Let an AI Choose a Tool

This module is a guided demonstration. You do not need Claude Code to complete the rest of the tutorial. Read the recorded exchange below; nothing here is required setup.

Until now, you chose every capability

In Inspector, you personally selected search_resources, entered its input, clicked Execute Tool, opened the resource, and retrieved the prompt.

The Practice Resource Desk did not decide what you wanted. Inspector did not decide what tool would help.

Now we introduce an AI model between your goal and the MCP tools. Instead of saying “run search_resources with research,” you can describe what you want to accomplish.

The tool has not changed. Who selects it has changed.

Three words to learn

Watch the handoff

For this demonstration we used Stanford Claude Code, because Module 7 established it is the Stanford-provided surface that can connect directly to the local practice server. It was run in an empty folder with nothing else in it, so the only place the answer could come from was the MCP connection.

Both tools were available to the model — the read-only search and the note-writing tool. Only the search tool was permitted to run.

Already use Stanford Claude Code? You can reproduce this demonstration with your downloaded Practice Resource Desk.

This was the entire human instruction:

What resources from the Practice Resource Desk could help me with research?
Do not create or modify any files.

Notice what the instruction does not contain. It never names search_resources. It never mentions a query field. It does not say which of the five fictional resources to look for.

Claude then made three tool calls of its own choosing:

search_resources  query: "research"
  -> 2 of 5 practice resources matched "research":
     Research Planning Workshop
     Source Evaluation Checklist

search_resources  query: "citation"
  -> 1 of 5 practice resources matched "citation":
     Citation Practice Guide

search_resources  query: "writing"
  -> 1 of 5 practice resources matched "writing":
     Citation Practice Guide

And this is part of the answer it wrote back:

Three of the desk's five resources are relevant to research work:

Research Planning Workshop (Workshop, any student) - a 90-minute session on
turning a broad question into an actual research plan...

Source Evaluation Checklist (One-page checklist, anyone) - criteria for judging
whether a source is current, relevant, and appropriate to cite...

Citation Practice Guide (Self-paced guide, first-year students) - a walkthrough
of citation formats, for the write-up stage.

Together they map onto the three phases: plan the question, vet the sources,
cite them correctly.

One note: the desk describes its entire collection as fictional...

In this observed run, Claude selected search_resources. Models can make different choices in another run, especially when several tools could plausibly help.

What changed?

1. You supplied the goal and constraint. Find useful Practice Resource Desk material about research, without changing files.
2. The model selected a tool. In the observed run, Claude chose search_resources.
3. The model supplied tool input. First research, then citation, then writing — three searches it decided to run.
4. The host invoked the MCP tool and received its result. The same text you would have seen in Inspector’s Results panel.
5. The model used that result to write its answer to you.

Step five is worth separating out. The Practice Resource Desk never wrote a word of that reply. It returned the same plain text it returned in Module 2; the model read that text and composed prose around it. The tool supplied facts, not sentences.

The schema became useful again

In Module 5, the schema helped Inspector build a form for a human.

Here, the same tool description and expected input structure help an AI host and model understand how the capability can be called. The host made the tool’s definition available as part of the model’s tool-use environment.

You no longer typed into the query field directly — but the tool still required valid input. Something had to produce { query: "research" } in exactly the shape the tool accepts, or it would have been rejected the way your 42 was in Module 5.

Your constraint mattered

The instruction ended with five words: Do not create or modify any files.

That was there for a reason. Since Module 6, this server exposes both a read-only search tool and a tool that writes a local file. The task did not require a write, so the human supplied a boundary before delegating.

A good delegation describes not only the goal, but also important limits.

Same tool, different decision point

Module 2 — Inspector

You chose search_resources. You supplied citation. You clicked Execute Tool.

Module 8 — AI host

You supplied a goal and a no-write constraint. The model selected search_resources. The host handled the tool interaction.

MCP did not become “more autonomous.” The host and model changed how the capability was selected.

What to notice

You stopped specifying the mechanism and started specifying the goal.

• In Inspector, you chose the capability yourself.
• In this demonstration, the model chose a capability from those the host made available.
• But the server remained the same.

Delegation moves a decision. It does not remove responsibility for understanding the tools, constraints, and results.

Look back at Claude’s reply once more. It reported that the collection is fictional. That was correct — but you only know it was correct because you read those same records yourself in Module 2.

Checkpoint

You are ready for Module 9 if you can:

• explain delegation in ordinary language;
• recognize a tool call;
• distinguish a goal from a tool-specific instruction;
• explain what the model selected and what the host executed;
• explain why a natural-language constraint is useful but not a security guarantee; and
• identify what still needs human review after a tool-assisted answer.