It might be actual industry interest, or it might be the buying power of two corporate juggernauts’ marketing machines, but the new Starbucks app within OpenAI’s ChatGPT that let’s you create a coffee based on “vibes” is generating a lot of media.
For you, the interest is that the Starbuck’s app is an MCP App that’s letting their customers access their products from inside the UI that more and more people are spending more and more time in. When in Rome…
We’re going to give you a quick overview of MCP Apps and help you decide if you should build one.
What are MCP Apps and where did they come from
MCP is Model Context Protocol. It’s a standard announced, and open sourced, by Anthropic in November 2024 that provides a way to give AI access to APIs. It provides detailed description of what API endpoints do and expected data and return types. This is enough information for AI agents to understand how to request data, and the necessary information for the agent harness (ChatGPT, Claude Code, Cursor, Copilot, etc) to pass the data back and forth between the API and the agent.
MCP was an instant hit. Now you could connect AI to anything – databases, Stripe payments, coffee machines… Best practices in security and management followed and with the addition of proper authentication and permissioning, it has been embraced by enterprise. It provides the perfect interface for giving employees secure and monitored access to inhouse resources through the ubiquitous ChatGPT/Claude clients.
In October 2025 OpenAI announced their OpenAI Apps SDK. It was MCP with added on UI. That UI is built and rendered like ordinary web content, and so can provide users with any kind of interface and feature you’d like.
In November 2025, the App extension to MCP was launched as a provisional addition to the standard. This wasn’t a competing standard. It was heavily aligned with OpenAI’s Apps SDK, and in January this year it became a part of the Model Context Protocol.
OpenAI’s App SDK still has its own proprietary elements. OpenAI App SDK let’s your MCP App access local files, trigger checkout flows and use a bunch of other ChatGPT specific APIs.
It is interesting that despite OpenAI’s App SDK providing checkout flows, Starbuck’s MCP App opens their own website (or app if on a phone) to actually complete the order. This might be to avoid the inevitable fees going through OpenAI’s payment gateway incurs. We can only wonder if this option will remain available or if they will pull an Apple require all payments to go through their gateway so they can take their cut.
How MCP Apps work in a nutshell
MCP Apps rely on the harness. Whether it is ChatGPT, Claude, Claude Code, Copilot, etc, etc. The harness is the middleman between everyone:
User ↔ Harness ↔ Agent
MCP App widget ↔ Harness ↔ Agent
MCP App widget ↔ Harness ↔ User
MCP server ↔ Harness ↔ Agent
MCP server ↔ Harness ↔ Widget
This is to ensure that MCP Apps are safe to use. The UI itself runs in a sandboxed iframe. It can only talk to the outside world, including the server that delivered it to the harness, via the harness. If it wants to load data, it calls a function in the harness. If it wants to provide the agent with fresh information for its context, it calls a function in the harness.
Microsoft has some good examples of MCP Apps with complex UIs that allow the user to navigate information manually, while also asking the agent to take actions on the data.
The user can explore with the UI. The UI can send updates to the agent. The user can ask the agent to take actions based on what is happening in the UI. And if the MCP App provides the right tools, the agent can carry out actions that are reflected in the UI.
What an MCP App can do is limited by what tools (ie API access points) and data you expose to them.
Do you need to build an MCP App?
That depends on your users and your product. Are your users heavy users of AI? Have you checked?
Is your product part of a broader process or workflow? Is it a “mission control” style system?
Does your product generate data of any kind that needs to be communicated outside of your product? Do you already have some export functionality?
Then your users might appreciate an MCP App.
How to build an MVP MCP App yourself
The MCP group have all the documentation you need to build an MCP App. And they recommend you start by installing a skill in your AI coding agent and asking it to build it for you.
Of course, you need to give some thought and planning to authentication and permissions and payments and all the other complexities that make a product viable on top of how you are going to implement your UI.
You probably already have a UI. Can you retarget it to the MCP App interface easily?
Welcome to the next and possibly last API
One of the major shifts in building web-based products was the separation of apps into APIs and UIs. It meant your backend could drive an Android App as well as a website. Or an iPhone App. Or an AppleTV app.
MCP Apps are the next target for your APIs. And given the way AI is eating software the way software was supposed to eat the world, it might be the last target. At least the last target you need to write yourself.
