Skip to content
The AI Agent Factory

20.9 A text answer behind every screen, in every AI app

Approved by
Panaversity ·
7 min read

In everyday life. A museum label has a picture and a sentence. Visitors who cannot see the picture still get the sentence.

What you will learn. Why every tool keeps its text answer, and how one screen works in many AI apps.

Why it matters. Not every AI app shows screens, and not every person's plan allows them. The product must still work.

One request, start to finish. The builder runs list_open_invoices in three places at 14:00.

  1. In Claude, the screen appears, with the text answer behind it for the model.
  2. In ChatGPT, the same screen appears, because ChatGPT implements the same standard.1
  3. In a coding agent that shows no screens, the same call returns the same text: five invoices, one per line, and the total of $21,329.00. The task is still done.

On the left, in gold, list_open_invoices, which returns a text answer and a pointer to its page. Three arrows lead to three kinds of AI app. An AI app that shows screens shows the screen, with the text answer behind it for the model. Another AI app that follows the same standard shows the same screen. A coding agent with no screens shows the same text answer, five invoices and the total. Below, build and test the text answer first, and the task is done even where no screen appears.

Figure 20.10. One tool, one answer, shown three ways.

Why keep the text answer?

The standard says so: servers should give every tool with a page a text-only fallback, and a tool must return meaningful content even when it has a page.2 It is 17.3's rule again: a tool never answers with nothing. An AI app that does not support MCP Apps treats the tool as an ordinary tool. The text answer is also what the model reads, so it is how the model explains the screen.

So build and test the text answer first, and add the screen second.

How does one screen run in many AI apps?

Build to the standard. The official code library, @modelcontextprotocol/ext-apps, registers the page and the tool with registerAppTool() and registerAppResource(), and App.connect() in the page finds the AI app by itself.3 Many AI apps support the standard, among them Claude, ChatGPT, Gemini Enterprise and Goose.4 20.18 lists which ones, and how an administrator adds a connector in each.

Python has the same standard. The MCP Python code library includes MCP Apps support, mcp.server.apps, so the Python connector can serve the same page.5

Some AI apps add features of their own. ChatGPT's are reached through window.openai, and a page that depends on them works only in ChatGPT. OpenAI's own advice is to use the standard first and add its extras only for what the standard does not cover. Check that a feature is there before using it.1,6

Does the answer to "whose click is it?" change between AI apps?

No. Each AI app shows the page in its own way, and may ask the person different things. But the call reaches your connector with the person's token in every one of them, and the connector checks it the same way.

At work. In Lab 20, Task 2, the text answer is tested before the page exists. In Task 8, the same screen is opened in a second AI app, and the screen matrix's rows C1 to C3 run again there.

The operating rule: build the text answer first and the screen second, to the standard. Before you use one AI app's extras, check that they are there.

Check yourself

Question 1 / 2

0 answered

Why build and test the text answer before the screen?

  1. MCP Apps compatibility in ChatGPT, OpenAI Developers, accessed 15 November 2026.↑↑2
  2. MCP Apps specification, revision 2026-01-26, Model Context Protocol, accessed 15 November 2026.↑
  3. Get started with MCP Apps, Claude documentation, accessed 15 November 2026.↑
Owner
Panaversity