Building ChatGPT and MCP Apps with All the Comfort of Modern TypeScript DevX

Hello, everyone. My name is Fred. I'm the CTO and co-founder of Alpik, the MCP hosting company, and today, we're going to learn more together about building JPG and MCP apps with modern development experience. Just a quick word about Alpik, we're basically a one-stop shop to build MCP apps, where we'll help you with our own framework, Skybridge, which we'll talk about a little bit later. We'll allow you to deploy on our infrastructure platform, monitor every usage of your MCP app to understand patterns and what our users are actually asking, and then to distribute across stores, especially on GPT and Cloud.

For those of you who haven't been around, following what MCP evolved into for the past few months, past years even, let's do a quick recap on where we are today. MCP was initially released back in November 2024 from Alpik. It quickly became adopted as a context enrichment protocol for the developer ecosystem, where devs would connect local server to their existing application, like Sentry for backtracking or AWS for infrastructure deployment. It grew in size, beginning of 2025, with additional support from many general clients, in addition to Entropiq Cloud, like OpenAI, Challenge GPT, Google, Mistral, with Russia, all in support for MCP, and brought their own MCP client into the generic chat.

But it really grew and went outside of the developer ecosystem when Challenge GPT announced back in October 2025, the release of Challenge GPT apps, a new ecosystem, a new platform of business built experiences, on top of MCP with the addition of UI in conversation, and a store catalog to browse across various applications available on the platform. The initial release back in October 2025 was made with less than 10 initial partners, like Booking or Expedia or Uber. But the adoption grew, and the quantity of apps is now exploding in the store. It was also extended outside of only the context of Challenge GPT.

The official MCP specs included an official extension called the app extension that was drafted in November last year and adopted at the beginning of this year, which opened the support of the same ecosystems that Challenge GPT introduced back in the days, back in October, to Cloud AI, and VS Code, and Goose, for example, that are now all compatible with these new specs. As we go further, more clients are adopting the protocol. And we see flourish a lot of ecosystem inside generative agents, conversational generalist agents. The two main ones, of course, are Challenge GPT, which is growing over 500 apps at the moment. Various from SaaS, productivity, travel, retail, e-commerce, insure tech, we have many different brands that are building the app and making them available on the Challenge GPT app store.

And the other one is, of course, Cloud Connector, which is close in terms of quantity of connectors, mostly from their heritage B2B and prosumer applications that dedicated to productivity. But recently, like two weeks ago, they announced a big push in the travel industry as well. So they are not that much lacking behind OpenAI Challenge GPT ecosystem. On our side at ARPIC, we've been busy, we haven't been there contemplating what Cloud and the Challenge GPT teams were shipping. We actually built our own fair share of applications that are currently available on the app store, on the Challenge GPT app store and the Cloud Connector store. And we learned a lot building those apps, which drive us into releasing an open source framework called Skybridge, which is a full stack TypeScript framework to build those new Challenge GPT or more generalistically speaking, MCP apps, which are compatible with the OpenAI Challenge GPT app ecosystem and Entropic Cloud Connector ecosystem.

And if you want to use Skybridge, and this will be the main focus of our conversation today, it's starting with a create Skybridge command from your terminal. The project is open source and will remain open source all the time. If you want to win one of those cool skin masks that we're starting to get known for, you can today. You just need to go on ARPIC AI Skybridge on GitHub. You can use this QR code if you want. If you prefer to go on the Skybridge repository, and give us a star. If you give us a star, you will enter a draft that we'll do at the end of the day, and somebody at random will be picked to receive one of those masks, who will contact you on GitHub straight away. So go ahead, connect to Skybridge and give us a star. This, of course, helps the project mature and grow, and also gives us confidence to continue working on it. Plus, we're sure it will help you build very easily your first app.

I've talked a lot about chat.gpt app, but let's dive in on a real example and real exploration together to learn a little bit more about what this ecosystem is all about. So here I am on chat.gpt. It will be slightly different from the one you're currently using, because I'm in developer mode, which allows me to use apps that are not yet published in the store. But everything I'm going to show you right now is accessible, just like it is here for you. You can see a new apps icon in the left side of your chat.gpt navigation menu. And when you go on it, you're basically on the homepage of the chat.gpt app store, where you can browse featured applications from various businesses or search for specific one.

For example, I like to look for Evanos, which is a travel company that helps me find a nice travel location and trips to plan for me and my family. But you can look for any applications that you might get used to using on the web or mobile, and find it on the chat.gpt app store. All applications that are connected, you can see this small connected icon. That means that I can use them in conversation. One that I'd like to show is the cottage application that we developed using Skybridge and that we're hosting on the Alpic infrastructure platform, which is the equivalent of Airbnb for the UK, countryside UK location. So I can basically ask to find a nice home for me and my family of two kids in Cornwall for next week. And I'm interacting with the app this way.

If the voice recognition feature from the chat.gpt works well, that would be much better. Otherwise, I can always cancel the dictation. And let's do that. So if I say cottage, I would like to go to Cornwall with my family of two kids next week. It's going to basically start conversation with chat.gpt, including cottage MCP app, which is an MCP server responsible for making the search of available cottages for next week for me and my family, plus some view components that are going to be rendered in conversation to provide a better experience than using, for example, a Markdown table that lists various options across various columns with detailed parameter. And this as a brand allowed me to keep controlling the way my products and services are exposed through chat.gpt to end users that might be looking for a location for their next vacation.

Those components are interactive. So I can browse, for example, here across the carousel, but you can implement anything that you want, except from carousel. And, for example, this is part of what we developed into cottage application. You can, when interacting with one specific card, go in full-screen mode. So I'm still on chat.gpt.com at the moment. But the app experience is taking the full available screen. I still have my prompt input at the bottom of the page. But otherwise, all of the experience that you're looking at here is powered by the app itself. So I can browse across various locations to see additional details. Of course, I can complete my checkout journey on the cottages.com site by itself. But what's nice is that I can keep on interacting with the UI that is displayed or switch back to conversation and say, finally, I want to go to Scotland instead, which are two ways to interact with apps that are available on the chat.gpt app store. And that you should always consider when you want to build this new type of apps in the ecosystem. It's not only you. It's not only your application and the user that are interacting together. You now have a three-body opportunity where the LLM is also a part of the interaction. And you always must account for people interacting with the UI or interacting with the LLM to ask specific questions. And there are many, many other apps in the chat.gpt ecosystem that you can discover. But for the remainder of this talk, we'll focus on what you need to know to start building your own and provide your product and services available in chat.gpt or inside Cloud Connectors, just like I showed you for cottages. The entry point that you should start with is the SkyBridge repositories that I was mentioning. Give us a star if you haven't already. Starting a new project is as easy as create SkyBridge with your favorite packet manager. Let's do just that. My favorite packet manager is pnpm. So I will do just that. And the name of the project is going to be just nation. First thing is going to prompt me to whether I want to package the coding agent skills that we developed inside SkyBridge together with the bootstrap project. And of course, we are going to do so because it's going to empower our coding agent with everything that it needs to know to be able to build an app without any problems. I proceed with installation. And now I'm going to install the SkyBridge project itself dependency and start the project. I'm downloading a few of them.

And here I am. So I get prompted with an information saying that my MCP server is running on localhost 3000 slash MCP and that I have a locally available dev tools to interact with my application. Let's check just that. So I will return on my navigator. Go on localhost 3000. And here I'm welcomed with this interface, which is a simple web app interface that allows me to get information about the app I'm currently working on.

So the demo application has a single tool called Magic 8 Ball that is a fortune telling tool for which I can ask a question. An answer will be provided by the server. And of course, using the official extension as your application extension from the MCP protocol. View of the Magic 8 Ball providing the answer to the question will be displayed. So here if I ask, am I going to do a demo effect and call tool, I will have the answer generated by my locally running MCP server using SkyBridge, plus the view of what will be generated in conversation to provide a better interface to the provided answer.

This dev tools are the go-to solution to iterate fast on your building lifecycle in order not to have an LLM already involved in the process, because it's much easier to control which would be the parameter of the tool calls that's reached your MCP server in order to have a look at the UI, other than trick an LLM into making the calls that you want him to do. So this is really the best interface that we want you to use to build those new kind of applications. Of course, this application keeps track of changes you make on your local running code. If I open the project I just generated, which is basically ts-nation, which would be around there. Okay. You can see the project structure with the source folder that contains the MCP server code itself with my Magic 8 Ball tool that handles specific components called Magic 8 Ball that you can find inside the views part of the source folder. If I change something, like for example, I don't want the answer to be the one provided by the server, but to be no, I can see the effect of this change straightaway inside my rendered UI. So every change I make on my MCP server during local development with SkyBridge can be inspected and controlled using this piece of dev tools, rather than connecting it to GPT, going through the trouble of making to GPT call my tool and check the impact of the modification I just made. I can switch back to the original answer and you can see here that it's switching back to the original answer that was provided by the server.

Okay, that's all nice and clear. But the Magic 8 Ball itself is not the thing that I want to do. I want to build a small capital explorer application. Let's do just that. Build a charge GPT app to explore capitals of your ops. I want this app to be a single tool that takes as input the name of the capital and outputs detailed information about the capital plus a view, which is a map with capitals represented as pins. The search capital is selected, meaning that the pin is highlighted. And whenever I interact with the map, the map goes full screen. When I click on the capital, it goes selected. The selected capital detailed information is displayed in the right side panel in the app.

Okay, let's see if this goes through without any issue. It's not having too much trouble understanding my English accent. But you see, straight away, it's using the GPT app builder skills that is packaged together with the project and will start making changes. It will interact with me to ask for confirmation about various changes I want to do. And this is really the easiest way to get started on building your own application. It's going to really know everything that you need to in order to build the best class experience of a charge GPT app.

Another thing I can show you in the remainder of time that we have together is what to do once you are finished building with your favorite coding agent, your charge GPT application. Things that you can do is expose your application so that you can render it outside of the dev tool inside charge GPT. In order to do that, we will go and is it did I kill my process? Yeah, looks like it. It will we will start a new process. It's okay. Which is basically skybridge. Now, pnpm skybridge, dev dash dash tunnel. Running this comment is still running somewhere. I don't know where. It's okay. So in addition to providing my dev tool and MCP app on my local machine, it's exposing it through a tunnel to the outside world, using a generated URL that is constant and connected to your Alpic account. You can go ahead and create your Alpic account if you haven't already. It's completely free of charge. And this URL, you can actually add it into charge GPT to start interacting with it.

So if I go back to charge GPT here and create a new application inside the apps and create apps, I'm able to use this URL as I've already done so, because I've already played along with apps that I'm building locally. But if I use this app inside any conversation, charge GPT, I will actually be interacting with my locally running skybridge instance. So if I say, will I do a demo effect, which was a question I was asking inside the dev tool, I should have charge GPT through the tunnel, call my locally running application, and we will see the application being rendered. If I make a change, once again, up to my application. So if I say no, it's not the answer generated by the server, but a constant, which is no. You can see the effect also reflected inside the conversation of charge GPT. So this is the second step of your building journey with Skybridge for charge GPT app.

You tested everything locally, the rendering of the UI is as per your wishes, the functionality looks fine. Now you want to see how it's going to behave while you're interacting with an LLM. The best way to do so is exposed with Alpik Tunnel, your Skybridge application, so that you can interact with it on charge GPT. And now that you have run locally your application, it's time to deploy it. It's as easy as running pnpm run deploy. This will go ahead and create a new project inside your Alpik account, where the app will be deployed. Because right now, the tunnel, if I stop exposing it, my session will end and nobody will be able to use this app. So you're satisfied with the first iteration that you've done with the skill and the application is as per your liking. You've tested it. Everything looks fine. Now you're ready to deploy. And you can do that in just a few clicks on Alpik. It's going to create a project and start the deployment.

Once the deployment is finished, you will have a demo environment, which we call the playground, which is the equivalent of charge GPT, that you can play with. I can show you quickly, for example, on one of the Skybridge example applications that we have, one, for example, for productivity. The playground is basically an emulation of charge GPT with your MCP app already connected that you can share with anybody. That allows you to test in condition your MCP application. So here, for example, is a productivity application that I can interact with, that is going to showcase the various functionality of what you can build with Skybridge. And of course, you can go through the submission process of charge GPT and cloud to make your app available in the store and usable by anybody. I hope this demonstration of what you can do on Skybridge helps you and encourages you to start building. We have an open discord that you can join if you have any questions about Skybridge. Once again, don't forget to give us a star. This helps us grow as a community around charge GPT and MCP app. And I wish you a wonderful rest of GS Nation. Bye bye.