MCP Apps: Extending the Frontier — Ido Salomon & Liad Yosef

0:01[music]

0:12>> Hi. So, hi everyone.

0:15We built this talk yesterday, so it

0:16might be out of date.

0:18I'm Ido Sadan, I am the creator of MCPY

0:21and co-creator and maintainer of MCP

0:24apps in the MCP steering committee. I

0:26also created Adam Craft if you were in

0:27the talk yesterday.

0:30>> I'm the Adi. I work with Ido on MCPY.

0:32I'm also the co-creator and maintainer

0:33of the MCP apps spec and recently

0:36co-founded Aura, which is a research lab

0:38for the agentic web. And we're going to

0:40talk a little bit more about it later.

0:45>> So, MCP apps are all around us. You

0:47might not even realize it, but all the

0:50fancy apps you have today in ChatGPT and

0:52VS Code and Slack are actually all based

0:55on MCP and the MCP app spec.

1:01>> And if we take a step back and we ask,

1:03why do we need MCP apps? What's the idea

1:05behind MCPY or MCP apps? So, when we

1:08work with chats, when we chat client

1:10clients, we used to text because that's

1:12the natural interface, but text is

1:15really the worst way to convey a lot of

1:17information, right? Because we don't

1:19want walls of text. And actually, this

1:21is the main blocker from companies to

1:24build an MCP server. They don't want to

1:26be reduced to a textual database. They

1:27don't want to lose their brand identity

1:29in the process. They don't want their

1:31data that they work so hard on um

1:34building the UX for to look something

1:36like this.

1:37So,

1:39instead of this,

1:42what if the apps could just send their

1:44UI to the chat, right? What if every

1:47service and every brand could just send

1:49their user interface to the chat? So,

1:51instead of us looking at something like

1:53this,

1:54we could just have the apps send their

1:57own identity, their own UI chunks into

1:59the chat, and then we take a look and we

2:01see, "Okay, yeah, I know this is Shopify

2:03in the middle. I know this is Hugging

2:04Face. I know this is Monday." And what

2:07if we don't want to do it only as a

2:09visualization? We also want to do it

2:11interactive. So, we want the users to be

2:13able to actually interact with Hugging

2:15Face, for example.

2:18And for Hugging Face to actually do

2:20something with it.

2:22>> So, we don't have to imagine the future

2:24as we said

2:26with MCPUI, which I created in May last

2:30year,

2:31and took that, which is essentially like

2:33an open protocol for interactive

2:35applications over MCP. So, it's not only

2:39how you transmit UI, but also how that

2:41UI, that application connect connect

2:44communicates with the host.

2:46And just a few months ago, we partnered

2:49with Anthropic and OpenAI to create

2:52the official extension to MCP, which we

2:55call MCP apps based on MCPUI, MCP SDK,

2:59and other solutions in the field. Their

3:02launch was pretty cool with Claude and

3:04VS Code supporting it to begin with, but

3:06now

3:08obviously also OpenAI and others have

3:10adopted it.

3:13>> Yeah, and there are a lot of early

3:14adopters to MCPUI.

3:1711 Labs, Shopify, Postman. Those were

3:20one of the first first companies to

3:22support it back like a year ago. They

3:24were the one believing in this spec, in

3:25this vision.

3:27And Goose also supported it. And it's a

3:30it's a funny anecdote because today

3:32Block released their agentic commerce

3:34solution that is based on MCP apps. So,

3:36a year ago Goose was the first client to

3:38support MCPUI, and now it is part of

3:40Block's

3:42product uh

3:44product. And today we have a lot more

3:47clients that are supporting MCPUI.

3:49We We Cursor, and we have

3:52co-pilot and

3:54GitHub ChatGPT support MCP apps. ChatGPT

3:58apps that you know are actually based on

4:00MCP apps and open eye actually recommend

4:03using

4:04MCP apps as the protocol to build

4:06ChatGPT apps.

4:08Postman

4:09and a lot more and obviously Cloud

4:10supports MCP apps. But we also have a

4:13lot large community around it, right? So

4:15people start to to build plugins to MCP

4:18apps and

4:19um integrations to different agents and

4:22also courses on how to build MCP apps.

4:24This is by integration for MCP apps.

4:26So we have a lot community around it. Um

4:29There's a repo X app which is the repo

4:31for MCP apps where everyone can just

4:34come and propose PRs and ideas of how to

4:37how to extend this spec and we have a

4:39work group in the MCP committee and

4:41we're convening every 3 weeks. We have a

4:43tri-weekly meeting on the future of the

4:45protocol and how to make the spec not

4:48just serve the bigger apps but also the

4:51community. So it's an open working group

4:52with Anthropic, Open AI and all the

4:55partners in in the MCP apps protocol.

4:58>> Okay, so let's look at a few of the core

5:00concepts of MCP apps.

5:03The first and most obvious one is how do

5:05we even transmit UI over MCP? So if we

5:09look at this example of Cloud like

5:12you know agent times like a few months

5:13ago

5:14and I would ask something

5:16best case scenario it would reach out to

5:19my MCP server and it would get back a

5:22textual response which is obviously

5:25suboptimal.

5:26So let's say I do want to get

5:28some something better.

5:31So now I can use existing MCP primitives

5:34like a resource and now return HTML. And

5:38I can take that HTML and since Cloud

5:41supports MCP apps it can turn it into an

5:44interactive application

5:46of the best soundtrack in the world.

5:50And what if you wanted to be really

5:52interactive, right? This is nice because

5:54it shows the best soundtrack in the

5:55world. What if I want to favorite one of

5:57the songs there? I want interaction. I

6:00want communication between the app

6:03and the host. So, when the user clicks

6:05on the favorite button,

6:07MCP apps actually standardizes this

6:09flow. So, instead of the app sending a

6:11message to to the backend, to Spotify's

6:13backend, it's actually sending a message

6:16to the host saying, "Hey, user clicked a

6:18button. Do something with it. I

6:20recommend you to call a tool in

6:22Spotify's MCP server." And the host

6:24decides what to do. The host keeps this

6:26control of the flow. In this case, the

6:27host can decide to actually call the

6:29favorite favorite tool. And MCP apps

6:31standardizes this flow.

6:35>> Okay, so seeing is believing. So, let's

6:37see an example from Claude.

6:42Yeah.

6:43Uh so, let's say that I'm a product

6:45manager to understand the status of my

6:47funnel. So, I would go to Claude and I

6:49would ask what's the status?

6:51In the again, old world of a few months

6:53ago, uh I would get back the textual

6:55response. Let's say that it's PostHog.

6:58So, it reached out to the PostHog

6:59server, got back the textual response.

7:01It's factually correct, but it's

7:03useless. I mean,

7:05how do I even take that and understand

7:07quickly what's going on? I would have to

7:08read, which I don't want to do. Uh and

7:10it's pretty challenging. Uh but luckily,

7:13because both PostHog server and Claude

7:17as a host support MCP apps, I can just

7:20say, "Show me."

7:21And now, instead of getting that block

7:24of text, I can actually get something

7:25useful, uh which is this interactive um

7:29widget that you would get, you know, on

7:31the PostHog uh uh server. And when you

7:34have that, you can at a glance see

7:35what's going on. And as you can see,

7:37it's branded PostHog. So, you're

7:39actually getting the PostHog experience

7:41within ChatGPT or Claude, etc.

7:44Uh but it doesn't really end there. As

7:47we said, MCP apps is also like an

7:48interactive photo call. So, not only can

7:51I see and and interact with it, I can

7:53also do stuff like

7:55ask him to explain what a funnel is. I

7:58might not even know that. So, again,

7:59instead of getting that huge wall of

8:01text explaining what a funnel is, I can

8:03just get this generative UI answer from

8:07Claude, which uses MCP apps. It streams

8:09like the HTML inside, and now I can get

8:11this nice interactive experience of

8:14learning.

8:15And not only is it visually nice and

8:18helps me understand, but it's also

8:21fully interactive. And when we say

8:23interactive, it actually means that

8:25clicking it would help me communicate

8:28with the host. So, let's say that I want

8:30to understand like a particular step in

8:32the funnel. Uh I just go and I click on

8:35it, and since it's an MCP app, it can

8:37send a prompt back to the uh model and

8:40say, "Okay, explain this specific step

8:43to me." And I can advance the flow.

8:46Uh so,

8:47this is a like an example of of how that

8:50uh looks. So, how does it actually work?

8:52If you look at the architecture of it,

8:54uh so, we started by prompting.

8:57So, we type something in. Uh we asked

8:59for the funnel information. A tool call

9:01went out. Since our server supports MCP

9:04apps, that tool call is actually linked

9:07to a resource. And if you look at the uh

9:10code here, then, you know, it's it's a

9:12it's just a resource with a

9:14uh some prefix. Uh we take that. It's

9:17pretty simple code. I could just add the

9:18but it's still the the resource with the

9:20HTML, and you're done.

9:21Uh that resource is then

9:24um consumed by the host.

9:27In practice, it's usually consumed

9:28beforehand, like it's preloaded. Uh but

9:31imagine that it's just consumed in real

9:32time. That same HTML then passed to the

9:35host that also supports MCP apps. MCP

9:38apps basically if you look at the MCP UI

9:40SDK, just a React component or a web

9:43component that just accepts that

9:44resource plus a callback which is how we

9:47implement that communication protocol as

9:50I said earlier.

9:51And renders it in a sandbox.

9:54So,

9:55like we said, not only is it

9:56presentational, I can click. So, what

9:58happens when I click? So, we click on

10:00it, it sends back through that callback

10:02the event all the way up. The model

10:04takes that event and then it can send

10:06out a tool call

10:08or

10:08call a resource or anything else that's

10:10completing the agentic flow.

10:14>> And this architecture actually brings a

10:16new philosophy or a new vision to the

10:17web. So, instead of us thinking of the

10:20web as tabs or

10:23services that we need to consume using a

10:26browser, we're now consuming it using

10:28our own personal assistants, right? What

10:30does it mean? It means that if I want to

10:32accomplish a task, for example, plan a

10:35um anniversary. So, up until now I had

10:37to open 20 tabs in the browser and I had

10:39to try to convey my intent to each of

10:42those services. And by saying conveying

10:44my intent, it means that I have to

10:45interact with the dashboards or the UIs

10:48of those companies. So, just to plan an

10:50anniversary, I need to convey my intent

10:53to Google Calendar and Amazon and

10:55Booking and Booking again and Amazon

10:57again and all and I don't need 99% of

11:01the UI that is shown there because this

11:02UI doesn't know me. It doesn't have the

11:03context on me. What if we could just

11:06take these UIs and just break them into

11:09atoms?

11:10And those atoms can be composed by my

11:12own personal assistant, right? Because I

11:14don't need

11:15the the UI. I need those atoms. So, if

11:18we can take these atoms and have my

11:21cloud or ChatGPT or OpenCloud just use

11:24them using MCP UI,

11:26we can have this flow. So, my proactive

11:29assistant can say, "Yeah, I know. I see

11:31that you have an anniversary coming and

11:33instead of just showing me data from

11:34Google Calendar, it can display a Google

11:36Calendar chunk. Now, for me it's good

11:38because I know Google Calendar, I trust

11:40Google. For Google it's good because it

11:41maintains their brand and identity and

11:44for the host it's good because they

11:46don't need to develop these capability

11:47themselves. And it goes even deeper

11:49because if I'm interacting with Amazon,

11:51instead of Amazon being reduced to just

11:53a list of items or or text,

11:56I can see Amazon. I can I can know that

11:58this is this is Amazon and I can

12:01complete my entire flow without even

12:03leaving my assistant. And this is the

12:06agentic web. This is how we're going to

12:08consume the web because my assistant

12:11will have the context on me. It It will

12:13know to pull the the map from

12:14booking.com. I don't need to know that,

12:16right? So, this is going to be the shift

12:18that we're going to see very soon where

12:20websites are going to shift into small

12:22chunks of UIs inside inside personal

12:24assistants. Um [snorts] and with that

12:26come new interaction mindset because um

12:29if I click on something in the Shopify's

12:32MCP app, then Shopify doesn't control my

12:34journey anymore. The host does. Um and

12:37no application will control the user

12:39journey anymore. So, Amazon won't be

12:41able to know to see my flow. It

12:43everything will go through the chat for

12:45auditability.

12:46Um and MCP apps actually standardizes it

12:48by defining this three level of control

12:52over the user journey. So, an app can

12:55notify the chat that something happened

12:56or an app can actually ask the chat to

12:58run a prompt and and releasing all

13:01responsibility to the chat. So, MCP apps

13:03actually standardizes it and this is the

13:05new software flow, the new flow of

13:07interaction that we're going to see

13:08between applications, the chats, and the

13:11users.

13:13Um in 2026, we had we had an amazing

13:16year of standardizing MCP UI and 2026 is

13:19going to be the year where it's going to

13:20be a global standard for UI.

13:23>> Yeah.

13:23>> But, it's still evolving. There's a lot

13:25of stuff going on. Even in these past

13:27few months, these are some of the things

13:29that are already in or already

13:31contributed or proposed uh, by the

13:33community. Uh, so you still have a lot

13:35of time and a lot of room to influence

13:37how this future will look like. Uh, so

13:40you can go to X apps. Uh, that's the

13:42official SDK and spec is also hosted

13:44there. It's under the official model

13:46context protocol uh, repository. There

13:48will be a QR code later. Uh, so you

13:50don't have to

13:51uh, to uh, photograph it. Uh, and also

13:55um,

13:56the the cool thing about using X apps in

13:58particular is that because it's

14:00maintained by us directly, uh, all

14:02changes to the spec are immediately uh,

14:04reflected in the SDK. So, if you use

14:06that SDK, then you automatically get all

14:08the new stuff out of the back.

14:10Uh, these are some of the issues that we

14:13have. So, please feel free to come and

14:15uh, contribute. So, what's next? Um,

14:17there's a bunch of stuff coming up. Uh,

14:20the first thing uh, that we get a lot of

14:23uh, uh, of asked for is kind of reusable

14:26views. So, if you have

14:28uh, um, companies like Autodesk that

14:31have really heavy apps like they have

14:33in, you know, the entire 3D render

14:35there. They don't want to keep

14:37re-rendering that over and over again

14:40because it just it takes time, it's

14:41inefficient. Uh, that is the way that we

14:43had to do it uh, for the MVP. But we are

14:46working on thinking of maybe we can pass

14:49some identifier from the server uh, in a

14:52way that would help the model actually

14:54keep updating the same view.

14:56Uh, the other way to do this is

14:59>> Um, app tools, which is something uh, if

15:02you've heard of web MCP, which is Google

15:04standard of how agents will interact in

15:06with web views. So, in MCP us, we

15:08actually standardize it into app tools.

15:10So,

15:11up until now we saw the flow where users

15:14does something in the app and the app

15:16talks to the host. But what if the host

15:18or the chat wants to speak to the app?

15:20If the user writes something, uh, fill

15:22out this form for me and the chat will

15:24fill out the form for the user. So, MCP

15:26apps actually standardizes this this

15:28flow which we call view tools. That's

15:29actually that's in the spec right now.

15:31It's also it's going to be released very

15:33soon.

15:34And we're working on this generative UI

15:37spectrum where you have predefined UI.

15:38That's MCP apps. That's like the black

15:40box iframe that renders

15:43all trace UI in that example. But you

15:45also have other things on this spectrum

15:47like declarative UI like JSON render or

15:50A2UI. These specs that say yeah, the the

15:53app just returns an instructions on how

15:55to build the UI, but the chat will

15:57actually build the UI. And you have

15:58fully generative UI on the other end of

16:00the spectrum. And if you know cloud

16:02apps, yeah, MCP apps is agnostic to the

16:04way the UI is generated. And if you know

16:06cloud apps imagine feature where you can

16:09just ask cloud to generate a UI for you.

16:12That's actually based on MCP apps. So,

16:14this is an MCP app behind the scenes,

16:16but it supports generative UI. So, we're

16:19working on interoperability with those

16:20other

16:21standards. And actually just a few days

16:24ago we released a guide on how to do

16:28A2UI versus a generative UI standard and

16:31MCP apps which is the standard. How to

16:33do interoperability. How can a server

16:35can write A2UI and ship it to Gemini,

16:38but also wrap it as an MCP app to ship

16:39to ChatGPT and vice versa.

16:42An MCP app is supported everywhere. So,

16:45it can run everywhere. If you build it

16:46once, it runs in Libra Chat which is an

16:48open source

16:49MCP app supported in ChatGPT. That's the

16:52same app that you're seeing the same

16:54code base that runs in in both which is

16:57pretty cool. Yeah.

17:00Yeah.

17:01>> So, this isn't just a technology or a

17:04cool feature. This is an entirely new

17:05way to distribute applications. So,

17:09if you look just a few months back then

17:10someone said that

17:12ChatGPT in particular has 800 million

17:15weekly users which is 10% of the entire

17:19world population. That's insane.

17:21So, if you think about the web in

17:23general, it took around 13 years to get

17:26to that number of users.

17:27So,

17:29if you look at that and you think that

17:30in the last few months we actually had a

17:33growth of over 1 billion. Just that we

17:35have like 170 times the total

17:37addressable market of the Apple App

17:39Store when it launched.

17:42So, MCP apps are everywhere.

17:43So, actually to list them, it is called

17:45cloud, open AI, etc. It's already there.

17:49So, how do you get started?

17:51You can clone those you can

17:53go to the X apps.

17:55As a host also go to X apps or the MCP

17:57website.

17:59>> Please visit the official repo.

18:01>> [laughter]

18:01>> The X apps repo to get involved.

18:04And yeah.

18:05>> So, embrace the new web. It's awesome.

18:08With MCP apps you can write once and run

18:10it everywhere.

18:11And the future is looking bright. Not

18:13quite Travis, but with MCP and MCP apps

18:16we're close.

18:17>> And come talk to us afterwards.

18:18>> Yeah, thank you.

18:36>> [music]