MCP Apps Explained for Owners: The Tool Opens Inside Claude, ChatGPT or Copilot (2026)

TL;DR: the chat window stopped being the whole product

For two years the AI pitch was "ask it anything." The problem was what came back: a wall of text about your spreadsheet instead of the spreadsheet. MCP Apps fix that one specific problem. The tool brings its own screen into the chat.

Anthropic pushed this into the mainstream on November 10, 2026: connected tools in Claude can now open beside the conversation or full screen, and Claude can edit the open page, row or file with the same approvals you already click for connector actions. What that changed for the connectors you may already use is on Claude connectors for small business. This page is about the standard underneath it, because it isn't a Claude feature. It's plumbing every major AI app now shares.

Connector vs MCP App: the one-line test

A connector (an MCP server) gives the AI tools: "find the overdue tasks," "create an invoice." The AI does the work and tells you about it in text. An MCP App is a connector that also hands the AI app a small web interface to show. The protocol docs describe it plainly: a tool declares a ui:// resource, and the AI app renders that page in a sandboxed frame inside the conversation.

Here's the test on something real. This site runs a public MCP connector at justinmckelvey.com/mcp with one tool, a free AI visibility check. Add it to Claude and ask about your company, and you get a text report. That's a connector. If the same tool returned a clickable scorecard you could sort and filter inside the chat, it would be an MCP App. Same server, same data, one extra resource.

Why the difference matters to a business: text is fine for answers. A screen is better when you need to check something before it happens: approving 14 expense lines, reviewing a schedule, picking from 30 options. The protocol's own "when to use it" list is exactly that: complex data, setups with many options, rich media, live monitoring and multi-step approvals.

Where MCP Apps show up: Claude vs ChatGPT vs Microsoft 365 Copilot

Lined up from each vendor's own documentation (read November 14, 2026):

The practical read: in Claude it's a consumer-style feature you switch on yourself. In Microsoft 365 Copilot it's an IT project, because apps arrive inside agents. ChatGPT sits in between. How ChatGPT's side of connectors works today is on ChatGPT connectors, and how Claude packages connectors with skills is on Claude plugins.

A worked example: Monday's pipeline review without the tab-switching

The numbers below are illustrative, not from a real business.

The situation. A 12-person marketing agency runs projects in Asana and its sales pipeline in HubSpot. Every Monday the owner spends about 45 minutes flipping between the two, asking "what's late and who's waiting on us," then pasting notes into Slack.

With connectors only (text). The owner asks Claude for overdue tasks and stalled deals. Claude answers in a list. The owner still opens Asana to reassign three tasks and HubSpot to log two follow-ups, because a list can't be clicked.

With MCP Apps. The owner opens the Asana project beside the chat and asks: "Move everything overdue for the Lopez account to Thursday and assign the copy tasks to Sam." Claude makes the changes in the open board, the owner approves them, and the board updates in place. Then the same for HubSpot follow-ups.

The time math. If the review drops from 45 minutes to 20, that's about 20 hours a year for one person, on tools the agency already pays for. The real gain is fewer "I'll do it after the meeting" items that never get done, because the change happens during the review.

Should your business build its own MCP App?

Usually not yet. The test I'd use:

• Use the vendor's app if the tool you need is on the list above or in your AI app's directory. You get it for the price of the plan you already pay.
• Use a plain connector if the answer is a fact or a short list. A screen adds nothing to "what's our open balance with this client?"
• Consider building only when three things are true: the data lives in a system no vendor has connected (an in-house database, a niche industry tool), people must review or approve something visual before it happens, and the same review repeats every week for several people.

Building one is real software work: an MCP server, sign-in, a small web interface, and testing in each AI app you care about, since each host supports a slightly different set of extras. Anthropic calls its newest pieces (app tools and embedded sites) a public preview for developers, so expect changes. If you already know which internal system you'd connect and want it scoped, that's the kind of job our AI implementation work starts with.

What to watch out for

• The app signs in as you. It sees what your account in that tool sees. A sandbox protects the browser page, not your data permissions.
• Approvals still matter. Claude asks before acting in an open tool the way it asks for connector actions. Don't train your team to click "allow always" on tools that change money or customer records.
• Admins can narrow it. In Claude, if an organization already allows all of a connector's tools, the new app features appear as they roll out. Restrict the tools first if that's not what you want.
• Pin what you use. Pinned tools sit in the sidebar next to your Claude artifacts, which is where Sheets, Whiteboards and Data now live too.
• Not every host shows every feature. Microsoft's docs warn that some of the OpenAI-specific extras are missing in Copilot. If an app looks wrong in one AI app and fine in another, that's usually why.

Frequently asked questions about MCP Apps

What are MCP Apps?

MCP Apps are an open extension to the Model Context Protocol that lets a tool show its own interactive screen inside an AI chat instead of only returning text. A tool's MCP server points to a small web interface, and the AI app (Claude, ChatGPT, Microsoft 365 Copilot and others) renders it in a sandbox next to the conversation, where both you and the AI can work in it.

What's the difference between an MCP connector and an MCP App?

A connector lets the AI read and act on a tool and answer you in text. An MCP App adds the tool's own screen: a board, a form, a document editor or a dashboard you can click, which the AI can also update. Every MCP App is a connector with a UI; most connectors today have no UI.

Which tools have MCP Apps in Claude?

Anthropic's November 10, 2026 announcement lists Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel and Webflow as the first tools that open in full, with examples from Adobe, Atlassian Confluence, Box, HubSpot and Notion. Figma, Harvey and Shopify are building theirs. It's rolling out on the web and desktop app on all plans.

Do MCP Apps work in ChatGPT and Microsoft Copilot?

Yes. OpenAI's developer docs say ChatGPT implements the open MCP Apps standard for UI returned by an MCP server, and Microsoft's documentation says MCP apps render as interactive widgets inside Microsoft 365 Copilot through agents that an MCP server-based plugin is added to. The same app can run in more than one of them, with small host-specific differences.

Are MCP Apps safe to use with company data?

The app runs in a sandboxed frame the AI host controls, so it can't read the rest of the page, and actions still go through the host's approval prompts. The bigger question is the tool account behind it: the app signs in as you, so it sees what your account sees. Admins can restrict a connector's tools in Claude the same way they do for any connector.

Sources

• Anthropic, "Your apps can now open in Claude, and Claude can work in them" (claude.com), published November 10, 2026, read November 14, 2026: full tool beside the chat, edits with approvals, all plans on web and desktop, the named tools, built on MCP Apps, developer public preview, admin controls.
• Model Context Protocol, "MCP Apps" overview (modelcontextprotocol.io), read November 14, 2026: ui:// resources, sandboxed rendering, two-way tool calls, when to use MCP Apps.
• OpenAI Developers, "Add UI to your MCP server" (developers.openai.com), read November 14, 2026: ChatGPT implements the open MCP Apps standard.
• Microsoft Learn, "MCP apps in Microsoft 365 Copilot" (learn.microsoft.com), read November 14, 2026: interactive widgets in Copilot via MCP server-based plugins in declarative agents, OAuth 2.1 / Entra SSO, host API differences.