MCP Apps unlock the future, here's how
After my last few articles, hopefully it’s clear that building MCP servers is about so much more than giving LLMs and agents access to your data.
The way to think about MCP is the same way any professional thinks about building websites and web apps: it’s all about giving users a great experience.
Today I want to showcase a really cool project called MCP-UI that is becoming one of the most important pieces of the MCP experience, and in my opinion really shows where this is going in the very near future.
What is MCP-UI, actually?
MCP-UI is an official extension to the MCP Spec which enables AI agents to display interactive user interfaces returned by MCP servers. Sometimes text is a good experience:
- Prompt: mark issue #452 as closed
- Reply: issue #452 has been resolved
Other times text is a really bad experience
- Prompt: start a timer for my cookies
- Reply: timer started…
In that situation, it sure would be nice to actually see the countdown and have buttons to pause and resume the timer. That’s what MCP-UI is for. And it’s getting built into the spec as “MCP Apps.”
The right UI at the right moment.
The amazing part is that this new addition to the MCP spec makes it possible for your MCP server to deliver interactive visual interfaces directly to the user’s chat window.
I’m not talking about formatted markdown. I’m actually talking about familiar UI elements from the webapps we build every day: structured inputs, data tables, forms, pickers, etc.
Yes, right inside the ChatGPT window!
And because you’re within an AI chat interface, these familiar UI elements are even more powerful.
Think of it this way: MCP-UI lets you present the right UI at the right moment.
Everything they see AND do stays inside the flow of the conversation. Now the agent itself has the ability to present visual and interactive structures, without needing to navigate a bunch of menus or figure out a custom filtering UI to decide what info they want to see.
This is a huge step forward in the agent UX story. Additionally, this reduces context bloat because the app gets to decide what to include in context and what to simply show the user (no sense to include the HTML for the “delete” button, but very useful to include a simple text representation of the state of the list so the agent stays in sync with the UI).
It gives users the speed of natural language when that’s the best tool, and the clarity of UI when text starts to fall apart.
As usual, one of the easiest ways to understand is to see it.
Here’s an example of the Figma app inside ChatGPT:
And an example of a task manager app inside Claude:
Isn't it awesome?
These screenshots are cool, videos are even better. Here are a couple if you’re curious what MCP UII looks like in action:
- This cool 30 second demo from gching.
- Or this in depth 30 minute walkthrough of a demo, complete with code, with me
It's still early days — the Claude example is from their recent announcement, and not available yet — but already a massive UX unlock.
So how does MCP-UI work?
- You declare a UI template in your custom MCP server as a resource.
- You attach that UI to a tool for the model to use.
- When the user's LLM calls the tool, the UI appears in the client.
- The UI is interactive: it can send data to your server using MCP transport, and your server can respond with updated UI or structured results.
That’s the whole loop.
Of course, under the hood there’s a lot of detail (predeclared templates, sandboxing, auditable communication) but the core idea is simple:
Your MCP server provides the UI. The agent host renders it. The user gets something that actually makes sense.
Cool, right?
Before + after MCP-UI
This is why I keep saying: MCP isn’t just an API surface…it’s a new user interaction model.
And MCP-UI gives you the tools to actually treat it that way.
You can't get that from a regular old API wrapper.
How to design a great MCP-UI
At a high level, an excellent MCP-UI is guided by a few core principles:
- Let natural language handle what it’s good at: Conversation, exploration, brainstorming, flexible inputs.
- Let UI handle what text is terrible at: Precise structure, complex interactions and visualizations, and interactivity.
- Keep the user in flow: The moment they leave the chat context, friction goes up and creativity goes down.
- These principles are your north star. Everything else hangs off them.
The best time to start is today…
MCP-UI is available now, and it works now, in ChatGPT (via their Apps SDK), Goose, Postman, Nanobot, and more.
You’ll learn how to build it yourself in Epic MCP.
In Epic MCP, you'll learn:
- How UI fits into the agent UX model
- How to choose when to use UI vs natural language
- How to design flows agents can actually navigate
- How to build UI resources that work across clients
- How to structure your server so the whole thing feels intuitive, not bolted on
- And you'll learn it all through building your very own working implementation.
…and for the future
We're still early days. MCP-UI works today, but an even better and more widely-supported version — MCP Apps — is coming down the pipeline, from a collaboration with Anthropic, OpenAI, and the MCP-UI maintainers.
I'll tell you more about it in a future article. But the best and most important news item is this:
MCP Apps builds on MCP-UI.
So when this next big upgrade officially lands, you’ll already be ahead of the curve.
Bottom line:
If you want your app to feel native inside ChatGPT, Claude, and your users' favorite agents… Epic MCP is how you get there.
Enroll in Epic MCP today and start learning instantly: Epic MCP: Go from Scratch to Production