Beyond Components: Designing Generative UI for MCP Apps — Ruben Casas, Postman

0:07[music]

0:16>> Hello everybody. So, I know I am the

0:17person standing between you and your

0:19lunch. And but this is going to be a

0:22very interesting talk that combines the

0:24the previous two talks into the future.

0:27And that's what I want to talk about

0:28today.

0:29So, back in November 2022, what we used

0:32to do was we used to go to ChatGPT and

0:34ask ChatGPT to

0:36create a component. And we would just

0:39copy

0:40paste. You have to ask reply in code

0:42blocks. Then, you know, again, fix it,

0:45repeat. And this is what I call the poor

0:48man's by coding.

0:50And we have come a long way.

0:53It kind of worked. It was very exciting.

0:57You could get models to could actually

0:59build some UI for you.

1:02And I'm sure that was not going to

1:05write better code than me, right?

1:07And then things improved very very

1:09rapidly, very very fast.

1:12What happened last year, and if you are

1:14aware of what happened in the the the

1:16last months of 2025 was this

1:18acceleration, an incredible inflection

1:21point where things changed. And it will

1:25go down in the history books as

1:27things changed very fast all at once.

1:30And this is in part because of the

1:32release of two very important models,

1:35which were

1:365.2 ChatGPT, sorry. It was yeah, a GPT

1:395.2 and Opus 4.5.

1:42And they were not just very good at uh

1:45most of the task, long horizon tasks.

1:48They were also very good at high

1:50fidelity UI generation.

1:53And they were producing very good

1:56working UI.

1:57Sometimes thoughtful, sometimes really

2:00really good. And also very fast.

2:05Now, I experienced this when I tried one

2:08of these models, tried to rewrite my my

2:10blog. I know people have used this in a

2:12more creative ways, but I just tried,

2:13you know, a single prompt, rewrite my

2:15blog. And then it did this, which I

2:18didn't ask for.

2:19It created a nice nice search box

2:22with a blur animation, with

2:25accessibility out of the box.

2:27And then that's when I realized that in

2:30the space of of 3 years from when

2:32ChatGPT was released to today,

2:35we went from

2:36you know, few lines of code is great.

2:40It can it runs. Oh.

2:43And now it can write

2:45better front-end code than me.

2:49And you know, I I don't mind. No no ego.

2:52Uh

2:53It's just reality. So, here's the

2:56question. If these models are so good at

2:58writing UI code,

3:01why are we still stuck in this mainly

3:04old

3:05paradigm of mostly static UI?

3:08And where is where is that Jarvis moment

3:11that we've been talking about earlier?

3:13Where are my floating UI windows that

3:15appear and disappear? And why we not

3:17there yet?

3:19So, my name is Ruben Casas. I am a staff

3:21engineer at Postman. And I've been

3:23looking at UI and generative UI for the

3:25past year, and I've been working with

3:26MCP apps as well.

3:28And today I want to show you what we're

3:30doing today and where we're going in the

3:31future.

3:33So, the news are we have a new computer.

3:36And as Andrej Karpathy put it,

3:39interacting with this new computer is

3:40like talking to the terminal. You have

3:43direct access to this operating system.

3:46And the GUI has not been invented yet.

3:48It's like we are in the '70s, where

3:50everything was just text.

3:52And we have a super intelligence, but we

3:54don't have a mature interface language.

3:57And today we are still trying to figure

3:59out what is this new interface for for

4:03this computer.

4:05And people ask, is it chat?

4:07I'll show you what we're what we're

4:08doing today. And actually this was a

4:09very recent tweet last week, where

4:12people were complaining that most SaaS

4:14companies have been adding

4:16chat to their their homepages, and

4:19everybody's just putting chat

4:20everywhere.

4:21And and that's that's fine. I don't have

4:23a problem with chat. It's not the final

4:26UI. It's okay for now.

4:29But the question is, if it's not chat,

4:31then then what is the interface

4:33for this computer?

4:35On the other hand, as we have seen with

4:39MCP apps,

4:40is there [clears throat] is another

4:42thought, which is we will have one app

4:44or super app to rule them all. And this

4:47is where MCP apps comes in, where

4:50instead of putting all of these chat

4:51windows into your homepages and and to

4:54every single app that you use, we will

4:56have a super app like ChatGPT or Claude

4:58or Gemini, where you will be interfacing

5:00with most of the the UI and the websites

5:03that we have today.

5:05And this is good. This is the way we

5:07using MCP apps today to to render

5:09third-party UI inside one agent

5:11environment.

5:13Now, these two options both could be

5:16valid. And and I believe these are part

5:19of the evolution towards finding out

5:22what is that new interface for that

5:25computer. And to be honest, I don't know

5:26which one is going to be the the final

5:29one. Consumers will tell us.

5:31But one thing is that these are two

5:33different questions.

5:35The question is where does the

5:37UI runs? In this case, is it third-party

5:40UI, a super app, or in this case, chat

5:42everywhere?

5:43But most interesting is what is the

5:45model generating. And this is what I

5:47want to talk about today in terms of how

5:49are we generating this UI? And we have

5:51seen this. We have mostly static,

5:53declarative, and generative UI.

5:57And I'm going to describe briefly this

6:00one. So, we have um

6:02to start with the the static components

6:04way of running running UI, which is what

6:06most agents do today.

6:09The agent is just an orchestrator. The

6:12agent makes a tool call via MCP apps or

6:15direct agent tool call. Then we will

6:18have some parameters

6:20and data passed to predefined static

6:23components that

6:25have been created by developers. And

6:27this is very similar to what what we

6:28have been doing

6:30for the past 20 years with with UI. And

6:33then the client renders the component.

6:34And if you see here, it's very similar

6:36to just getting a server to send some

6:38data, and then the UI will be rendered

6:40by the client. But in this case, the

6:42agent will be generating that data and

6:44the props to to do this. And some

6:47examples that we have today are the AGUI

6:50protocol. They have an SDK when you can

6:52register a client tool that maps to a

6:56React component. The tool call will

6:59receive some props. Those props will be

7:01mapped to a static component that then

7:04will be rendered to to the user.

7:06Another example is Goose.

7:09Goose is a

7:10an MCP client where you can try most of

7:12the MCP features. And Goose has this

7:14really interesting feature called Goose

7:16Auto Visualizer, where you can just pass

7:18any type of data to Goose, and Goose

7:22will try to match that data, organize

7:24it, and then pass it to a set of

7:26predefined components that the Goose

7:28team have created. In this case, we have

7:30a few interesting components that you

7:31can

7:32use to visualize your data.

7:35So, that's the static way. That's the

7:38most common way of generating UI today.

7:41But I have seen an evolution recently

7:43where we call now declarative UI.

7:46And declarative UI takes it to the next

7:50level. So, we will still have some

7:52predefined static components that

7:54developers build, and it contains your

7:56design system and all these

7:58components that you have.

8:00But instead of the agent just passing

8:02the props and the data, the agent uses

8:05a descriptor that could be either JSON

8:07or YAML. Or I've seen Python as well

8:10with with fast MCPs, where they have a a

8:12descriptor in Python that maps to these

8:15predefined static components.

8:17And then you have this translation

8:20rendering engine that takes those

8:22descriptors and converts them into the

8:24final UI.

8:26How is this different? Well, in this

8:27case,

8:29it is more dynamic. They are still

8:32static components, but it's more

8:34personalized.

8:35And you if you look at this and you

8:37think that this might look familiar as

8:38well, it's because it is not new.

8:41Netflix has been doing this for a long

8:43time since the the personalization and

8:46server during UI era, where when you go

8:48to the Netflix homepage, you will get a

8:51UI that is completely personalized to

8:53you. But that's still mapped to the

8:56Netflix components and UI elements.

8:59Another very

9:01good um

9:02tool that I've seen recently is JSON

9:04Render. JSON Render is being built by

9:07Vercel. And it is a way to map your

9:11components using JSON and also YAML.

9:13They released the YAML support recently.

9:15And create all of these very dynamic,

9:18very good UI interactions that you can

9:21use today.

9:22But now JSON Render still, they say, is

9:27constrained to your static components.

9:32And yes, they're still static

9:34components. The LLM is not generating

9:37these components. The LLM is generating

9:39the JSON.

9:40However, I think in at this point in

9:42time, declarative generative UI is

9:45probably the perfect balance today in

9:47terms of flexibility and consistency.

9:51Because you would still want your design

9:53system. You still want to have uh,

9:55predictability of what the UI is going

9:57to be generated, also faster and also

9:59potentially cheaper at this point. Uh,

10:01so you don't create and use a lot of

10:03tokens to create the UI.

10:06But as I mentioned at the beginning, why

10:07why are we still stuck here?

10:09And what's the next level? I think the

10:11next level will be uh, generative

10:13components.

10:14And generative components

10:16uh,

10:17goes into like the the premise I I put

10:19at the beginning where the models are

10:20good at writing front-end code. They are

10:22good at writing React. They're good at

10:23React creating in JavaScript, CSS.

10:27And the question is why we don't let

10:29them just write that

10:32on demand at runtime.

10:34What could possibly go wrong with that,

10:36right?

10:37This model um, of generating the UI uses

10:41the agent capabilities. And in this

10:43case, you can also use a tool call, but

10:45instead of calling this uh, layout

10:48rendering engine, you can call the same

10:50model with reverse sampling or you can

10:51call another model that will generate

10:54the HTML, CSS, JavaScript

10:57on demand and then it will be passed to

10:59the client.

11:00I did this experiment um, I work at

11:02Postman on this experiment where I

11:03created uh, this weather agent that goes

11:06to the API, the weather API. It creates

11:08a joke. It creates the HTML, CSS,

11:11JavaScript, all in one tool call.

11:13And you get presented with this random

11:16but very uh, imaginative UI where

11:20everything is created by the agent.

11:22There is no component. There is no

11:23translation.

11:25So there is of course a problem with

11:27this approach. Uh, and the problem with

11:28this approach is uh, if we

11:31don't trust third-party code,

11:34well, we should not trust um,

11:36code that has been generated by LLMs and

11:38then just present it to the user.

11:41Uh, generative UI and and this level of

11:44generative UI needs a distribution

11:46model.

11:47And this distribution model requires a

11:49boundary, requires containment and

11:51requires a sandbox, which is what we

11:53were talking about earlier.

11:55This is what I think MCP apps matter a

11:58lot because MCP apps are the best uh,

12:01delivery mechanism uh, for

12:04generative UI.

12:06We have the features provided by MCP,

12:09including authentication and tool

12:10calling and message passing between the

12:12UI and the agent. Uh, it's sandboxed by

12:14default with that double iFrame. Is the

12:17default for third-party UI delivery

12:19today. Does this become the standard?

12:22And one interesting thing is it's not

12:24just for

12:25for third-party UI. It can also be used

12:28for first-party UI.

12:31And this is why I think what Anthropic

12:33is doing with the the visualizer feature

12:36is very interesting strategically

12:38speaking because

12:40they could have just created their own

12:43rendering

12:44um, and an architecture mechanism for

12:46delivering this in interaction in in

12:49cloud,

12:50but they decided to go with MCP apps

12:52because MCP apps provide most of those

12:55um, features that I mentioned earlier

12:57uh, out of the box.

12:59So if Anthropic decided to use MCP apps

13:01for their first-party UI, uh, you can

13:03ask yourselves why cannot we do the

13:05same? It is

13:07um, a very very strong protocol. And

13:09especially when the UI is being

13:10generated on the fly by the agents, by

13:13the the code uh, coding models,

13:15then is the best um,

13:17mechanism for delivery.

13:21Now,

13:22today is probably not the final form.

13:26And people keep saying

13:28is chat the final form? Is MCP apps the

13:31final form?

13:33We're still we're still trying to figure

13:34this out.

13:36And the obvious future

13:39is probably too obvious.

13:41And we said about where is my Jarvis?

13:44Where is my

13:45floating windows?

13:47And if you think about it, that's the

13:50obvious things thing that people would

13:51think how we would look like if we were

13:53generating or creating a new uh, user

13:56interaction.

13:57But what I think is

14:00we don't have enough imagination yet.

14:02And and this analogy I heard recently is

14:05very interesting when when uh, radio

14:08came out in the 30s,

14:10um, the the the first um, sorry, the the

14:13TV came out, the first

14:15uh, TV shows were radio shows

14:18with cameras

14:20because they could not imagine what you

14:22could do with this new technology. So

14:24this new technology that we have today

14:26is very similar when television came out

14:28and we are still in the radio era where

14:32we don't know all the amazing things

14:34that we will do in the future with this

14:36new media, with this new power that we

14:38have with the with the new computer.

14:40And we can see that we cannot even

14:43imagine what it's going to look like.

14:46Uh, of course this is a speculative, but

14:47what do I think is actually going to

14:49happen

14:50is we are going to be moving uh, beyond

14:53components and more towards a

14:55collaboration uh, through human agent

14:58collaboration.

14:59If you haven't heard about the

15:00Excalidraw MCP app, uh, definitely check

15:03it out because

15:04the Excalidraw MCP app

15:06is not just for um, output and

15:09visualization of diagrams.

15:11The Excalidraw MCP app does something

15:13very interesting, which it creates a a

15:16shared artifact.

15:18It creates a canvas where a human and an

15:22agent can collaborate together into a

15:25shared space where you can go back and

15:27forth with the agent and ask, you know,

15:29change this, but you can also click

15:31around, modify the UI the way that you

15:33are used to. And that becomes the new

15:36way of interacting, the new way of

15:40experiencing the the agent um,

15:42powers.

15:44And and at the moment again, we are very

15:46constrained to our imagination. And and

15:48I believe these agents are very very

15:50powerful for just to just use them as a

15:54orchestrator and a delivery mechanism to

15:56show me some visualizations.

15:58So I believe beyond components,

16:01it will be the future of generative UI

16:04will be more a collaborative experience

16:06where yes, we will have some generative

16:09UI, but that UI is going to be super

16:12personalized and it's going to be

16:14collaborative.

16:16So we we are still early.

16:19Um,

16:20we don't have the answer. People say,

16:22you know, what is what is the future of

16:25the user interaction, the user

16:26interfaces? We don't know yet.

16:28But we can shape that future uh, and

16:31create this uh, new computer.

16:34And that's me. Thank you so much uh, for

16:36listening to this.

16:37>> [applause]

16:37>> You can find me

16:39and ask any questions.

16:42Thank you.