MCP UI: Extending the frontier — Liad Yosef and Ido Salomon, MCP Apps

0:07[music]

0:14>> Okay, hi everyone.

0:17We built this talk well not really

0:18yesterday we built this talk this

0:19morning and it might already be out of

0:21date. I'm Eda Solomon. I'm the creator

0:25of MCPUI,

0:26co-creator of MCP apps and maintainer.

0:30Also creator of Agent Craft if you were

0:31on the previous session.

0:34I'm Liad. I work with Eda on MCPUI,

0:37co-created the MCP app spec with Eda and

0:40I'm also co-founder of Ergo Labs which

0:42is like human agentic interfaces a

0:44company.

0:46So MCP apps are all around us. You might

0:49not even realize it but the interactive

0:51applications you see today in ChatGPT,

0:53Claude and others are actually based on

0:55MCP and the MCP app spec.

1:02But why do we need MCP apps and what's

1:04MCP apps? I mean you heard David this

1:05morning talk about it a little bit.

1:08We used to text, we used to

1:10MCP tools sending text to our chat

1:12agents.

1:13But that's not ideal, right? Because

1:15chat chat is really text is really bad

1:17and actually this was one of the main

1:19blockers of companies or

1:22tools not to send their data to ChatGPT

1:24because they didn't want to be reduced

1:26to like

1:27this thing, this wall of text where you

1:28don't have identity. You don't know if

1:30if this information came from Shopify,

1:32Booking, Expedia or any other company.

1:35But what if

1:37every tool or every company could just

1:38send its own UI to the chat? So instead

1:41of us looking at

1:42this thing we can just imagine that for

1:45the parts that are relevant

1:47we can have the relevant UI, the

1:48relevant UI for Shopify, from Hugging

1:51Face, from Monday and this can can be

1:54not only presentation as this can be

1:56interactive. So we want to be able to

1:58respond to a user click on this Hugging

2:01Face widget.

2:04So we don't have to imagine it anymore.

2:06So back in May last year I released

2:08MCPUI. The concept was pretty simple.

2:11There were a bunch of stuff around it

2:13but the concept was pretty simple of how

2:16do we take UI and find some way to pass

2:19it over MCP. We need some general way to

2:22do that so we can both have MCP over

2:26sorry have UI over MCP and have the

2:29communication between the UI and the

2:31host. Obviously it also had community

2:33SDKs and the general motivation was

2:37we don't need to throw away everything

2:39we know about UI and UX just to get into

2:41this new world of agents.

2:44We can simply adapt and use that and

2:47preserve our branding and identity and

2:49still be practical.

2:51And

2:52just a few months back

2:54MCPUI we partnered with Anthropic and

2:57OpenAI to really put this into the MCP

3:00standard as the first official extension

3:03called MCP apps.

3:05And as you can see here it made kind of

3:08a big splash. We had support from a

3:09bunch of hosts. Now VS Code and Cursor

3:12and Claude and ChatGPT and like

3:15Microsoft Copilot and a bunch of others

3:17already adopted it and you have these

3:19really cool interfaces built in right to

3:22your your assistance.

3:25And going back a little bit there were

3:27early adopters for MCPUI. So these are

3:29some of the companies

3:31shout out for Hugging Face Sean if

3:33you're here that adopted MCPUI. So even

3:36a year ago Shopify was already sending

3:38MCPUI chunks of all of millions stores

3:41of Shopify online stores send MCPUI

3:43chunks. Hugging Face all of Hugging Face

3:45spaces were MCPUI widgets.

3:49And now once it's standardized we have

3:51much bigger adoption.

3:53So we have a VS Code, we have a Cursor,

3:56we have Copilot,

3:58GitHub ChatGPT are supporting MCP apps.

4:01Not only that ChatGPT are recommending

4:03MCP apps as the way to build ChatGPT

4:06apps. So it's really standardized.

4:09Obviously shout out Postman and Goose

4:11and Claude

4:13the first one that released Claude apps

4:15that actually supported MCP apps. But

4:17it's not just support from the big

4:19companies or the big hosts. We also have

4:21huge community adoption.

4:23We have people that building plugins

4:26around MCP apps, building workshops

4:28around MCP apps

4:30building all all kind of support around

4:32MCP apps. Spy just announced support in

4:34MCP apps which is amazing. It's like a

4:36terminal, right? But we have UI in the

4:37terminal right now and we have all of

4:39these advocates that that are speaking

4:42about MCP apps. There are even companies

4:43that are built around MCP apps to help

4:46other businesses build those apps.

4:48We have an official MCP apps repo with

4:52with Anthropic and OpenAI. There's a an

4:54amazing community engagement. We

4:56recommend for you to check it out. We

4:57have a

4:58a work group public work group meetings

5:01and we're meeting tri-weekly once every

5:033 weeks just to push the standard

5:05forward because as we can see it's going

5:06to be the global standard for UI inside

5:09chat apps. And we're going to talk a

5:11little bit about the concepts behind

5:13MCPUI.

5:14Yeah, so let's talk about the core

5:15concept. So the first and obvious one is

5:18how do we even pass UI over MCP?

5:20Um

5:22so in the old world of a few months back

5:25when we wanted to do whatever let's say

5:27we wanted to create the best playlist

5:29ever we would type something into the

5:31chat and it would send out a tool call

5:34to our MCP server.

5:35So far so good. Um what we would get

5:38back in response

5:40would be text

5:42and as you know text is sub-optimal. But

5:45if we are using MCP apps what we can do

5:48is instead return a resource. So we can

5:51return this actual HTML back to the

5:54host. The host supports MCP apps so you

5:56can take that HTML and transform it into

5:59an interactive application.

6:01And when we say interactive we mean

6:03interactive. So this is not just

6:04presentational. MCP apps also

6:06standardize the way that this UI can

6:09talk to the user and to its back end

6:11because just imagine if the user wants

6:13to favorite this song then the

6:16sub-optimal thing that would happen is

6:18that for this UI to speak to Spotify's

6:20back end and and favorite this song and

6:23then later when the user asks Claude

6:25remind me which song I favorited Claude

6:27wouldn't know because the UI spoke

6:28directly to the back end. But MCP app

6:30standardizes this message passing so

6:33that every UI chunk sends message back

6:35to the host. The host gets this message

6:37message in this case like a tool call

6:39and the host decides what to do. In this

6:41case it decides to actually call the

6:42server tool but the control is in the

6:45hands of the host and everything stays

6:47in context.

6:48Okay, so seeing is believing. Let's see

6:51a quick example of what that looks like.

6:53So this is Claude like actual Claude.

6:56And let's say that I want to do

6:57something like analyze my funnel. So I

7:00type that in in the old world it would

7:03go out to let's say PostHog and I would

7:06get this textual response which is

7:08accurate

7:09but it doesn't really help me understand

7:10what's going on. I have to read this

7:11whole thing now and kind of try and see

7:14what's the deal.

7:15But with MCP apps instead of doing this

7:18I can just say show me.

7:20And now

7:23Okay, the clicker is not yet up to par

7:25but now what we'll have is this nice UI

7:29visualization actually created by

7:31PostHog. So they control the identity,

7:33the experience. It's actually their

7:35component that you would see in the

7:37website.

7:38And now I have like a really cool way to

7:41just see the funnel in one glance.

7:44That's not all. I mean the MCP app isn't

7:46just UI generated by the server.

7:49There are also really cool innovations

7:52from Anthropic and other companies

7:55to do generative UI on top of MCP apps

7:57or even first party UI in general.

8:00So for instance this Claude feature

8:02says

8:03let's say that I don't know what a

8:05funnel is which is

8:06reasonable.

8:07I can ask what a funnel is and instead

8:09again getting that long textual answer

8:12what would happen is that Claude would

8:14be able to generate this UI for me

8:17explain exactly what I need or create

8:19some UI that I need to do some action

8:22and present it to me in a way that is

8:24very digestible.

8:25This is obviously applicable to a bunch

8:27of other stuff and you'll see it in

8:28other hosts as well.

8:32So let's look at

8:33like another cool thing is here is that

8:35this is not just presentational. Like we

8:38said it's interactive so I can just

8:39click on it and it would give me a

8:42follow-up on the specific step of the

8:44funnel

8:45that I have a questions about. So you

8:46can imagine how this goes into into a

8:49bunch of other directions when you want

8:51to do interactive exploration.

8:54So how does it work

8:56in general?

8:57So

8:59let's go over the stages.

9:01We went to the host and we prompted

9:04something. We asked for

9:06funnel data. What happened is that it

9:09sent out a tool call to our MCP server.

9:11And again instead of just returning text

9:14that tool was actually pointing to a

9:17resource.

9:18That resource was our UI.

9:21So if you look at the

9:24code for it then it's super simple like

9:26you just register a resource and you

9:28just have it. So return that resource

9:31back to the host.

9:32The host because it also supports MCP

9:35apps can take that

9:37transform it, put it if you see like

9:39look at just code wise if you want to

9:41build a host it just

9:42like

9:43react component that accepts that

9:45resource and also this callback which is

9:48the way that we handle messaging between

9:50the UI and the host.

9:52So you take that and you render it

9:53inside a sandbox so it's secure.

9:56Like we said it's not presentational, so

9:58we also click on it. And once you click,

10:01what happens is that there's a bunch of

10:03events going back from the UI, from this

10:05view,

10:07all the way back to the model.

10:09So, it can actually take out to do other

10:13tool calls or even follow up messages on

10:16your behalf or fetch additional

10:17resources,

10:19really completing this end-to-end

10:20bidirectional flow.

10:23So, when we look at that, when we look

10:25at this flow, when we look at this

10:26architecture, it's not just technical

10:28change. It's not not just a technology

10:30that's changing. It's also how we

10:32perceive the web. Because this is

10:34ushering a new web, a web where

10:37we don't need websites. We don't need

10:40all of those tabs just to organize

10:43an anniversary. We don't need to

10:45familiarize familiarize ourselves with

10:48bunch of different UIs.

10:51We don't need to

10:52force ourselves to pass our intents to

10:55dashboards of companies where 90% of

10:57this UI is not relevant for an agent. If

11:00I have a personal assistant, I don't

11:01need most of it. I can just take this

11:04and I can just decompose it to atoms and

11:06let my agent build them for me, right?

11:10Because I I have my assistant I convey

11:11my intent to. So, for example, my agent,

11:14my proactive assistant can say, "Yeah, I

11:16see that you have an important

11:17anniversary coming." And instead of

11:18Google just sending the data, Google can

11:21actually send a chunk right of the

11:22Google calendar. And now, this is a

11:24win-win-win because for Google, it's

11:26amazing. It gets to keep its identity.

11:29For me, it's good because I know I know

11:30this interface. I I recognize that it's

11:32Google.

11:33But it's good for the host as well

11:35because the host doesn't need to render

11:36that.

11:37We have domain experts. We have

11:39companies that spend decades in

11:41perfecting user journeys and we can't

11:43expect Claude or ChatGPT or any host to

11:46automatically generate all those UIs.

11:48And if I continue and I ask something

11:50for Amazon, so instead of Amazon just

11:52sending me the data of the product and

11:54that's we reducing itself to be just a

11:56database, it can just send this chunk of

11:58Amazon. And I look at it and say, "Oh,

11:59it's Amazon. Okay, I know."

12:01And then, I can complete the entire

12:03planning of my anniversary,

12:07the entire planning in just one

12:09assistant chat, right? And you can see

12:11that it pulled just the relevant parts

12:13of it because it pulled

12:15the venue from Booking, but it knows me.

12:17It knows that I prefer something that's

12:18close to nature and not in the city. So,

12:20it also knew to pull the the map from

12:22Booking. That's because the assistant

12:23knew me. Booking doesn't know me that

12:25well, but Booking knows how to how to

12:27book a venue. So, this is real synergy

12:29between those. Um

12:31and we have to think about this new

12:33interaction mindset. Why? Because we

12:35have to remember that in this flow,

12:37the apps, the services, the tools, they

12:40no longer own my journey in the in the

12:42platform, right? If I click something in

12:44Booking, it doesn't go to Booking's

12:46backend. It goes to it goes to the host

12:48like we said. So,

12:51what we did with MCP apps is that every

12:53click, every interaction actually sends

12:55this kind of like message back to the

12:57host.

12:58Um and like we said, it this is breaking

13:01the model for for all of the for all the

13:04companies. So, this is like a new

13:05philosophy.

13:06But the messages can can be put on a

13:09spectrum.

13:10So, this spectrum represents how much

13:12control the UI wants for itself and how

13:15much control it gives to the host. So,

13:16for example, notification, that's the

13:19highest level of control the UI has. It

13:21just notifies the host that something

13:22happened. For example, if I increase the

13:24number of items in my cart, it doesn't

13:26need to to go to the host. It goes back

13:28to Shopify. But just notifies the host

13:29that something happened. A tool call is

13:32the UI telling the host call a tool. And

13:34prompt, that's like the the UI just

13:36releases all control and say tells to

13:38the host, "Just run this prompt and see

13:41what happens."

13:42Um so, MCP apps really standardizes this

13:45new software flow and that's something

13:47that we need to remember.

13:48Perhaps in 2 years, we won't have

13:50browsers as as we know them. We won't

13:52have websites as we know them. We'll

13:53have a personal assistant that accepts

13:56only small chunks of UI and this will

13:58replace our our web journey.

14:01Um 2026 is going to be the year that

14:04we're going to standardize

14:06MCP apps as a global standard for UI.

14:08And um

14:09Yeah, but the spec is still evolving. I

14:11mean, there are a bunch of stuff

14:12happening. Just in those last, I think 2

14:14months, we shipped all of those or

14:17almost all of those based on community

14:19feedback, based on community work done

14:21by the work group, which you can join.

14:23So, you're encouraged to do this.

14:25There's the official SDK, XApps. You can

14:28just use that to build your

14:29applications. It's very simple. There

14:31are built-in skills. So, you just let

14:33your coding agent do it for you. You

14:35don't actually need to code anything.

14:36God forbid.

14:37So, you have this and then,

14:40it's important to remember that the

14:41reason to use this SDK is that it's just

14:45always compliant with the spec. Like we

14:47always update both.

14:48So, feel free to use it.

14:51You can see that just the issues and

14:52stuff that people open on it. So, please

14:54feel free to do it.

14:55So, what's next for MCP apps? Obviously,

14:58there are a bunch of stuff in the

15:00pipeline. But just to give you like a

15:02taste.

15:03So, we have reusable views. The idea

15:05here is that today, for simplicity,

15:08whenever you render an app, we actually

15:10render a new one. So, let's say that

15:12you're working with the same app

15:14multiple times. If you keep re-rendering

15:17it and you have some heavy applications.

15:19So, for example, Autodesk had this

15:21problem.

15:22It just takes a really long time and

15:24your experience will be bad. So, we are

15:26working on ways to solve it. The first

15:29one is just why can't we just reference

15:31that same view and push data into it?

15:34But the second one is actually to take

15:37this and flip the script.

15:40So, another thing that we've been

15:42working on is interactions not for the

15:45user to interact with the with the app

15:47and then the app sends it to the model,

15:48which we just saw.

15:49But what if we want the model to be able

15:51to interact with the view? We want

15:53Claude to be able to click on buttons or

15:55to fill forms or do anything inside the

15:58UI. So, today we have solutions like

16:00WebMCP and things like that. We are

16:02working on a standardized way. So, when

16:04the user interacts with the model, the

16:05model the app can actually expose tools

16:07for the model to interact with it, thus

16:09closing this loop. Um and you can you

16:12can check out the PR. It's still an open

16:13PR, but that's something that we work on

16:15in the in the committee.

16:16And the most important thing is that MCP

16:19apps supports all ways of all the ways

16:23of generating UI. Because that's a

16:25question that we always get asked. What

16:27about generative UI? So, if we put it on

16:30a spectrum, then we have the predefined

16:31UI. That's like the classic MCP app.

16:33That's like Airbnb building its own UI,

16:35sending it to Claude or to ChatGPT.

16:36That's predefined. That's a black box.

16:38That's good for 8% of the cases. But we

16:41have things that are a little bit more

16:43structured like declarative UI. Like if

16:44you know JSON render or

16:47things like that where the app can just

16:50declare the the structure of the UI, but

16:52the components are being rendered by the

16:54host. So, the host and the app are

16:56sharing the UI functionality and

16:59visibility. That's good for hosts that

17:01want to control the look and feel of the

17:04apps. For example, just imagine Claude

17:06probably doesn't want to have a Booking

17:07UI then an Airbnb UI then an Expedia UI

17:10in the same chat flow, right? So, this

17:11is pretty pretty good middle ground. And

17:14in the other hand, you have the fully

17:15generative UI, which is what

17:18Claude Anthropic released a few weeks

17:20ago

17:21where the model just generates the UI

17:23out of thin air. Now,

17:25the nice thing is that MCP apps is

17:27really agnostic to how you generate the

17:29UI. MCP apps doesn't assume that Airbnb

17:32created created the UI.

17:34Any any part of this process can create

17:36the UI and the feature that Claude

17:38released, which is the generative UI on

17:40the fly, actually uses MCP apps under

17:42the hood, right? So, it's a generative

17:45UI that's being streamed into an MCP app

17:47and then MCP apps closes that closes

17:49that loop. So, it's good for third-party

17:52UI, which is the black box, but also

17:53first-party UI. So, that's

17:55that's something that we're working on

17:57standardizing. Um we're doing a lot of

18:00work to do interoperability with other

18:02UI protocols like A2UI, which is the

18:05generative UI protocol by Google, WebMCP

18:07like we said, and we just want to build

18:10a unified standard for UI in chat apps.

18:13Um yeah, and that's a like a cool

18:16summary about MCP apps. If you build an

18:18MCP app, it runs everywhere. LibreChat

18:19is an MCP MCP app client. ChatGPT is a

18:23ChatGPT app client, but the same

18:25application works for the same codebase

18:27works for every

18:29every host.

18:33Cool.

18:34So,

18:36if you think about it, this isn't just

18:38some tech, right? This isn't some

18:39protocol. This is a new way to

18:41distribute applications.

18:43>> [clears throat]

18:43>> So, if you look at just a few months

18:45back, Sam Altman said, I think it was

18:47October, that 800 million people are

18:49using ChatGPT on a weekly basis. That's

18:5210% of the world's population.

18:54It's insane. The internet took like 13

18:57years to get to that number of users.

18:59And

19:01now, it's not even 800 million. It's a

19:03billion. And it's not just ChatGPT. It's

19:05also Claude and VS Code.

19:07You have a potential audience that is at

19:09least 160 times the number of users that

19:12iPhone had when the App Store launched.

19:15So, how do you get started? There are

19:18two main ways. As a server, like if

19:20you're developing an app. So, like I

19:22said, you go to XApps repo. There's a

19:24key QR code if you want to do it

19:25quickly. You have the skills. Just use

19:28that.

19:29The other way is in details. The other

19:33way is that if you're a host, so if

19:35you're building an application that

19:36actually hosts application, you can just

19:40take MCP UIs SDK, which is the

19:42recommended client SDK. It's also just

19:44fully compliant with the spec. You just

19:46take that React component and you're

19:47done. It just supports apps out of the

19:49box and then get hundreds of apps from

19:52booking and other providers out of the

19:54box.

19:55Um so just just to emphasize

19:59there was a slide about skills. So it's

20:00really easy to create an MCP app. Just

20:02if you visit the site we just pass

20:04through it but it's just a skill. You

20:05just push it to cloud code and you

20:06generate an MCP app out of thin air. If

20:09you want to get involved in the spec

20:10itself in how MCP apps are going to

20:13operate and if you want to help build

20:15the future of UI in in agents

20:18then obviously visit the official MCP

20:20apps repo that's X apps. Open an issue,

20:23open a PR, participate in the

20:24discussion. We also have the official

20:27discord for the MCP apps committee where

20:30we do surveys and we interact with the

20:32with the community and with other hosts

20:34to to decide on things that relate to

20:36MCP apps. And there's the community

20:38discord which is I think the coolest

20:39place to be because you have all of the

20:40users of MCP apps be it

20:43people that build servers, companies

20:45that build hosts that just talk to each

20:47other, share tips,

20:49troubleshooting,

20:50asking for features.

20:52That's the place to be if you're

20:52interested in MCP apps.

20:55Uh so we said some scary stuff along the

20:57way

20:58like the web is dying and there's like

21:00all the websites are meaningless at the

21:02point at this point but

21:03I kind of hope that you don't look at

21:05this as a threat but more as an

21:07opportunity like basically a once in a

21:1020 years opportunity to think about your

21:12apps again and think what is the core

21:15user experience that we're looking to

21:17get and imagine it not as a monolithic

21:20single app where people go to but

21:21actually a part of a new web of

21:23applications, these chunks of UI that

21:26allow you to communicate between each

21:27other using a smart model in between.

21:30That's that's pretty insane.

21:32And with MCP apps even this early in the

21:35ecosystem in this early in how agentic

21:37apps work we already have

21:39standardization and they all work the

21:41same. You can write your app once and it

21:43will run everywhere.

21:45So what does the future look like?

21:48We're not yet at Jarvis

21:50but with MCP in general and MCP apps in

21:52particular you can bring experiences

21:54that were impossible just a few months

21:56ago to every host in the world including

21:59your own.

22:00So thank you. Thank you very much.

22:03>> [applause]

22:09[music]