Skip to content
The AI Agent Factory

Chapter 20. Building Interactive Connectors: MCP Apps and Extensions

Approved by
Panaversity ·
29 min read

The point

Software is moving from web and mobile apps to AI apps. For twenty years, a company that wanted to reach its customers built a website, then a phone app. Now many people start their work in an AI app such as Claude or ChatGPT. They say what they need in plain words, and the AI does the steps. So a new kind of app is growing: one that lives inside the AI app, has its own screens, and works together with the model.

Anthropic and OpenAI agreed on one standard for these apps: MCP Apps. Anthropic, OpenAI and the team behind the open-source MCP-UI project proposed it together in November 2025.1 On January 26, 2026, it became MCP's first official extension: an optional add-on that both the AI app and the connector must support.2 You build a screen once, and it runs in every AI app that supports the standard, among them Claude, ChatGPT, GitHub Copilot in VS Code, Microsoft 365 Copilot and Goose.3 Two more official extensions sit beside it. One handles sign-in, so a whole company, or a program with no person present, can connect safely.4 The other lets a connector share its skills with the AI app.5 This chapter teaches all three.

Can you build something like Claude Docs or Claude Slides?

Many readers ask this first. Mostly yes. An MCP App has its own screens, buttons and forms, like a web app. It is also joined to the model: the model can open it, read what is on it and make changes in it while you watch. Since November 2026, a connector's screen in Claude can open beside the chat or full screen as a full editor.6 Claude calls these apps interactive connectors, and its directory lists them in their own section.7 Well-known software companies already ship them, among them Notion, Atlassian's Confluence, Box, Adobe, HubSpot and Canva.6

The difference from Claude's own Docs is where the work lives: on your server, so storing, sharing and protecting it is your job. Once you know the parts, 20.3 compares the two point by point. The optional Lab 20b builds such an app.

Where this chapter fits

Chapters 16 to 19 built Brightline's product one layer at a time. Chapter 16 wrote Ana's method as skills in a plugin. Chapter 17 built the connector those skills call. In Chapter 18, the connector learned who is calling, and for which company. In Chapter 19, the database learned to refuse another company's rows. All four chapters assumed one thing: every call starts with the model. This chapter adds calls that no model chose, and it ends the arc. The connector gets its own screens, and Chapter 16's skills travel with it.

This chapter teaches the extensions a builder who sells a connector needs. The first, and the biggest, is MCP Apps: a connector with its own screens. When a tool runs, the AI app shows a small web page inside the chat: a table, a form, a chart, buttons. The person reads it and clicks. Nobody has to write a good request to get the next page of invoices.

From Skills to Income named four things a reader can build and sell: a skill, a plugin, a connector and an MCP App. This chapter builds the fourth on top of the third. Brightline's AP connector from Chapter 17 gets a screen for open invoices, for Brightline's customers who would rather click than type. Then it learns three more things through extensions: to share its skills with any AI app that connects, to let a customer company decide which employees may connect, and to let a program with no person present sign in as itself.

Two ways to ask for the same invoices, side by side. Left, as text in the chat. Someone types show my open invoices, and the reply lists Cedar Ridge's 16 open invoices, $21,329.00 in all, then the first rows as lines of text, INV-6201 V001 $4,120.75, INV-6210 V001 $3,250.00 and INV-6204 V004 $2,688.00, and says to type next five to see more. A note says each next step is a new request the person has to type. Right, as a screen in the chat. The same rows in a table with columns for invoice, vendor and amount, a Vendor details button on each row, and a Next page button. A note says one click gets the next page. Below, same tool, same data, same checks. The screen is another way to ask.

Figure 20.1. Same tool, same data, same checks. The screen is another way to ask.

But a button is not a picture. In Chapters 17 to 19, every call to the connector began with the model. The model chose a tool, and the AI app sent the call for it. A button on the screen calls a tool too, and the model chooses nothing. So the chapter asks one question before it builds anything:

When Luis presses a button on the screen, who is that call acting as? And could the page ever show him one of Brightline's invoices?

The answer, in one line: a screen is a second caller, and it calls with nothing of its own. The page asks the AI app. The AI app sends the call with the sign-in token the person gave it. The connector checks that call exactly as it checks the model's: the person, the company, the permission. Hiding a tool or a button protects nothing.

The same question follows every extension in the chapter. A skill from the connector: whose instructions is the model following, and who approved them? A company's identity provider: who decided that this employee may connect? A report job that runs at night: who is acting when no person is there? Each time, the answer is something the connector can check on every call, never something it takes on trust.

After this chapter you can:

  • Say what an extension is, what its stage promises, and when to fall back or refuse.
  • Say when a screen helps, and when a plain text answer is better.
  • Trace one button press from the page, through the AI app, to the connector and back, and say whose call it makes.
  • Use tool visibility correctly: to decide which tools the model is offered, never who may call them.
  • Keep the page in its box, show the data it receives as text, and keep the model up to date with the screen.
  • Build one screen for many AI apps, with a useful text answer behind every screen, and let the AI shape screens only from approved parts.
  • Share a skill from the connector, with a manifest the AI app checks and an approval that a changed file cancels.
  • Let a company's identity provider decide who connects, and give a job with no person its own identity.

Who does what. This is a co-build. Ana Morales, Brightline's AP analyst, decides which work deserves a screen, what it must show, and which skills Brightline shares. You, the builder, direct the coding agents and decide how the screen and the skills reach the connector. Managers and experts can read this overview and Concepts 20.1, 20.2, 20.4, 20.5, 20.13 and 20.14.

Words this chapter uses

The words of Chapters 17 to 19 keep their meaning: AI app, connector, tool, sign-in token, scope, role, tenant, handle. In the stories, "the builder" is you.

WordMeaning
MCP AppA connector's own screen: a web page that the connector offers, which the AI app shows inside the chat. MCP calls the standard behind it the MCP Apps extension
ExtensionAn optional add-on to MCP. Both sides, the AI app and the connector, must say they support it before it is used
PageThe web page an MCP App shows. Its address starts with ui://
ScreenWhat the person sees: the page, drawn inside the chat
SandboxA sealed box the AI app draws the page in. The page cannot reach anything outside it except the AI app
Message channelThe browser's built-in way for two separate pages to send each other messages. The page and the AI app talk only through it
VisibilityA setting on each tool: offered to the model, callable by the page, or both
Text answerThe plain text every tool still returns, for the model and for AI apps that cannot show screens
ManifestThe list of every file in a skill, each with its size and a fingerprint that changes if one letter changes
Identity providerThe system a company uses to sign employees in to work apps, such as Okta or Microsoft Entra ID
ID-JAGA pass from a company's identity provider that an AI app trades for an access token. Short for Identity Assertion JWT Authorization Grant
Client credentialsA program's own sign-in, with no person present
TaskA handle a connector returns for long work, which the caller checks later

What you will build

Where you are. At the end of Chapter 19, Brightline's connector checks a sign-in token on every call and takes the company from it. It reads each company's rows from a database that refuses the others. Two test suites prove it on every build: the sign-in matrix (18.13) and the data matrix (19.12). Cedar Ridge Supply, Brightline's second customer, has 16 open invoices, $21,329.00 in all. Luis Moreno, its AP lead, reads them by asking in a chat. Three things are missing:

  • Luis has to type. "Show the next five", "now the ones from V102", "what are V102's terms?" Each step is a new request.
  • Nothing on the connector is drawn. Every answer is text, even a list of 16 invoices.
  • Nobody has asked who a click is. A screen would add calls the model never chose.

What you build instead. As in Chapters 17 to 19, coding agents write every file, and you specify, predict and approve.

  1. A page for list_open_invoices. One ui:// page, a table of five rows a page, with Next page and Vendor details buttons. The tool's text answer stays exactly as it was, and is checked first.
  2. One tool the page may call and the model is not offered. It shows the table's next page from the sealed handle.
  3. A screen matrix, the third after Chapter 18's sign-in matrix and Chapter 19's data matrix. Every row is one click by one person and the answer it must get, and each row runs as a test. The rows cover Luis's next page, Marco's refused vendor details, a request for Brightline's invoice from Cedar Ridge's screen, and Priya, who belongs to both companies.
  4. The same screen in Python. As in Chapter 17, Claude Code builds the TypeScript connector's page, Codex builds the Python connector's, and OpenCode writes the suite that proves both answer the same.
  5. Two skills served from the connector: invoice-triage and vendor-terms, with their manifests.
  6. If you have time, a test identity provider for Cedar Ridge, the month-end report job with its own identity, and an app tool that lets Claude jump to the overdue invoices on the screen.
  7. Optional, a second lab: a full interactive app. Lab 20b builds Cedar Ridge's review workspace, an editor that Luis and Claude use together, with notes saved on the server.

Why it matters

This story is invented, but each failure in it is common. A small software company in Tucson, Arizona, sold a connector for timesheets to about thirty firms. It added a screen with a Team hours button. The button called a tool that only the page was offered, so the developer reasoned that nobody else could call it. That tool skipped the scope check and read the firm from a field the page sent.

A contractor who worked for two firms called the same tool from a small script, with her own token, and put the other firm in that field. She saw the other firm's timesheets. She reported it. Others might not have.

Three design choices made it possible. The tool believed a firm the page sent. A tool hidden from the model was treated as private. And no test ever pressed a button as the wrong person. This chapter fixes all three. The company comes only from the token (20.4). Every tool gets the same checks, whoever is offered it (20.5). And the screen matrix presses every button as the wrong person, on every build (Lab 20).

Three failures and their fixes, in three rows. First, the tool believed the firm the page sent, fixed by taking the company only from the token, in the connector. Second, a tool hidden from the model was treated as private, fixed by giving every tool the same checks, whoever is offered it, in the connector. Third, no test ever pressed a button as the wrong person, fixed by a screen matrix that presses every button as the wrong person, in the test suite. Below, a button is a caller, so check it like one.

Figure 20.2. Three failures, three fixes. Two live in the connector, and one in the tests.

A screen makes a connector easier to use, so more people use it, and some of them will never type a request. That is why it sells. But every button is a new door into the connector. If you treat a click as trusted because "the page is ours", you have built a way around Chapter 18. If you treat it as an ordinary call, checked like every other, the screen adds nothing a person could not already do by asking.

How the plugin grows

ChapterWhat it adds to the plugin
16Skills, a command and a ready-made connector
17Brightline's own connector: four read-only tools, written by coding agents, and put online with a shared key
18Real sign-in. Each person signs in and picks a company, and the connector knows who may act for which company
19The business data in a database, each company's rows kept apart by the database itself, each company's own rules, and policy search by meaning with citations
20Its own screens inside the chat, a read-only review of open invoices, and its skills served over MCP. Every button press is checked as the person signed in
22An agent that runs on a server inside a web page Brightline's staff already use, holding each person's tokens for them

The plugin's files do not change in this chapter. The connector's address stays the same, and its four tools keep their names and their text answers. list_open_invoices gains a pointer to its page, and the connector gains one tool that only the page is offered, next_invoice_page. The sign-in server and the database do not change.

In this chapter

Your path through this chapter.

PathReadThen do
Core, for everyoneThis overview, 20.1 to 20.9, 20.12, 20.13 and 20.16Lab 20, Tasks 0 to 12
Going further20.10 and 20.11 on screens the AI shapes, 20.14 and 20.15 on sign-in, 20.17 and 20.18Lab 20, Tasks 13 to 15, Lab 20b, and the field guide when a task or an error sends you there

Every lesson builds on 18.6, 18.7, 18.11 and 19.9. If those are not fresh, read them again first.

Part A. Extensions, a screen, and whose call it makes. Then Sitting 1.

  • 20.1 What an MCP extension is
  • 20.2 When a screen beats text
  • 20.3 How an MCP App works
  • 20.4 Who a click acts as
  • 20.5 Hidden is not protected

Part B. The page in its box. Then Sitting 2.

  • 20.6 The sandbox
  • 20.7 Untrusted text on the screen
  • 20.8 What the model knows about the screen
  • 20.9 A text answer behind every screen, in every AI app

Part C. Screens the AI shapes.

  • 20.10 Screens the AI designs
  • 20.11 A2UI: screens described as data

Part D. Skills through an extension, and two ways to sign in. Then Sitting 3.

  • 20.12 Skills that travel with the server
  • 20.13 Approving a skill from a server
  • 20.14 Employees through their company (going further)
  • 20.15 Machines with no person (going further)
  • 20.16 How the pieces fit

Going further.

  • 20.17 Charging and the screen
  • 20.18 How today's tools realize it

Then the field guide (F0 to F7), Lab 20, the optional Lab 20b, and Check yourself.

Three sittings. The lab takes about 6 hours 10 minutes in three sittings: a screen and whose call it makes, the page in its box, and skills over MCP. Its three optional tasks add about 1 hour 45 minutes. Lab 20b, the full interactive app, takes about 5 hours 50 minutes more.

  1. MCP Apps: Extending servers with interactive user interfaces, Model Context Protocol blog, 21 November 2025.↑
  2. MCP Apps: the first official MCP extension, Model Context Protocol blog, 26 January 2026.↑
  3. MCP Apps, Model Context Protocol, accessed 16 November 2026.↑
Owner
Panaversity