947: S-Tier MCP Servers for Developers

Scott and Wes break down the top-tier MCP servers developers are using right now. From browser automation to debugging superpowers, they explore how these servers are changing what’s possible in modern dev workflows. Show Notes 00:00 Welcome to Syntax! 01:52 Brought to you by Sentry.io. 02:46 Submit your Spooky Stories! 03:37 Syntax San Francisco Meetup. 04:11 S-Tier MCP Servers. 04:17 Sentry MCP Server (not sponsored). 05:01 Pause, what the heck is an MCP server. 05:45 Back to the Sentry MCP. 07:07 Spotlight JS. 09:08 ElevenLabs MCP hack. 11:19 Docs Tools. 11:29 Context7. 15:55 GitHub MCP Registry. 16:10 Nuxt MCP Server. 16:25 Svelte MCP Server. 20:35 Cloudflare MCP Servers. 22:34 Chrome Dev Tools (Playwright) MCP Server. 25:17 Stripe MCP Server. 27:55 Mastra. Hit us up on Socials! Syntax: X Instagram Tiktok LinkedIn Threads Wes: X Instagram Tiktok LinkedIn Threads Scott: X Instagram Tiktok LinkedIn Threads Randy: X Instagram YouTube Threads

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Scott Tolinski?

  Welcome to Syntax Today. We're going to be talking about S tier MCP servers for developers. These are going to be MCP servers that work really well in general development flows to help you build stuff, to make things, to vibe code, any of that. And by no means is this an exhaustive list or anything like that, but these are a handful of MCP servers that I've used personally, that I have really enjoyed and that have made my time working in code a lot better with agents. My name is Scott Tolinski. I'm a developer from Denver. With me as always is Wes Boss. What's up Wes?

• 0:38Wes Bos?

  Hey, stoked to talk about this. It feels like we've had a bazillion MCP episodes, but there's just like this never ending well of stuff to talk about, especially now like OpenAI just announced a whole bunch of their agents SDK. We'll have a show on that, but it's never ending.

• 0:55Scott Tolinski?

  Yes, it's never ending. And one thing that I've personally felt is that like, yes, MCP to me always seemed like an interesting idea, something that could be really advantageous to use, but we have sort of reached a point where MCP servers are doing so much and so many good things they've gotten really good that I'm finding their utility to be more and more something that I rely on every single day when I am using agents to code. I actually, we should have Kent C. Dodd's on the show. He's been doing a ton with MCP servers and he and I chatted a whole bunch at the VS Code Insider Summit about MCP. And his thoughts

• 1:40Scott Tolinski?

  on MCP overall had like really shaped how much I've dove into adding these tools into my workflow lately, and it has really paid off. So before we get going, let's talk about Sentry. This show is presented by Sentry, sentry.io/synthesis. Sign up again two months for free. What is Sentry? It's the perfect place to fix your bugs, make sure your application's running smooth. I'm going to talk about a newer MCP server that They've released recently beyond the just the already awesome Sentry MCP server. That is like awesome S tier development tool. So that that's not even going to be part of this sponsor

• 2:23Scott Tolinski?

  read, right? That's going to be part of the normal episode because it's that good.

• 2:27Wes Bos?

  It's just good. Yeah.

• 2:28Scott Tolinski?

  It's just good. And I more and more rely on Sentry's tools to fix my stuff, man. Seer, is there AI thing that that goes through and finds the root cause of your bugs? And I've had just such good luck using that thing. So check it out. Sentry.io forward slash syntax. Sign up two months for free. Tasty treat. All right.

• 2:47Wes Bos?

  Real quick, before we get going, two things. We need your web development horror stories. And we got a meetup in San Francisco on October 27th. So if you have a horror story for our yearly Halloween episode, where we tell stories of people who have you've dropped the database by accident, you deployed a fix that DDoS'd yourself, you wrote a bad React hook, you accidentally forgot the where clause in a database.

• 3:12Scott Tolinski?

  About the werewolf clause.

• 3:13Wes Bos?

  Oh, man, these stories are awful. I want to stick my head like I'm so stressed out reading these stories every year, but it's so enjoyable. Please, please send them to us. Go to syntax.fm forward slash spooky. Submit your story or send me an email, [email protected], send me a DM, wherever. We would love to read it. We make everything anonymous. We're not going to say who you are. But we'd love to read them off. And then also, October 27 in San Francisco, we are doing a meetup and we want to invite you. We are at Bear Bottle Brewing in San Francisco. Go to syntax.fm forward slash meetup. You can grab some tickets. We're going to have some merch there.

• 3:55Wes Bos?

  We're going to have some brewskis there. We're going to have Scott there, myself, CJ, the whole team. It's going to be a hoot. So come on down.

• 4:12Scott Tolinski?

  This is the end, end ad read. Let's talk about S tier MCP servers. Now, the first one that I have on here is the Sentry MCP server. The Sentry MCP server is great because you're, if you use Sentry, which is again for tracking errors and all that stuff, you want to be able to integrate and use this tool to use your Sentry from your agent, from your CLI in terms of like Claude code, whatever you're using. All these examples that I'm showing you on screen today will be done with Claude code. But if you cannot see, you're listening on audio, you're not missing anything. We're mostly just kind of showing what they, how they work. So.

• 4:55Wes Bos?

  And we should say, before you even get into that, for anybody who, who hasn't heard, these, these MCP servers are a way to surface additional tools to AI.

• 5:08Scott Tolinski?

  Thank you.

• 5:08Wes Bos?

  And that could be coding, right? Via Claude code or VS code or cursor. It could be via, via chat, a chat app, and it could be by simply making straight up API calls to a like OpenAI or Anthropic, but you provided additional tools to do things that it, it might not be able to do, it might not be able to access specific things.

• 5:33Scott Tolinski?

  Yeah, you're providing the AI with specific tools to do specific things. So I've loaded up the Sentry MCP and I've looked at the tools. There's 19 tools available from find teams, find releases, get issue details, get trace details. You can create a team through this. You can create a project. You can update a project. So what that means is that oftentimes when you are in the past, you're working with Sentry or any of these types of services like this, what do you do? You go to their website, you create a new product, project, you maybe you get an API key or a DSN in Sentry's case. And then you run through the wizards or you Sentry has a CLI that does it for most platforms. And the docs are great, but you're going through and you're copying paste and stuff from docs.

• 6:18Scott Tolinski?

  Getting set up, you're having to click around. With this, you can create a project directly from here. After that project is here, you can then search through issues. You can analyze issues with their C or AI platform to help fix them. And ultimately, what that means is that you're spending less time leaving your tools, which is Claude Code, your editor, whatever, anywhere that you're using this MCP server, VS Code, Cursor, or whatever. You're spending less time leaving that to go to Sentry's UI to do things and more time just staying where you are. As somebody who's allergic to context switching like that, it's made a big impact for me.

• 7:01Scott Tolinski?

  So that is the Sentry one because again, we use those tools. For me, this, this works so well. But they also released a newer MCP server, which is the Spotlight JS, spotlightjs.com. Basically, this is like a local development tool where like Sentry primarily, I'm using it mostly for production bugs and things like that. Where Spotlight is something that's collecting your errors. The MCP server in regards to this allows you, again, you're getting local errors. You're getting local logs, traces, get events for traces. And this is still brand new. But if you're working in local debugging, I know me personally,

• 7:43Scott Tolinski?

  I'm oftentimes having the MCP server or the, the agent having to run the process to then read the console to then, or maybe even look into Chrome Dev Tools MCP to read the console logs.

• 8:00Wes Bos?

  If you're simply just copy pasting it like a sucker and then there's no

• 8:04Scott Tolinski?

  Copy and pasting it.

• 8:05Wes Bos?

  Additional context on top of that in terms of like, how often did it happen? What's the most common error? Yeah, this is awesome.

• 8:13Scott Tolinski?

  Yeah, so if you have this set up, you could say, you know, read this error, whatever, and you'd be able to get that error directly from the MCP server without having to. You know, instantiate other tools or, or, you know, run additional process. So I've really liked this. And on top of that, there's also like this spotlight UI that you get. Um, so it's not just an MCP server. There's also this UI for being able to debug your app. So, uh, this has become a

• 8:42Wes Bos?

  It's like a, like a local sentry.

• 8:45Scott Tolinski?

  It's a local sentry for sure. And I, I've really, really liked this. Uh, and they're working on it pretty hard. Cause, um, I gave them some feedback on the install process and sure enough, this landing page has changed since I've done that. So, uh, shout out to the sentry team for knocking this thing out of the park.

• 9:01Wes Bos?

  Um, look, can I talk about the 11 labs SDK, um, or the 11 labs MCP real quickly? Uh, this is not necessarily something that you'll do, but if you are the type of person who's like, I don't really see that. The, the point, I don't really see the purpose in this type of thing. What I did is I took the 11 labs MCP server, which 11 labs does like voice, uh, text to voice, voice to text. Um, but they also have like an agent's SDK, which is able to kind of spoof like, or make a phone call, right? You can queue up an agent with a whole, but whole knowledge base. You can give it a bunch of tools. You can give the agents MCP servers as well, which is crazy.

• 9:44Wes Bos?

  I guess that's like your MCP server can have MCP servers inside of them.

• 9:50Scott Tolinski?

  Totally.

• 9:50Wes Bos?

  Um, but the crazy thing was, is I had a bug in my code. And instead of like asking, I had it set up in cursor and I said, instead of like asking cursor to fix the bug. I said, call Wes and ask him how to fix the bug. And it literally rang my phone.

• 10:09Scott Tolinski?

  Oh, yes.

• 10:10Wes Bos?

  Asked me what I would do in this situation. I replied on the phone with what I would do. It transcribed it, put that back into cursor, cursor then parse the whole transcript and then put, applied the fix. Obviously a stupid thing to do.

• 10:26Scott Tolinski?

  So funny.

• 10:26Wes Bos?

  But that just goes to show how, and I'm often just shocked at how you don't have to write any code for, for any of these.

• 10:34Scott Tolinski?

  Yeah.

• 10:35Wes Bos?

  They just all work together. Like so much of my career has been wiring things together. And with these MCP servers, it's just.

• 10:46Wes Bos?

  Most of the time.

• 10:47Scott Tolinski?

  Language is wiring it together.

• 10:49Wes Bos?

  Yeah, yeah.

• 10:49Scott Tolinski?

  Yeah.

• 10:50Wes Bos?

  Exactly. So I thought that was, that was a pretty funny and like, just like if, if you haven't seen the benefit, try that. Like it's, it's eye opening as to like how these things can work together. My cursor called me on the phone.

• 11:02Scott Tolinski?

  Cursor called me on the phone. That seems like a, a nice little like Dev Horror story. I was trying to fall asleep and cursor called me on the phone. All right. So next up is a class of tools that I'm calling like docs type of tools where you are getting the, the documentation for various libraries. Now, Context7 is the big one for this. And Context7 works super well because what you do with Context7 is you're basically saying, give me the documentation for this library. Like use Context7 to give me the documentation. And what it does is it goes off and it gets the ID first and then it goes and gets the, the docs.

• 11:43Scott Tolinski?

  And that way it has that in its context. So the agent is able to, in it, in the context of what it's working on, able to take in the documentation of the project that you're working in. Now, this can sometimes lead to context bloat where it's bringing in a lot of too much context. And I have some solutions for some of that here, but I found that this is like one of the, the best things to do when you are finding that the agent or the AI is writing code that is not in adherence to the actual library. Uh, because if, if you say use Context7 to get the up to date docs or whatever, it's going to be able to have that in its context and actually, uh, work on it.

• 12:28Scott Tolinski?

  One thing that I've been doing lately is anytime I, I have a library, whether that's like, I, I wrote the, um, zero svelte bindings for zero sync. I threw that on Context7 and it's been so nice to be able to be like Context7 to make sure that you're using the documentation for this.

• 12:45Wes Bos?

  It, it's so key, especially for, obviously for libraries that don't, like the, the AI doesn't know about, or it's just hallucinating it, especially like sometimes it just hallucinates method names when it clearly has access to the types, which drives me nuts because like you shouldn't. Maybe you shouldn't even need docs if the types are good enough.

• 13:06Scott Tolinski?

  Yeah, maybe. Yeah. Sometimes it's just for best practices or yeah.

• 13:11Wes Bos?

  Yeah, or examples of like how to how to configure it and whatnot. That's really good as well. But like even just like AI still writes old tailwind all the time.

• 13:21Scott Tolinski?

  Old every.

• 13:22Wes Bos?

  So frustrating.

• 13:23Scott Tolinski?

  Yeah.

• 13:23Wes Bos?

  It writes old everything. Yeah. And you have to like explicitly tell it use these features. And like, that's a bit of a, like, we're kind of stuck in this, like, 2023 version of coding things.

• 13:36Scott Tolinski?

  Totally.

• 13:36Wes Bos?

  Unless you explicitly tell it to use the newer version of something. And like, I worry about like React, all these new React features we talked about last week.

• 13:45Scott Tolinski?

  Yeah.

• 13:46Wes Bos?

  And like, it's

• 13:47Scott Tolinski?

  Svelte, yeah.

• 13:48Wes Bos?

  It doesn't know about it.

• 13:48Scott Tolinski?

  Svelte too, yeah. There's just so much of this stuff, you know, I, and for me, it's not even that, it's CSS, man. Like, the amount of things that I have to throw into my, my context, or I have to have special CSS agents to write good CSS.

• 14:04Wes Bos?

  Yeah.

• 14:05Scott Tolinski?

  AI models today, all of them. And people that I hate that wouldn't leave the comments, they'd just say AI, but they don't even say which model. Well, that's usually because it's most of the models, right? They, a lot of them have the same issues. They all write dogshit CSS. They all add background color to stuff that does not need a background, right? You, this is, you made, you set the background of this input to be the exact same as the background of the container. Why?

• 14:31Wes Bos?

  Yeah.

• 14:31Scott Tolinski?

  They don't need that. That's just making this unmaintainable. Or you used transform, translate instead of translate. Gosh, that's the MCP server I need is the, do not ever use transform, transform, translate. Use translate. Alrighty, gosh.

• 14:48Wes Bos?

  Yeah. What does Contact 7 have for like modern CSS as well? Because like, we've been blessed with so much new CSS in the last like two years, you know? And I don't see a lot of it having, having huge uptake. And I think partially that's because obviously browser support. I think also partially just because like, the AI just loves to absolute position everything instead of use pop over and anchor positioning.

• 15:15Scott Tolinski?

  Let me throw an ass ton of JavaScript in here when you could use the pop over API, like. Okay, I know, I've been working on something for this last personally, because I'm so fed up with the CSS that it writes. And I have my own like CSS systems and you could obviously train it on those systems, but I want it's like a fundamental, like CSS fundamentals, because obviously the models are all trained on dogshit CSS because most people write bad CSS. So it's a, it's a, it's a problem, but I, I personally want to solve this problem because it's a problem. Now, addition to context seven, there's something new kind of popping up where libraries are doing their own MCP servers.

• 15:56Scott Tolinski?

  I have the MCP registry from GitHub pulled up. We'll link this in the show notes, github.com forward slash MCP. One thing that you'll see on here is there is a Nuxt MCP server specifically for Nuxt itself, working in Nuxt. And I haven't used the Nux one, but it is by Ant Fu. So I'm going to put it in S tier just based on that alone. But the one that I have used in this regard is the new Svelte MCP. This is their official MCP server from the Svelte team. It's in the svelte.dev. And all of their libraries and projects need to take note of how stinking good this thing is.

• 16:36Scott Tolinski?

  Because one really cool tool that it has besides Git documentation, you know, the big problem with Svelte is that they decided to release a completely new syntax for everything, like the, the day before, you know, models were released. So none of the, like even modern models, anything that comes out, GPT-5, the Nusana 4.0, they're all bad at Svelte still. They're not, they're not terrible. They're getting a little bit better, but they're still bad at Svelte 5. So this fixes that problem because you can say use Svelte MCP, implement remote functions. It's going to go off. It's going to get the documentation for remote functions specifically. And it is going to use that to work on it. So it always has that context. The same thing that you'd have with context seven, but this one's very targeted.

• 17:20Scott Tolinski?

  It's going to know the exact feature that you're asking it about. But one of the coolest features in here is the auto fixer. The auto fixer takes in a string of an entire component and then it

• 17:36Scott Tolinski?

  Returns what's wrong about this component. That way, the the agent, the AI, whatever you're working in can then use that information to update your thing locally. So this is sending a string of your component and it's saying, hey, you used on colon click instead of on click. It fixes the component. That rules. I think a lot of these things need that.

• 17:57Wes Bos?

  How does that work? Is it? Is it just like providing like a prompt to the AI? Or is there literally like a linting server hosted somewhere that it's it's a linting against?

• 18:09Scott Tolinski?

  Oh, man, that's a great question for the Svelte team. I think I think it's providing a prompt. It's sending your string of your component to a prompt somewhere. I don't know. Yeah.

• 18:21Wes Bos?

  Because it also could be like, like, running it through like this. Doesn't Svelte have like a Svelte linter where it will like, it'll flag stuff that is is maybe a possibly a bad practice?

• 18:32Scott Tolinski?

  Yeah, yeah.

• 18:34Wes Bos?

  Along with how to fix it.

• 18:36Scott Tolinski?

  Let's see Claude use Svelte auto fixer MCP to lint. Let's find a dot Svelte component. Let's find the let's do menu here and we're going to see what it does really quick. Basically, what it's going to do is it's going to, like I said, take a string of that. It's going to send it off to the MCP server.

• 19:00Wes Bos?

  The docs for this say it uses static analysis to provide suggestions for code.

• 19:05Scott Tolinski?

  So Cool.

• 19:06Wes Bos?

  What it

• 19:07Scott Tolinski?

  Thank you for looking it up instead of just guessing.

• 19:09Wes Bos?

  Run all of that through a like static analysis where it will look at all of the code and there's there's likely things in that as they they loop through the tree of the Svelte code that say, ah, I noticed that on colon click.

• 19:23Scott Tolinski?

  Yeah.

• 19:24Wes Bos?

  Let's let's throw a warning for that as well as a suggestion of how to fix it.

• 19:28Scott Tolinski?

  Yeah, it returns in an object with issues and suggestions. No issues.

• 19:34Wes Bos?

  It probably just runs it locally. So most of these MCP servers are simply just spinning up like NPX commands in the background. Some of them run Docker containers. Some of them are hosted externally and

• 19:46Scott Tolinski?

  This is a remote one.

• 19:47Wes Bos?

  Run it. This is a remote one. Okay.

• 19:49Scott Tolinski?

  Yeah, cool. So I think other frameworks libraries should take note. That rules. I know people don't have necessarily the resources or whatever, but like Auto Fixer thing, man, I love that. I would love that for CSS. I would love that for people who write React code to do something that one line of CSS would do. Like, hey, have you considered learning CSS before installing something?

• 20:12Wes Bos?

  Hey, maybe this is outdated or just like a MCP server full of dead simple examples that can then be projected on your codebase, especially for a lot more of the modern CSS stuff. One MCP server I love, or probably 15, is the Cloudflare has like 15 different MCP servers for each of their products. So they have one for docs, which is good if you need docs on how how workers run. But then you can also use it just like an API to like spin up a container or

• 20:49Wes Bos?

  create a binding to a database in this project. So where you might have previously had to like figure out what is the like wrangler command that I need to type exactly to make this work. It's so beautiful just to be able to type, you know, spin up a new D1 database and and put it in my wrangler bindings.

• 21:10Scott Tolinski?

  Yeah, I love that. Because again, otherwise, yeah, having to navigate the Cloudflare UI does drive me nuts sometimes. I don't know about you, but like as I use more and more of their tools, I'm just like, man, I feel like they could they could rethink some of this stuff.

• 21:26Wes Bos?

  They're having the same problem that Amazon had years ago, which was like initially Cloudflare was like dead nut simple. And then as more products get added, it's just like.

• 21:39Wes Bos?

  Where do I even go for the like, I was looking for my tunnels the other day and I just, it took me like five minutes just to find the link to find where tunnels were, but their search was actually really good and I ended up just typing in tunnels into the search.

• 21:54Scott Tolinski?

  So I happen to feel like it's a prank. They're like, we have a tool called tunnels and we're going to make you go through countless tunnels and mazes to actually get to where you're trying to go.

• 22:05Wes Bos?

  Find where it is. I don't know that I could design a better version of that. Like that's where like proper application designers come into play because like trying to keep things simple, but also still surface all the knobs for when people want them is a very hard thing to design around.

• 22:25Scott Tolinski?

  Yeah, yeah, for sure. That is tough. Next one here is the Chrome Dev Tools or Playwright MCP servers. I've been using Chrome Dev Tools over Playwright MCP servers for doing the same type of task now since they were released. So this is a brand new one. If you haven't heard about it, Chrome Dev Tools have released an official MCP server. It is, we did a show on it last Monday. So you should have heard it. We, if you're listening, we should have heard it. This thing rules. Again, this one is running into just an NPX command and it is firing this thing up. And let me tell you, I've had such good luck with getting into

• 23:07Scott Tolinski?

  loops where I'm saying, test this in the Chrome Dev Tools MCP server. That way it can find console logs. It can find like visual things that are off with it. It's taking screenshots. It's doing all kinds of stuff and you can see it work too. So it's not just like happening blind in the background. Yeah, like emulate network, list console logs, click, drag, fill form. I mean, there's 26 tools here getting performance. So if you want to evaluate performance, finding animation issues or key frame or a key, key frame, frame drops. Yeah.

• 23:46Wes Bos?

  There's keyframes too. I don't know if it's accessible via the dev tools, but you know what is one move that I really like in this is if you're working with an API that doesn't have types and you don't know what the API response looks like, normally what you do is you just console log it and take a look at it. And then maybe you copy paste that object out into some converter and convert types for it. But what you can do in this case is simply just say, hey, run the fetch, console log it, see what you get, then write types for that, then change it to be like an error, like make it intentionally break and then look at what the response for that looks like.

• 24:30Wes Bos?

  And then all of a sudden you have like this, these beautiful types and you can write a whole function that will go ahead and fetch everything for you.

• 24:37Scott Tolinski?

  Yeah. And people who are allergic to AI stuff, like you've got to be able to see the utility in that, you know?

• 24:45Wes Bos?

  You're doing that anyway.

• 24:46Scott Tolinski?

  You're doing that anyway. Also, that's like such a bad waste of time. It's something that you need, but you're like, the process of you going to do that, look, that's not helping you. So like, use the tool that just goes and does it for you. Verify, validate, all that stuff. But use the tool. It's great. This is, this one to me couldn't be any more S tool, S tier, S tool, S tier, Chrome Developer Tools MCP. Here's one that I used recently the other day that I was just blown away by is the Stripe MCP server. My gosh, I told Stripe to create a product, add it to my, my code base or whatever, make a buy now button, just for fun.

• 25:29Scott Tolinski?

  And man, it went through and just instantly, it took no time at all. It made the, I got all the, every single thing right about it. I've used Stripe a lot myself, so it's not like something that I'm just like, oh, okay, we'll see if this works. No, I told this to create this product in this specific way. And within, you know, a minute I was, I had a working product in my sandbox and I was just...

• 25:55Scott Tolinski?

  Wow. Okay. This to me saved me so much time again of going back and forth, hot potatoing things from the UI to my code, adding IDs, whatever. And the end result couldn't have been better. I found the Stripe MCP server to just be shockingly super good.

• 26:16Wes Bos?

  You know, another neat thing, like, of course, this is good for, for development, but I'm thinking about like, the amount of tooling, like, internal tooling I've written against Stripe for refunds, partial refunds, taxes, stuff, payment, you know, being able to change somebody's address on an invoice, all of that, so much, like crappy internal tooling as well, you know, like, it's not, it's not great UI. It works, you know, and, and my assistant knows how to use it. But I just look at this and think like,

• 26:52Wes Bos?

  how sick would that be just to give her a chat box instead? Or don't even have a person, you know, and, and like, of course, you just approve these things. But just like before, like, as the request comes in, just queue up the chat and propose refund $6 in, in Canadian or something like that. And it just hit the button and it will go off and do it for you. It's just all that wiring up time is gone with a lot of this stuff.

• 27:21Scott Tolinski?

  Time. Yeah. And, and even like I said, you can create products, create customers, list customers, create prices, invoices, finalize invoice. You can do so much stuff with this, create coupon. This, the entire Stripe dashboard is essentially here as an MCP server. And I found it to be just, just top tier, excellent, worked really well. If you're working with Stripe in any capacity, throw this in, you won't regret it. I, I found this to be just really, really great. Last one I have here is less of an MCP server and more of a platform, Maastra, Mastra, however you say it, folks. This is a tool that allows you to build like orchestration pipelines of potentially tools talking to other.

• 28:04Scott Tolinski?

  Tools, but the cool thing about MASTRA and the cool thing about it is you can expose and create MCP servers with this. So I can make my own MCP server that maybe I'm writing with just straight up JavaScript to go and scrape pages or do this or that. Or perhaps I wrote a tool that scrapes a page. That tool then uses another tool to save it to a SQLite database. And then I have another tool that's exposed as an MCP server to use information from that database that I'm then able to call for my application.

• 28:41Scott Tolinski?

  So like I built a tool that was scraping MDN and putting the things that I wanted to have from it in a SQLite database. Then I could then use MCP from Claude Code to say, uh, go, go grab this, um, MDN doc. It looks it up from my database, looks up the URL, then uses web search to load that page, then read some of that information. So you can really get totally customized with this thing. The MCP, again, the, the server that you're writing can just be straight up JavaScript. It can be the result of an, uh, an agent. It can be the result of other tool calls. And you can build these things all up and then expose it as an MCP server.

• 29:24Scott Tolinski?

  And this isn't like a brand new tool. It's newish and it's really super good. Another thing that some people might like about this is that there is UI in all of this stuff. So as you're building it, when you're testing these things, you can click around, you can see the results, you can see the returns. You can, you know, run things in parallel branch loop. I mean, it's very powerful tool. Even if you're not using to be its most powerful, uh, features, this thing rules and I've been having a lot of fun building my own stuff with it.

• 29:56Wes Bos?

  Yeah, they, OpenAI actually just released something yesterday at the time of recording that's very similar to this as well, which is just wiring things up. You know, you can click and drag things together. I still think like if this and that kind of dropped the ball because this, this was their game, you know? But yeah, this is, this is really cool. Just to be able to drag and drop and, and click things together as you need it. I was, I was looking into this as well because a lot of people are thinking like, okay, I need something like this for my business. And OpenAI is trying to like get you to to bring that into ChatGPT, or you can host your own. They released this thing called ChatKit, which kind of seems similar to this.

• 30:40Wes Bos?

  But yeah, a lot of people want to build these custom flows for themselves, but also give them to possibly the non-technical audience that they can just wire things up.

• 30:51Scott Tolinski?

  Yeah, this to me feels like firmly a dev tool, but then you can expose it as an MCP server to then connect to whatever application you're using it in. So there's so much here. This thing is really pretty neat. And I found working with it to be generally, you know, there are some iffy things here and there in the docs. It's some new stuff here and there, but like in general, I found this to be especially for like developers to to feel like, wow, this is a really cool tool. I've been really, I don't know how much you've had to deal with this, but I've been really getting into context management. People call it context engineering sometimes. And it's like, how do you give the AI enough context to do its job, but not too much?

• 31:35Scott Tolinski?

  Because the more and more you pollute that context, the more and more you get bad results, hallucinated results. You fill up the context and then it has to summarize. And then, man, I was doing so much with agents and subagents that were first calling and reading large MD documents. And then it just like wouldn't be able to do anything because that context window would fill up so quickly. So yeah, a lot of interesting stuff here. These can solve some of those problems, probably worth an episode of its own. Folks, is there an MCP server that you use in your development that really, really, really helps you out? Is there something that we didn't talk about that you feel like is necessary in day-to-day development?

• 32:17Scott Tolinski?

  I want to hear from you because, man, every single time I find a new one of these, like the Stripe one that makes me really that excited. I just like, man, this is a, this is a really cool place to be where I'm no longer having to go learn and explore some UI. I'm now getting to interact with it in the language. I need a product. It needs to look like this. Give it to me now. You know, love it.

• 32:41Wes Bos?

  Beautiful. All right. Thanks so much for tuning in. Catch you later.

• 32:44Scott Tolinski?

  Peace.