Skip to content
The AI Agent Factory

Field guide: interactive connectors, product by product

Approved by
Panaversity ·
14 min read

Optional. Read a section when a task or an error sends you here. Every fact on this page was checked on November 15, 2026. Check it again before you build, because these products change often.

F0. The versions this chapter was checked against

PartVersion or stageNote
MCP AppsRevision 2026-01-26, StableThe standard behind every screen in this chapter1
Skills over MCPBetaIts proposal, SEP-2640, is final2
Enterprise-Managed AuthorizationStableCompany sign-in, 20.142
OAuth Client CredentialsStableSign-in for a job, 20.152
TasksBeta, planned for the coreLong work, 20.152
@modelcontextprotocol/ext-apps1.7 or later for app tools. 2.0 is outSee below

Which version of the code library? Version 2.0 of @modelcontextprotocol/ext-apps works with version 2 of the MCP code libraries for TypeScript. Its messages are the same as in 1.x, so a page built with 2.0 works with an AI app built for 1.x.3 Claude's own quickstart still loads 1.7.5.4 For Lab 20, keep the version that matches the MCP code library your connector already uses from Lab 17, and use 1.7 or later if you want app tools.5 Write the version you chose in docs/connector-spec.md.

The skills for your coding agent. In Claude Code:

/plugin marketplace add modelcontextprotocol/ext-apps
/plugin install mcp-apps@modelcontextprotocol-ext-apps

Other coding agents can install the same skills with npx skills add modelcontextprotocol/ext-apps, or by copying them from the project's repository.6

F1. Claude in detail

  • Who can use screens. Claude shows screens for everyone on the web, in the desktop app and in the iOS and Android apps.7 A connector must be added on the web or the desktop app before it appears on a phone.
  • The first time. Before a page first appears, Claude asks the person: Allow, or Always allow for a server they trust.8
  • The sandbox address. Claude serves each page from its own address. Set the page's ui.domain to the first 32 hexadecimal characters of the SHA-256 of your server's address, followed by .claudemcpcontent.com.8
  • Display modes. A page starts inline, in the conversation. It can ask for full screen, and Claude then opens it in a side panel beside the chat, or full screen where there is no side panel. Claude grants the request only after a recent click or key press in the page. Lay the page out again whenever its size changes.9
  • Full apps. Since November 10, 2026, a person can open a supported tool beside the chat or full screen, open it without a prompt, or pin it to the sidebar. Claude can change things in the open tool, and the person approves those changes as for any connector action. It is rolling out on the web and the desktop app, on all plans.10
  • A page's own tools. A page registers its own tools with app.registerTool(), which needs version 1.7.0 or later. The page must check every argument itself, because nothing checks it for the page.5
  • Embedded sites. A page can frame a website of yours. Claude must allow the site's address for your connector first. For a connector in Claude's directory, its listing must allow that address, and a site that loaded while you tested may still be blocked under a Community listing. Never list a wildcard sandbox address, because then other connectors' pages could frame your site.11
  • The directory. A connector with screens needs screenshots when you submit it to Claude's directory.12

F2. ChatGPT in detail

ChatGPT implements the MCP Apps standard. A page calls tools with the shared tools/call. ChatGPT keeps window.openai for older apps and for its own extras, such as file uploads and its checkout. Check that each extra is there before you use it, so the page still works in other AI apps.13 On ChatGPT's Business, Enterprise and Edu plans, an administrator adds a custom MCP server and publishes it to the workspace.14

F3. Gemini Enterprise in detail

Gemini Enterprise shows MCP Apps in a sandboxed frame in the chat, and can open them full screen or in a side panel. An administrator registers the connector as a custom MCP server, and its tools start turned off until the administrator turns them on.15

F4. Open source in detail

  • Goose is listed by the MCP project among the AI apps that show MCP Apps.16 It is free, and Lab 20's Task 8 can use it as the second AI app.
  • The basic host is a small AI app for testing, in the project's repository. Clone the repository, run npm install, then npm start, and open http://localhost:8080/.3 It is a test tool, not a product. The project says the repository has no supported host beyond this example.
  • OpenCode shows no screens. It uses the text answer, which is why Lab 20 uses it for the tests.

Which AI apps support which extension changes often. Check the MCP project's support table before you promise a customer anything.17

F5. Testing a page

  1. Text first. Test the tool's text answer with no page at all (Lab 20, Task 2).
  2. The basic host. Open the page in the basic host before any AI app (F4). It shows the messages between the page and the host.
  3. Developer tools. In the desktop app or the browser, open the page's developer tools. The console shows blocked addresses and the page's own logs. Claude's troubleshooting page also covers developer tools on iOS.18
  4. A custom connector. Add the staging connector to Claude as a custom connector, and test as each test person.
  5. The matrices. Run the sign-in, data and screen matrices on every pull request.

Claude's troubleshooting page covers the common problems: a page that does not appear, a tool result that is too large, requests that fail only on iOS, and ui.domain errors.18

F6. The interactive-connector checklist

Before you sell a connector with screens, check each line:

  1. Every tool returns a useful text answer, tested without the page.
  2. The page holds no key, no token and no password.
  3. The page bundles its code and declares no outside address, or only the few it cannot avoid.
  4. Every tool a page may call is checked like any other call: the token, the scope, the company and the handle.
  5. A tool marked for the page only has the same checks as every other tool.
  6. Every refusal has plain words on the screen, never an empty box.
  7. Every value from the data is shown as text, never as page code.
  8. The page tells the model what is showing when it changes.
  9. A token that expires or is removed while the screen is open stops the next call (row C7).
  10. The page uses only the standard, and checks before it uses any one AI app's extras.
  11. Served skills have complete manifests and names that will not collide.
  12. A job that signs in as itself has its own identity and the least scope it needs.
  13. Every AI app and plan you promise has been tested, phones included.

F7. What to check again, and when

Check these again before every lab run and every customer release, and at least by the date in this page's front matter:

  • the MCP project's support table, for which AI apps support which extension17
  • the stage of each extension in F02
  • the version of @modelcontextprotocol/ext-apps and the install commands in F06
  • each AI app's own pages in F1 to F3, especially anything marked as rolling out or in public preview
  1. MCP Apps specification, revision 2026-01-26, Model Context Protocol, accessed 15 November 2026.↑
  2. Extensions overview, Model Context Protocol, accessed 15 November 2026.↑↑2↑3↑4↑5
  3. @modelcontextprotocol/ext-apps 2.0.0 README, npm, accessed 15 November 2026.↑↑2
Owner
Panaversity