MCP Apps: your software inside AI assistants

Last week the Model Context Protocol, the open standard that lets AI assistants use outside software, got its first official extension. It is called MCP Apps, and it lets a tool send back a small interactive interface instead of just text. That interface appears inside the conversation: a form, a chart, a map, a preview with buttons.

It went live straight away. Claude now shows interactive versions of Slack, Asana, Figma, Canva, Box, monday.com and a handful of others, in its web and desktop apps. You can ask Claude to draft a Slack message and see it formatted, edit it, then send it. You can turn a messy list into an Asana project timeline without opening Asana. Block's goose agent supports it too, as do preview builds of Microsoft's VS Code, and OpenAI, which helped write the standard, said ChatGPT support would start rolling out the same week.

In November we wrote about generative UI, the idea that AI should answer with interfaces as well as words. This is the practical, standardised version of that idea, and it arrived faster than we expected.

Why one standard matters

Until now, putting your product inside an AI assistant meant picking a platform. OpenAI's apps in ChatGPT, launched in October, used their own SDK. Other assistants had their own approaches, or none.

MCP Apps was written jointly by people from Anthropic, OpenAI and the open-source MCP-UI project, so the same app can appear in several assistants. You build the interactive piece once, as a small web page bundled with your MCP server, and each assistant that supports the standard displays it.

The security model is sensible. The interface runs in a sandboxed frame, separated from the conversation and the rest of the assistant. The assistant can check it before showing it, and can require the user to approve any action the interface tries to take. Messages between your app and the assistant go through a channel that can be logged.

What it is good for in a customer product

The best uses are moments where a customer is already asking an AI for help and a little interface beats a paragraph.

• A property search that returns a map and a shortlist, not ten addresses in a list.
• A quote calculator where the customer adjusts quantities and sees the price change.
• A booking widget showing real slots, which books on a tap after the customer confirms.
• An order or shipment tracker with the next step as a button.

All of these could live on your website too. The difference is that the customer does not have to leave the assistant, find you and start again. Whether that is worth building depends on how many of your customers ask AI tools for help in your category, which is worth checking before you spend money.

The quieter opportunity: internal tools

The bigger change for many businesses may be internal.

Teams already spend part of their day in Claude or ChatGPT. Most of the data they need lives elsewhere: the CRM, the stock system, the sales spreadsheet, a custom dashboard someone built three years ago. Until now, AI tools could fetch that data as text, which was useful but awkward.

With MCP Apps, your internal systems can show up properly inside the assistant. A manager asks "how did the Sharjah branch do this week?" and sees the branch dashboard with its usual charts. A salesperson asks to prepare a quote and gets your quote form, filled in from the conversation, ready to check and send. Staff stay in the assistant they already use.

For a business with a few custom internal systems, this is often a better first project than a public app. The users are your own staff, access runs through your organisation's AI accounts, and the time saved is easy to measure.

What to watch out for

The standard is days old in production. Some things are still settling:

• Each assistant decides how and when to show your app, and those rules differ.
• Discovery is uneven. ChatGPT has an app directory, Claude has its connectors directory, and how users find apps beyond that is unclear.
• Anything that changes data, such as sending, booking or paying, should always need a clear confirmation from the user. Design for that from the start.
• Treat third-party MCP servers like any software you install. Connect only ones you trust, with the narrowest access they need.

Where we would start

Pick one task that your team or your customers do often, where they currently switch between an AI assistant and another screen. Build the MCP server for it first, with the actions it needs and the data rules around them. Then add a single interactive view for the moment where text falls short.

If the assistants change their rules, you keep the server and your data model. If they keep moving in this direction, as they have every month since October, you are already in place.