How to Build MCP Servers

Welcome to Syntax. Today, we had an episode on building MCP servers. We talked about MCP, I don't know, maybe a couple months ago, and so much has changed since that episode that I thought we need to do a brand new episode on explaining what it is, cleaning all the updates to it, and kind of diving into all of the different pieces and tools and and helpful things around it. MCP is the model context protocol.

It's for exposing tools to AI. We'll talk a little bit more about that, but it's gonna be a great episode. You're certainly not gonna wanna miss this one. What's up, Scott?

Oh, hey, man.

I'm sick.

All this stuff. Aw. There's there's been somebody jackhammering I wanna say this. Somebody's jackhammering a patio for, like, two straight days now, and this is the finally, the first little window where there's been some silence. I mean, they were jackhammering at, like, 6PM. We were gonna eat Node outside, and it was still going on.

I don't understand how you could spend that much time jackhammering a patio. That's unreasonable time to jackhammer. Yeah. Yeah. You can't do that. You can't when people are trying to eat dinner and you're jackhammering.

So yeah. Node. Not a fan of that. And, yeah. So my besides that, I've just been, like, banging my head against some code stuff, and I'm sure the jackhammering has not been, like, helping my brain function today. Yeah. So that's where I'm at. Yeah. One of those days. Right? Well, I'm I'm gonna fill you with MCP

nuggets, and I I feel like you'll be you'll be back to to your bright self by the end of this.

Love that. Yes.

So alright. Let's let's get in on into it. So MCP, it's a standard for being able to surface tools to AI.

So whether you are using, like, an IDE editor like Versus Node or Cursor or whether you're using a chat application, like Claw Desktop or whether you are building your own thing with with AI libraries, like the OpenAI TypeScript library.

All of those things have the ability to surface additional tools to it via something called MCP.

And I'll I'll give you an example of of where I ran into it recently. I I built my second MCP server, a couple days ago, so I'm feeling feeling fresh on that. And I used a lot of the the new features that have come out since we did the first episode on it. So the application that I was building was sort of like this little scratch pad that will will do calculations for you. So I find myself often doing things like just like writing notes in a text editor and then having to do quick little math. You know? Like, this thing costs a $100, and then you add on tax and shipping, and then I'll divide it by there's two of them. You divide by two.

Or alright. We have to record six episodes, and and that's this many hours on average. They're between this and this. And I find myself just having to, like, copy paste that into a calculator or something like that. And I'm like, I want, like, a natural language UI for doing currency conversion and and and addition and and whatnot.

So there's a couple apps out there that do it, but none of them have are using AI. And I thought, AI is so good at natural language processing. So I I built it up, and it worked really Wes, except there were two problems. First of all, AI sucks at math.

It was just imagining Yeah. Totals up. Like, some of them were correct, and then some of them were just like, no. That's that's not right. That's that's not right at all. Or I would do things like t shirt vendor one has a $50 setup.

And then at a 100 t shirts, it would cost $7 per t shirt. And at 500 t shirts, it it would cost $5.55.

And then vendor number two, I would go ahead and give all the Scott. Which one is the better deal? And it would go ahead and do all the math for me, find the averages, and then it would tell me at a hundred and five hundred T shirts, which was the better vendor.

And it it sounds like a math problem from, like, a textbook, but it was just imagining the actual answers. You know? And I was like, this is kinda frustrating. Be so what you have to do in these situations where AI doesn't necessarily excel at doing these things is you have to give it tools. Right? And one thing that AI sucks at is math. So I implemented something called tool calling, which is not MCP. I'll talk about the differences later on in the show. And, basically, you give it access to a tool that can do math. So I exposed it. It's called MathJS library. And then what it does is anytime it needs to to do math, it simply just returns the function call that needs to be run. It runs that function, then it passes it back to the a with the actual answer. So we're actually, like, evaluating the math Scott with AI, but with with a computer.

So that worked really well. And then then the next step I needed was, like, I needed, like, currency conversion. You know? And but I needed, like, real up to date rates.

And the AI doesn't necessarily know what the the currency rates are. So what I did is I created a MCP server that will expose a Vercel tools for for getting rates, for converting currency, for getting historical rates.

So when the AI realizes it needs to fetch a rate or do some do some conversions, it can simply use the tools that I have supplied to it via MCP. So here I have this little MCP jam, which is a yeah? Cool. Yes. Answer or Wes question there, student Scott. What I don't wanna do I don't wanna interrupt you. I how does the AI know when

to use one of these specific tools? Yes. Are you having to pass specific keywords, or is context enough?

Yeah. So when you're loading a tool, when you're defining a tool, you you give it a description. Actually, hilariously enough, I was just you just asked me that question, and I went to show you my description. I don't even have descriptions for the four tools. But if you go into, like, the Century, MCP and load all of their tools, you can see here that there's, like, a find teams tool here, and it says finds teams in organization Century. Use this tool when you need to view all teams in a Sentry organization.

Find the team's slug to aid in other tool requests. So that both the name of it, obviously, was enough for me, as well as the description will be made available to the AI, and then the AI will be able to decide when it wants to use a specific tool,

and wants to run it. Okay. I've always found that it will be a little bit interesting with MCP. Maybe it's gotten easier to always get the the right call. I know this is calling tools.

Yeah. Yeah. Because you I think what what can happen is if when you have too many, tools, you just, like, load it up and give it a bazillion, It won't necessarily know which one

to run. Too many tools.

Syntax the syntax, dot f m team.

Too many tools. Yes.

So that was the, MCP server that I made, and I obviously have a couple other MCP servers here. We got the CloudFlare Worker MCP server. What this will do is it you authenticate, which is a new part of the MCP spec that we did not cover last time. So the ability to use OAuth flows inside of the MCP server is now possible, meaning that you can hook it up to your CloudFlare or your Sentry or whatever.

And then when you try to use that tool, you have to first do that whole, do you wanna allow this application? You you click yes, and then it it will then connect it to the thing that you have. So authentication is something that is relatively new as well as the transport protocols. So there's two ways that MCP currently works. The first one is via standard IO, meaning that the MCP server is going to be running locally on your machine or on the same machine that the AI process is running. So if you have, like, a Node script or something that's that's handling incoming requests, you might also have a MCP server running that will be able to input and output data. So that's that's standard IO. And then the other way that MCP works is via HTTP and server sent events, which is more probably what people listening to this podcast are are going to wanna use because we're we're web developers. Right? You wanna be able to fire up a server. You wanna have URL endpoints, and then people can connect to it. So that's generally how I'm connecting to all of these. You'll see I have this Node, currency-mcp.westboss.com.

When it connects to that, you can list all of the different tools that are available as well as no. There we go. They're they're listed.

As well as some of the other things like resources and prompts, I'll talk about in just just a second. So HTTP MCP servers will either return one of two things. First of all, you you send a post request to it, and then what is returned is either going to be just a JSON payload, like, give me a list of all of your tools, or let me go ahead and use one of these tools. So, for example, get latest rates, Go ahead and run that tool. It's gonna give me a list of all the the rates, but you can also change, like, the Bos currency.

Go ahead and rerun that, and it will give me the Canadian rates.

So that will either return just JSON back or it will give you a stream which uses server sent events.

So it will either return immediately or a stream that that then can send data as it's made available over server sent events.

Server sent events are kind of a a interesting one because it requires your server to be long running.

You know what's not very good at being long running computers? Serverless.

Yes. Correct. Yeah. Functions. Turns out Wes kinda need servers. Yeah. So the solution to that is, CloudFlare uses durable objects. Vercel uses fluid compute.

Mhmm. Basically, anyone who is offering hosting of MCP servers will either be running it on, like, a long term storage or they'll have some sort of other way, to get around that via durable objects or or fluid compute.

This tool, that I'm using or if you want to be able to explore an MCP, you can use what's called the MCP inspector, and that will allow you to connect to an m p MCP server, and you can view all of the things that it has available to you. So, tools is probably the the most popular one. Right? Like, here's an example MCP server that will surface a whole bunch of of different tools, you know, add, print environmental variables, annotate message.

But there's also things, like resources, which are allowing you to supply files, to an AI, which is really cool. So last time Wes talked about MCPI said the server that it built was something that would go off and download an SVG icon and then Yes. Bring it back to the AI because that's something the AI sucks at. It sucks at getting icons. You know? Like, put a Chrome icon here. So I I hooked it up to an API that went ahead and downloaded the SVG icon and brought it back. But at the time, not many of these apps supported resources.

So there was actually no way to give the AI the actual SVG icon. So what I did is I told it just to, like, run a a cat command and and pipe it into to a SVG. Now you can give a you can properly give it a file or an image or or any other, type of resource that you would be expecting. So if you say, like, go off and download six sound effects of cats meowing, you could write an MCP server that would that could do that. That is, something I would like to do. I think,

if if I'm reaching for AI, that's the type of thing typically that I might wanna do. I wanna have some cats meowing. One one time, at I used to work at a a a record label, Ghostly International, for people who might have heard of it. We each used to get like, there would be a rotation of who got to pick the album that we'd be listening to. And there's, like, six people in the office, and these, mind you, are very into music people. I'm very into music person myself, but these guys were very serious. And it was often like, I'm gonna put on this album that is, like, touching my soul right now. And, one time I put on The Beetle Barkers, which was just dogs barking in The Beatles, and I think they all wanted to fire me, because we listened to the whole thing, and they were so they were so mad about it by the end of the by the because they're like, they were gonna throw me out the window, I I swear, but I really liked it. That's hilarious.

Yeah.

Another part of the MCP spec is the ability to save reusable prompts.

And this is really cool because, like, you know how we have, like, cursor rules, you know, like, kinda things that you wanna tell the AI over and over again? Yeah.

It's not the exact same thing, but the idea of a server being able to ship prompts along with other tools is is really neat and really useful. So for example, on CloudFlare Workers, if you load their prompts, you see they have workers prompt full, and you can go ahead and load that sucker. Basically, what what they're going to return is this is a massive prompt for working with Cloudflare workers. Right? It just says, like, use TypeScript unless explicitly told to use JavaScript, always add appropriate types and interfaces, use ES modules. So, like, that's kinda like cursor code code guidelines, but then they also have output formats, bunch of integrations, and then all kinds of different code examples that you could possibly have. Right? That's a massive prompt that's gonna help your your AI agent that's generating code to generate code that is much more in line with it.

So they provide those prompts. Because if you were to tell the AI agent, hey. Create a CloudFlare worker that does x, y, and z, chances are it's going to go ahead and grab the prompt for that CloudFlare worker because it now knows I'm building that, download it, and then put that as part of the the prompt before it goes off.

So this, like, MCP Jam, MCP Inspector, these are tools that are just for exploring it. This is not actually something that you would you would be using it directly.

Like I said earlier, the way that it generally would work is you would add these MCPs to your chat application.

You would add it to your your cursor or your Versus Node IDE, anything that interacts with ID. In my case, I was using the OpenAI TypeScript SDK.

And in order to surface this currency conversion, you simply just add it as a MCP. And then when the AI goes off and and runs it, it says, oh, that's neat. I have access to these tools. Right? And if you're using ChatGPT, you you can tell that it like, sometimes it just searches the web. Right? That's a tool that they have integrated into it, so it knows that it can go ahead and and do these things.

Actually, coding these tools let's talk about that Node, Why that what that actually looks like. The way that it works is you can fire up a, MCP server. This this one, in my case, is using the, MCP server from, Anthropic. They're the the makers of the spec.

And then in order to actually make it hostable and long running, I'm using the Cloudflare agents MCP that just extends that. But most of this code is going to be standard MCP pnpm.

And then you simply just when it fires up, you can add resources or prompts. In my case, I'm just adding tools.

So, convert currency or get latest rates, and then you give it a schema. Now the schema is really important because its schema it does several things. Right? It it tells the AI how to give it the information.

Right? Because the AI has to know how to structure the input. It does the validation of the data coming in.

So in my case, I am using Node, so generate latest race schema.

Right? You you can take the base, which is a currency Node, and you can go ahead to describe it.

And then you can give it different symbols.

Maybe if we take a look at the convert currency, There is a from Canadian to USD and amount.

And in order to specify that schema, I've, done it all out inside. You can use any anyone that you want, but you must validate that the data coming in from the AI is of the format that you expect it before you go ahead and and pass it through. And then you can also specify the output types. This is something that I have created, but I'm not using it yet because I have not yet found anything that can use these things. So the output schema is just a new part of the spec, which I have not found to be implemented anywhere yet. And if you want to see all of the errors in your application,

you'll want to check out Sentry at century.i0/syntax.

You don't want a production application out there that, well, you have no visibility into in case something is blowing up, and you might not even know it. So head on over to century.i0/syntax.

Again, we've been using this tool for a long time, and it totally rules. Alright.

So you're calling tools, which we said at one point isn't MCP. Yarn is is tool calling an MCP related, similar, the same? What what's the the lingo there? It it's kind of a confusing thing. So

before we had MCP, we had this idea that, like, the OpenAI docs, you can give it you can have function calling JS is what they they they called it.

And then we have MCP, which kinda seems like the same thing but also isn't.

So the difference between the two, as far as I understand it, there's there's a lot of overlap between them. The tool calling is something that's going to happen on the same server that your your AI code is running. So for a very simple example is is I was I gave it a math tool calling. So it's able to do things like add and subtract and average and max the the different numbers, and it will and we'll be able to run those.

Whereas the MCP server can live somewhere else. The MCP server can live on the same machine as well, but it also it can live absolutely somewhere else. And, also, it seems to be a little bit more more stateful. You know? There's authentication.

There is long lived sessions that are streaming data back and forth.

There is lots of back and forth with the the actual server. So once you're logged in, then you can say, what are my things? So I I feel like the tool calling is for very simple functions that need to get run, whereas the MCP server is going to be for in in my example, I'm connecting to the Cloudflare workers once to get docs and be able to create databases. I'm connecting to my century m MCP to be able to get a list of my issues or to to kick off the their AI agent to actually fix the things for you. I've got my currency one, which is going off and fetching the currency rates and and able to, calculate them for me. So I feel like they have quite a bit of overlap, but the MCP can be remote, and the MCP is a standard specification.

So I would unless it's just a very simple function that I'm running, I would probably throw that off to a separate MCP server. One more thing. I've been talking about remote MCP servers, and and that's a relatively new thing to the spec. Mhmm. Meaning that, previously, your MCP server had to be running locally on the same machine that your your actual AI code is running, and and that's really helpful for something like Puppeteer.

Right? There's a Puppeteer MCP server that can has the ability to fire up a browser, navigate through the page, take screenshots, count, see if something's on the page. You know, like, anything that that you could normally do, you can sort of robot via an MCP server Wes you're basically when you fire it up, you're saying, hey. Here is and I think the the Puppeteer MCB server has about 40 or 45 tools, and you say, these are 45 things you can do in a browser.

Go nuts. You can you can do any of them. You can control the entire browser. Right? Like, that makes a lot of sense. You run that in a Docker container, and it runs on the same machine that you have. But then other other things like the the Sentry one, the Cloudflare Workers one,

that makes more sense to log in and be able to communicate back and forth because Yeah. They have all the info information. Right? Yeah. In fact, that's primarily how I've often used them. Now I have used, like, the Sentry one, but, you know, the the MCP server I've been using the most lately is the Godot one. And there's like I I will tell you, folks, if you're out there looking for good MCP servers, it's it's really the wild, wild west with a lot of them right now. There's, like, six or seven different Godot MCP servers, and I would say, one or two good ones. And even though two good one or two good ones vary pretty widely in terms of what they're able to do. So And how are you interfacing with that

that server?

Via cursor currently.

Okay. Because I I I what I what I'm doing is I'm working on a game, and Yeah. Godot has its own interface that also has, an API.

And the Godot MCP server allows it so that I don't have to work inside of Godot, but I can control the Godot interface via cursor if I want to automate certain things, create resources, update things through the UI, those types of things.

Yeah. That's that's great. I even think about, like, OBS, which we use quite a bit. OBS has a WebSocket API to do literally anything.

And I I built an app that allows us to start and and turn off recordings and record screens and do a whole bunch of, like, neat stuff. But I I often think like, wouldn't that be nice if it was just a MCP Vercel? Because then you can just you could just type to what you want. Or, like like, a lot of times, I look at at something like like the Cloudflare Workers one Wes let's take a look at the tools here.

OBS, I would like you to have these, settings, the settings that are Wes from YouTube. Give me the YouTube settings.

There we go. Find issues in the syntaxfmorg on on the Century MCP, and then it it just fires them all out. But, like, I often think like, it could probably do, like, generative UI. Like, if you just give it a list of all the possible things that you can do Mhmm. It could probably do a pretty good job at just building out a UI for you. And I built a whole UI for the the, OBS Node. And I like, don't you think it could do a if you were to have an MCP server of these are all the possible things you can do Yeah. It could probably generate a pretty neat Node demand UI Yeah. For you. Yeah. Especially because OBS is one of those things that sucks to work in. I don't know if it would make a better UI than than what it is. I don't think the AI will know better. No. I'm just saying, like designers. But You would have to interface with it less. Anyways, what I was saying about the remote MCP servers is that not all clients support them just yet because it is a relatively new this whole MCP thing is moving at, like, a breakneck speed right Node. And I imagine that this episode in in three months will be relatively out of date, and we'll have to have to do another one. But if you are using a client that doesn't support remote MCP servers yet, which is most of them, then what you'll have to do is you have to use this little Npm MCP remote. And what that does is it just fires up a little proxy server on your machine, and it will sort of translate them from one to another. It's it's a pretty common thing. Anytime, like, you see any documentation for for integrating

them, you'll you'll have to be using this MCP remote. I was gonna say most of the documentation for remote ones is gonna include that anyways for Yeah. Common providers.

Let's just rattle through a couple really handy ones as well. Obviously, I've been talking about the CloudFlare ones. Look at how many they have, though. Yeah. They have one for their docs, one for workers One one for worker builds.

That's kinda worker builds would be kinda cool because you could just chat, hey.

Why did my build fail? Can you explain to me why you would need one for the docs? Yeah. Documentation is so key because Yeah. But, like, why Node why would you need an MCP server for the docs themselves? Basically,

you're working on something. Right? Like, you're working on Cloudflare Workers. You're working on Svelte. You're working on whatever, and you need to be able to give the AI the the specific docs.

If the AI knows, oh, I'm gonna go grab the documentation for this specific library, it can go ahead it'll just simply fetch it, and then it can go through it and and pull pull the parts out that it needs and understand how it works. And that's that's huge. So you basically just have to throw a MCP server for the docs of what you're working on.

And then, like, the Cloudflare and Vercel both do an awesome job where they have all of their documentation as markdown.

And that's a great thing just to feed directly into an AI, but it's it's one step further where you don't even have to go copy paste it. You can simply just give it the tools like, hey.

Check the docs for x, y, and z. Or if it doesn't understand how something works, because so many of these things that are not like, it's AI has been is really good at writing Next. Js and Tailwind, but there's a lot of other stuff that doesn't have a lot of examples out there that it's not very good at writing. And it's it's hard when things change, like Svelte five.

It doesn't know about Svelte five Gotcha. Just yet. Right? Because the the the models haven't been trained on it. But if you can prompt it with a whole bunch of examples and and and TypeScript types and how these things work, then it's gonna do a much better job at at how that works.

There's this use MCP, which is a React hook, which is wild to me because you basically just take any MCP server, feed it into this React hook, and it will give you all of the methods and state for using it inside of our our React application, which is is really cool, which a lot of this feels like we're just recreating

Wes APIs. Yeah. Right. Yeah. Of course. Yeah. That is the the weird thing.

With the currency one, I felt like I was just hot potatoing a lot of the requests from from one to another. David Kramer had a really good tweet thread on it. Kenzie Dawes also replied. He he had a blog post on it. He says, like, the MCP is not really your API.

The MCP is more of a UI, which I thought was kind of an interesting way. It's a connecting thread.

Yeah. Exactly.

Instead of interfacing with your service or your API via buttons, you're interfacing the service via MCP. Right? Or or it's generally going to be something like chat. The the AI is going to be doing that. So in some cases, those will be the same thing.

You want your entire API to be available to the AI. But in in other cases, like, if you think about your UI, sometimes your UI will have two or three different API requests or rest requests, and then you'll need to do do things in a specific order. Yeah. And and surfacing it via MCP request is is more the the move there.

So that's that's all we got on the the MCP right now. There's this really great, GitHub here called awesome MCP servers you can kinda rip through. It seems it really feels like the early days of APIs.

In the early days, everybody had an API, and everybody was building stuff and remixing and connecting things together, and and it was really good. And over time, a lot of those APIs disappeared, partially because people don't want you building your own version of of Twitter or or whatever social network they want you or or Reddit or whatever, and and partially because, like, people use them to spam, and you can't give people programmatic tools for for pushing things. However, it does feel like everybody's coming out swinging with their own MCP server right now.

And it's kind of an exciting time to to see everybody like, there's a Telegram one and a welcome bot. Like, do we need a syntax MCP server?

What would that do? What would that do? What would that do?

You could get the latest show.

You could search for a like, again, we're just describing It's just the Node. Yeah. Yeah. Here. You're just giving it a different means of interfacing with it. Yeah. I think the the real power of a lot of these APIs is is both when you can pull content, but also when you can can do things. Right? Like, when you can start stuff or save things or modify.

Yeah. Especially things that have, like, a a UI because it gets you out of doing click ops for things. Yeah. But, also, I mean, you're getting into AI ops. But, like, yeah, if if you're clicking around a UI to trigger things and now you have a tool to interact with that via natural language, that seems interesting to me, or you're able to connect several tools together, you know, to work with each other. Yeah. Those seem like the biggest,

bonuses for me. Yeah. And I and probably the the one other one I'd add to that is is being able to get better results out of your out of your AI application. Because at at the end of the day, the whole reason I I stumbled into all of this is because I simply just needed better math and better currency conversion in the app that I was building.

And the so this Wes not like a chat app that I was using directly. It's just I just the app that I was building was not very good. And in order to make it better, you have to give the AI more tools.

Well, that's great. You Node, maybe we could have a MCP server for some sick pics Wes it could just send you right to wherever we we bought these things from. Sick picks are the things that we are enjoying. Could be anything, a podcast, could be just just about any product type of thing or or whatever.

Yeah. Sick picks. You know what? I will sick pick. This is, like, not really a unique sick pick in any regard, but, man, the new Mario Kart on Nintendo Switch two is just a lot of fun. I've only been playing with the kids, and we'll we'll find some time to to rip around that thing. And when they announced, like, giving, like, modes where it was, like, give, like, steering assistance and stuff like that, I didn't really understand it. But, like, now that I'm playing with a six year old, man, the steering assistant stuff really helps them have a good time. Because even in the beginner settings, like, yeah, they're learning how to play the game, and they're doing pretty decent, but they're finishing in the bottom half of the standings every time. And they have a lot less fun doing that. And then when you turn on the steering assist, they're finishing in second place and fourth place, and that just makes them, like, wanna get better at the game and not get discouraged. So I I've been really enjoying the game. I'm historically someone that plays Mario Kart games by picking the most obnoxious characters. Like, who who are the biggest, most annoying characters in the game? And let me tell you, this game has got some obnoxious characters.

So I'll be the cow. I'll be Chargin' Chuck. I'll be just, like, the goofiest characters because that's, that's how I like to play.

So, yeah, the new Mario Kart is just lovely, yeah, as you'd expect it to be.

That's good. I I downloaded or purchased Mario Kart for the Wii, and That's I didn't realize that. There's, like, a there's, like, a hacked version of Mario Kart called Mario Kart Black.

And, like like, literally, somebody, like, went and replaced all the music, all of the, like, infographics, even, like, the loading screen. Oh, creepypasta.

Yeah. Wow.

And I just, like, was like, hey, kids. Let's play Mario Kart. And I was like, this is

dark. This is what you downloaded? Oh, man.

That's so funny.

But, apparently, somebody just, like, like like, literally modded the entire image for Mario Kart. Yeah.

I'm looking at it right now.

Mario Kart black. Yeah. It's

Just got scandalous pixelated, princess peach on the back cover of this thing. This is what you this is what you, you loaded up. My gosh.

But that just goes to show how, like, not dialed I am on video games. I was like, Mario Kart black? That sounds like the best one. That sounds like the best one for sure. Yeah.

Wow.

Yeah, man. Modding a Wii is the move if you don't have a Nintendo Switch because that's we may need them a game. I modded my Nintendo Wii when I was in college maybe. So It's still it's still a good thing to do. It's great. Our kids love it. They don't know any better. That's the most hilarious thing. They don't realize that it can't be better. Probably for the best.

I'm gonna sick pick a kid's smartwatch.

So my daughter I don't know. Probably about six months ago, and my son just got one for his birthday. They got the AnyLoop kids smartwatch.

Cool. And it's, like, the right amount of smartwatch. It's the battery lasts, like, five or six days, which is awesome.

Like, it lasts like, it's not every day they're charging it. It does, like, step counting, which they love every day. How many steps you got? You Node? They're trying to compete with us. And it's also interesting as a parent to see how many steps they have. It does heart rate. It does weather if they're close enough to their iPad. It's just Bluetooth. There's no data on it or anything like that. No Wi Fi. But it's it's really cool. And they just I just loved having a watch as a kid. And now they have a watch which has a couple of neat features. They're setting timers on it and everything, and it's, like, something like $30.35 bucks. When we first got it, I was like, this thing JS probably gonna suck. You know? Like, how good can a $35 smartwatch be? But it, like, it's awesome. I'm I'm very impressed at how good it is for for a kid's watch. It's waterproof. It works great. So it's it's from any loop.

Cool. Sick.

Yeah. Love that.

Alright. That's it for today. Thank you so much for tuning in. We'll catch you later.