AI Learning Hub › Learn AI › Self-Paced Courses › MCP Without the Mystery › Use Your First Tool

Module 2 ยท Beginner

Use Your First Tool

Inspect what a tool says it does, identify the input it expects, then call a read-only tool by hand and examine what comes back.

By the end

  • Explain what an MCP tool is in ordinary language
  • Inspect what a tool says it does before invoking it
  • Identify the input a tool expects
  • Call a read-only tool manually through MCP Inspector
  • Examine the output before an AI model is involved

A tool does something

In Module 1, you saw that the Practice Resource Desk exposes a tool named search_resources.

You did not run it.

Before invoking a tool, first ask:

In this module, the answer to the last question is intentionally simple: search_resources only searches the fictional records inside the practice server.

Three words to learn

Tool

A capability an MCP server exposes that can be invoked to perform a defined piece of work.

Input

The information a tool expects before it can run.

Output

The result the tool returns after it runs.

Step 1 — Read before you run

Start the Inspector exactly as you did in Module 1. From inside the mcp-practice folder:

Command

npm run inspect

Open the printed address, connect to practice-resource-desk with the toggle switch, then click the Tools tab. The left column lists the server’s tools; click Search resources.

Stop there. Do not run it yet. Read what the Inspector is telling you:

The Tools tab of MCP Inspector with Search resources selected. The panel shows the title Search resources, a green READ-ONLY badge, the tool's description, an Edit as JSON toggle, a field labeled query with a red asterisk and the guidance text A word or phrase to look for, the word citation typed into that field, and a blue Execute Tool button.
Everything the tool says about itself, before it runs. The field and the button are the whole interface. Open full-size image: the tool's input form (opens in a new tab)

Find each of these on the screen

  • Its readable title: Search resources
  • Its technical name: search_resources
  • Its description, which says it searches the fictional collection by keyword and never changes anything
  • The one input field it offers, labeled query
  • The red asterisk beside query, marking it as required

The asterisk is not the only sign that the input is required. With the field empty, the Execute Tool button is grayed out and cannot be clicked. The tool has told the Inspector that it cannot run without a query, and the Inspector is enforcing that before anything is sent.

You may see additional technical information about the tool, including an Edit as JSON switch. You do not need it for this exercise.

The form did not come from this tutorial

Notice what just happened. Nobody told the Inspector that this tool takes a single piece of text called query. The Inspector asked the server, and the server described the information it expects.

MCP tools describe the information they expect. The Inspector can use that description to present an interface for supplying the input.

You did not need to read the server’s code or construct a technical request yourself. That is the part worth remembering: a tool is self-describing, so anything connecting to it — an inspection tool today, an AI application later — can find out how to use it by asking.

Step 2 — Call the tool yourself

  1. Type citation into the query field.
  2. Click Execute Tool.
  3. Wait a moment for the response.
  4. Read what the server sent back.

The panel switches from the form to a section headed Results:

MCP Inspector showing the Results panel after calling search_resources with citation. The returned text reads: 1 of 5 practice resources matched citation, followed by Citation Practice Guide with its format Self-paced guide, audience First-year students, and a one-line fictional summary, ending with the line All of these are fictional.
The output, exactly as the server sent it. One fictional match: the Citation Practice Guide. Open full-size image: the first tool call and its result (opens in a new tab)

The form disappears when the result arrives. That is expected, not a glitch. To run another search, click the × beside Results and the form comes back with your input still there.

What happened?

  1. You supplied the input: citation.
  2. Inspector invoked the MCP tool.
  3. The Practice Resource Desk searched its fictional records.
  4. The server returned the result.
  5. Inspector displayed the output.

No AI model chose the tool, rewrote the request, or summarized the answer. You called the capability directly.

What you are reading in the Results panel is the server’s own words, unedited. Nothing stood between the server and you.

Step 3 — A successful call can return “nothing found”

Close the result with the ×, replace citation with spaceship, and click Execute Tool again.

The server replies:

What the server returns

No practice resources matched "spaceship". The fictional collection holds 5 items; try a broader word such as "research".

Look at how that arrives. It appears under the same Results heading, in the same panel, styled the same way as the match you got for citation. The Inspector does not mark it as a failure, because it was not one. The tool ran, searched, and correctly reported that nothing matched.

That distinction is worth naming:

Successful tool call
The tool ran correctly.
Useful result
The returned information answered your need.

Those are related, but they are not the same thing. A tool can work perfectly and still tell you nothing you can use — and if an AI application had made this call on your behalf, that is a difference you would want it to report to you plainly rather than paper over.

Tool does not mean “write”

A tool is a capability, not a synonym for an action that changes data.

This practice tool only searches information. Other MCP tools may create, edit, send, delete, publish, or trigger something elsewhere.

Later in the tutorial, you will learn to distinguish those consequences before allowing a tool to run.

What to notice

You can understand a tool before letting an AI decide when to use it.

You inspected what the tool claimed to do, the input it expected, and the output it returned.

You also saw that tool success is not the same as answer quality. A search can run perfectly and still find nothing useful.

Later, an AI application may decide to invoke a tool for you. The underlying capability is the same one you just called yourself.

Checkpoint

You are ready for Module 3 if you can: