, and none turns tools off. Forcing a tool is handy for structured extraction.

7

What are the host, client and server in MCP?

The host is the AI app, like a chat app or a code editor. Inside it, an MCP client connects to each MCP server. The server wraps one system, like GitHub or a database, and offers its tools and data.

8

How do you build an MCP server?

Pick an official SDK, like the Python or TypeScript one, define your tools with names, descriptions and input schemas, connect them to your system, and run the server over stdio or HTTP. Test it with an MCP client before sharing it.

9

What is the difference between an MCP resource and a tool?

A tool is an action the model decides to call, like create_ticket. A resource is data the app can read and give to the model, like a file or a customer record. Tools do things; resources provide information.

10

MCP vs A2A: competing or complementary?

They work together. MCP connects an agent to its tools and data. A2A lets agents talk to other agents and hand over tasks.

11

What is 'code execution with MCP'?

Instead of loading every MCP tool into the context and passing every result through the model, the agent writes code that calls the tools as functions, and only the final result comes back. It saves a lot of tokens when there are many tools or large data in between.

12

Structured outputs guarantee the shape. Do they guarantee the truth?

No. The JSON shape will be right, but the values can still be wrong. Always check amounts, dates and IDs against your real systems.

13

How do parallel tool calls work?

The model asks for several tool calls at once. Your code runs them at the same time. The model then combines all the results into one answer. It saves time.

14

What three things can an MCP server offer?

Tools: actions the model can call. Resources: data the app can read. Prompts: ready-made templates.

15

Should tools mirror API endpoints or whole tasks?

Usually whole tasks. One tool like refund_order that does the checks inside is much easier for the model than five low-level API calls it must chain in the right order. Fewer, task-shaped tools mean fewer mistakes.

16

How should a tool report errors?

With a short, clear message the model can act on: what went wrong and what to try, like 'Order not found. Order IDs look like ORD-1234.' Don't return raw stack traces, and never fail silently.

17

What is the most important security rule for tool calling?

Who the user is and what they can do must come from your login system, never from what the model says. Otherwise someone could trick the model into acting as another user.

18

How do you make tool calls safe for money-moving actions?

Give tools only the access they need, check limits on the server, ask a human to confirm risky steps, stop the same payment from running twice, and log everything.

19

Tools keep getting called at the wrong time. Fix?

Give tools clear names. Write when to use each one and when not to. Use fewer tools. Test with real questions until the right tool is picked.

20

Why is code execution such a powerful tool for agents?

With one tool the agent can do exact maths, clean and analyse data, use any library and draw charts, instead of needing a special tool for each job. It must run in a sandbox with no secrets and limits on time and network.

More free interview questions by topic

ready to practise?

Now test yourself

Play the free quiz, or try a mock interview where you answer out loud and get a score.

Your next AI interview is closer than you think

Interview-ready. From ₹999.

Four courses: FDE, AI Agent Developer, AI Engineer and Claude Certified Architect. Narrated lessons, labs where you write the code, and every interview answer written out. Pay once and start today.

Enrol now ▸ Take the free quiz