Pod Insights Syntax - Tasty Web Development Treats 893: Everyone Is Talking About MCP About this episode Scott and Wes break down the Model Context Protocol (MCP), a new open standard that gives AI agents secure, tool-like access to your dev environment. They cover how it works, why it’s a big deal for AI coding workflows, and real-world use cases like GitHub, Sentry, and YouTube.
Show Notes 00:00 Welcome to Syntax!
00:49 The lore of ICP. Wes MCP Shirt.
03:09 Brought to you by Sentry.io.
03:33 What is MCP?
05:06 The steps of AI coding. 07:11 MCP hosts.
07:28 MCP clients.
07:35 MCP servers.
08:24 Why you might want to do this. 10:39 How this works in VS Code.
14:10 Wes built an MCP server. SVGL.
14:57 Playwright.
17:24 Sentry’s implementation. Building Sentry’s MCP with David Cramer.
18:54 YouTube implementation.
21:19 DaVinci Resolve implementation. Smithery.
23:02 Postgres.
24:40 Transport protocols.
24:49 STDIO.
25:19 SSE.
25:32 Streaming.
26:24 Writing you own MCP server.
26:28 FastMCP.
27:00 Cloudflare.
28:01 Data validation.
28:47 Standard schema. Episode 873.
29:27 Other parts of MCP. 29:35 MCP resources.
30:37 MCP prompts.
30:48 MCP roots.
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
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Scott Tolinski? Welcome to Syntax. In today's episode, we're talking about MCP. In fact, a lot of people, yeah, MCP. You know MCP, you know me, all that stuff. A lot of people on Twitter have been talking about MCP. A lot of people on YouTube have been talking about MCP. This is a thing that is seemingly, I would say, having a moment right now. So I figured we would, we would talk about it. Now, it is an AI topic. So those of you who are AI adverse, I think you should still know what this thing is. You should still take the time to understand it. But it is going to be a topic of AI. I know that is upsetting for some people.
0:49 Scott Tolinski? Wes has created an ICP t-shirt that says MCP, which is very funny.
0:56 Wes Bos? Oh, that is great. I don't know how much longer I'll laugh at the AI-generated t-shirt designs, but that is funny. ICP, maybe for, I don't know, maybe the Europeans don't know. Can you give us the lore of what ICP is?
1:12 Scott Tolinski? Well, hilling from Detroit, Michigan, my stomping grounds. Yes, the Insane Clown Posse or ICP. Yeah, they're an American hip hop duo. Yeah, basically, it's like, I don't, I don't really know how to describe the music other than just kind of, hmm, not great. It's rap music in a way, but it's rap music. It's like a very specific type of rap music. Man, I don't know the best way to describe it other than just being kind of like.
1:41 Wes Bos? It's like a lifestyle, like the people that, people that listen to ICP are Juggalos, right? Just go search for Juggalo documentary. Yeah, it's a cool world.
1:52 Scott Tolinski? It's a cool world. Hold on, I actually have an ICP story, Wes, related to web development.
1:57 Wes Bos? Lay it on me.
1:58 Scott Tolinski? Yeah, I, so I worked for an agency in Michigan for a long time and it was a small agency and, oh, we used to bid on projects all the time. One of those projects that came up, ICP's website. And we pitched them on it and we won the project. We won ICP's website. But then the owner of the company was like. Guys, I've been doing some thinking and I decided to listen to a lot of the ICP records and I think we're gonna have to turn down this project. And they, let me tell you, they were a cool company, too. I, I like, it was all cool people that worked there. People were like, but he was just like, I really just don't know if we want our company's name on this website. Like, I really just don't know if that's what we want.
2:43 Scott Tolinski? And so we didn't get to do it. I, I might have been the developer on that site. I might have had that on my resume. Uh, so.
2:49 Wes Bos? You would have.
2:50 Scott Tolinski? Sad.
2:50 Wes Bos? Man, that's funny. That's like the, the poor guys that did the Kanye website. They're all so proud of it and then they just had to come out and say, scrub it from their records. Oh, that's funny.
3:02 Scott Tolinski? I wonder if you would have gotten, uh, free tickets to the, uh, whatever carnival thing they do.
3:07 Wes Bos? Yeah, I don't, I don't think you would want to go to one of those.
3:09 Scott Tolinski? And if you want to see all of the errors in your application, you'll want to check out Sentry. At Sentry.io forward slash 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 Sentry.io forward slash syntax. Again, we've been using this tool for a long time and it totally rules. All right. So MCP, what is this thing now that we're done joking about the insane clown posse? Basically, it is a protocol. You think REST, you think GraphQL, these are protocols for communication, right?
3:48 Scott Tolinski? So this is a standardized protocol for LLMs to communicate with tools. So whether that is your database, whether that is, um, you know, even like certain APIs, whether that's tools like Sentry at Sentry.io.
4:06 Wes Bos? Nice.
4:06 Scott Tolinski? This is a protocol for AI to be able to communicate with these tools in a, like a more predictable, but also a more connected way so that you can do things on behalf of yourself with LLMs or even just in general AI clients, whether that be, you know, Claude Desktop or Cursor or WinSor for any things like that. You can do things in a more connected way. Via this protocol.
4:33 Wes Bos? Yeah.
4:33 Scott Tolinski? Which was created by the folks in Anthropic and Anthropic in like November of 2024. So this is basically a protocol that's made this to be more standardized for these tools to talk to each other. And we'll talk throughout the course of this episode about a little bit about some of these terms, but also what you can actually do with this stuff. Why is it useful? Why are people talking about it? But in general, just think about it as MCP is the protocol for the connection between an API and the LLM.
5:06 Wes Bos? Yes. If, if you think about like the steps of, of AI that we've had, like we had, at first we had autocomplete, right? Oh, it knew that I wanted to type this thing and it autocomplete it. And then, and then we were able to like ask questions about chunks of code. And, and then we got agents which could do things like, like run terminal commands. They could relent code. They can check outputs and inputs. It can, and it kind of recursively calls itself until it's, it's finished. And then that got us pretty far, but there's, we always hit a point where the AI agent is not able to, to do all of the things that, that we want to do. It's not able to, to query a database. It's not able to book a flight for us.
5:50 Wes Bos? It's not able to log into your GitHub and leave a comment for yourself because it doesn't have access to those things and it, it doesn't know how to do those things. So the MCP, the main point of MCP is to provide a whole bunch of tools, which you basically code a whole bunch of tools, like leave a GitHub comment, merge pull requests, close pull request, like a tweet, invite Scott, attach the, you know, it could, could literally be anything. And then you, you code up all of these different tools and then you provide them to the LLM and say, hey, these are things you can do.
6:33 Wes Bos? And if you want to do them, here's what they look like, you know? Here, it's essentially just giving it a library of functions. That it can possibly run for you.
6:44 Scott Tolinski? Yes. And what that means in regards to running functions and things for you is that if you're a developer out here, you might be like, oh, it's just running stuff on my behalf. Yes, in that regard, you kind of have to be aware of what these things have access to, what they could possibly do. And you might be able to get yourself in some sticky situations this way. So it is important to be careful with some of these tools. So some of the the terms that you have here would be like an MCP host. This is a program like Claude Desktop. This could be WinSerp. This could be Cursor.
7:23 Wes Bos? It's in VS Code as of yesterday.
7:25 Scott Tolinski? It is. Yes, it is in VS Code. You also have MCP clients, which is a protocol clients that maintain a one to one connection with an MCP server. Okay, the server is basically a it's a server like you know, a server, but it's a lightweight program that exposes specific capabilities through this protocol. Basically, we write servers to do API to to host your API or whatever. This is a very similar to that. It's a server that, again, just establishes what are these functions that the AI can access and do.
7:58 Wes Bos? Yeah. And then the the client would be the way that you connect to it. So if you think about like a traditional API, your client might be fetch or or Tansac router or a web socket connection. But in in the case of like the people who are building like VS Code, Cursor, all these IDEs, they would be implementing the client library. If you would only be doing that if you are building your own agent.
8:24 Scott Tolinski? Correct. So what are some things like, why? Why do people care about this? Why is this interesting? What? What does it matter? Well, let's give you some examples of some stuff you can actually do with this just broadly. Before we talk about like what it takes to actually make one or use them or anything like that. So GitHub released their official MCP server. And you could imagine that this might be useful in terms of being able to communicate with an LLM or communicate with any sort of client here to do things like.
8:58 Scott Tolinski? Open a PR, look at issues, maybe even analyze issues or things like that. Give me, give me the first 20 issues in this repository and tell me what they have in common. I don't know, you know, but basically it can do things like that. It can do anything that you would want like a GitHub API to do. In typical API fashion, you're calling like a function to do things like potentially merge a PR or open a PR or any of that kind of stuff. Now you could integrate that and do it via an LLM rather than via code.
9:29 Wes Bos? I don't even, yeah, I don't even think of it as like things that you would typically use an API for. It's things that you would typically use you for. So for example, I hooked up the GitHub MCP server and I said, hey, how many, how many issues are currently open on the, or how many pull requests are open on Westboss? And it went up and went in and said, ah, you have 28 pull requests open. Most of them are minor grammar and spelling mistakes. So I was able to say, all right, go through them and merge any that are clearly spelling or grammar mistakes. Take the entire blog post into context, leave a nice reply. And I went through and it merged a couple for me.
10:10 Wes Bos? It left a nice little comment, a little nervous doing that, but it worked really well in that like I,
10:18 Wes Bos? you're able to use the like flexibleness of flexibility of an LLM. We're like, like, look at the pull request. Is it just like a simple spelling mistake or grammar? Then yeah, go ahead and merge this thing and leave a little comment. And it went through and it basically what it does is it runs all of the different tools. So the, here, let me, let me share my screen real quick. So here I'm, anyone who's not listening, I'm just looking at a list of all of the tools that are available. So the GitHub MCP server, what it does, it fires up a little Docker container locally. And then it has what 20, 30 different things. Create a pull request, update, create a branch, add issue, comment, get issue, et cetera, et cetera.
11:01 Wes Bos? So basically we'll look at all of these tools and their description and say, all right, I think I will use this one. To get a list of all the issues and then I'm going to loop through each of those and and read them, read through them and then I will decide if if I would like to close it or merge it, then I will I will use the merge tool, etc., etc. It works really well. I was I was surprised at how nice and smooth it was to be able to do that. And I previously would not have used AI to do that. Or I would just like copy paste everything in into like an LLM and say, which of these are which of these are worth merging?
11:41 Scott Tolinski? Can you tell me how you got to this again?
11:43 Wes Bos? Yeah, so this is VS Code.
11:47 Scott Tolinski? Oh, it's VS Code.
11:48 Wes Bos? With their MCP. I was just just trying it out because it it VS Code launched agent mode and MCP servers and their next edit suggestion. And those are like the three huge things that people have been like people left for cursor for and and windsurf for. So I was like, all right, let's let's give it a shot. And it's really nice. So I have this little MCP dot JSON and I have a list of all of the different MCP servers I want. And then it tells you how many tools are available in each of those servers.
12:23 Scott Tolinski? Yeah, that's the thing that I've had a hard time with is making sure that the the chat, whether it be Claude desktop or anything, actually uses the MCP call. Like, I don't know if you've had any trouble with that, but like Claude, I had for the YouTube MCP, I had it all set up correctly. And Claude desktop was like doing everything in its power to not use it. And I was just had to be like, you have the you have the following capabilities. Like, use these capabilities, which is crazy where cursor had a much easier time with it, where I'd say specifically on YouTube. Hey, give me the the five most recent videos from our YouTube and output their their stats or something like that. And Claude would be like, I don't know what your YouTube channel is.
13:05 Scott Tolinski? And then like cursor would be like, yeah, here you go because it had my API key. It knew what channel I had via the configuration of it. So it's just like getting it to actually utilize the tools I found to sometimes be a giant pain, but I haven't tried this VS Code limitation.
13:19 Wes Bos? In VS Code, you can do a pound and then you can actually explicitly name them. Ideally, you would just use English to say what you want. It would figure them out. But if you have too many tools, then it might pick the wrong one. So you can also explicitly tell it which one you want. So here, I just went ahead and said, how many issues are on the syntaxfm/website, GitHub repo? Summarize what they are about. And it's going off and writing that. We'll see. Let's talk about one that I actually built and we'll come back to that. Cause it is a little bit slow, which is not an issue when it's doing work for you in the background. But when you're kind of just sitting there, waiting for it to do things, it can be a little bit sluggish.
14:01 Scott Tolinski? Yeah. So you wouldn't want to do it for things that are easily accomplishable by you in one second. Right?
14:06 Wes Bos? Exactly.
14:07 Scott Tolinski? Yeah. It's not exactly a savings at that point.
14:10 Wes Bos? So I actually built one because one of the problems that I have with AI is that it can't ever get logos for you. You know, and I often want to be like, show the Chrome and whatever logo. So I wrote an MCP server for SVG L, which is, you told me about this. It's like a, it's a website. It's a raycast extension for getting logos for coding things. Right? So I wrote a little MCP server that just hit their API, fetched it, and then returned the SVG code. That's a perfect example of like, go to this service, download this thing and bring it back into my, my project.
14:55 Scott Tolinski? Yeah, no, that's a great idea as a demo for that.
14:57 Wes Bos? Playwright is another really good example for one. You can kind of describe it as it gives your AI agent hands. So if I can go into my history, here we go. Can you visit syntax FM and tell me if there are any console errors? So that's kind of interesting, right? Visit this actual website and open the console and see if there are any errors. And if this was the syntax repo, it could maybe even find the actual issues that were happening. Right? And then it, so what it did is it fired up a headless or head full playwright browser for me. And, and then it visited the website and it says there are no issues. And then I said, are there any accessibility issues?
15:41 Wes Bos? So then it ran another tool called Browser Snapshot, which gives you sort of a accessibility tree of all the HTML. See, that's the Syntax FM homepage.
15:54 Wes Bos? And then it ran through that and it found a couple of issues that we could possibly fix. So that was good. This is, by the way, this is why I get so many garbage emails these days. Because people are using these tools to just like, hello, Mr. West boss. I found these following issues on your website and they're just churning them out. So like these, I don't know, as much as I love this stuff, I also, it's just has created such a problem with, with people emailing me.
16:21 Scott Tolinski? We got one the other day was like, I love this episode and the episode had not been released yet. I was just like, okay, you're just like, you found it in our GitHub or something. I don't know.
16:30 Wes Bos? Oh, really? Oh, I think I saw that. Was that on Twitter?
16:35 Scott Tolinski? No, it was on email. I got an email.
16:36 Wes Bos? Oh, really?
16:37 Scott Tolinski? Yeah.
16:38 Wes Bos? Well, that's interesting. Yeah, they, they probably scraped the, the markdown files. But then I asked it, go ahead and give me the last 20 episodes. And it knew to click on the show's page. And then it knew to click on the show's page two, to get another 10. And then it went ahead and summarized all of them for us, which is, is amazing. So yeah, it's almost like another scraping agent as well. Like, you know, visit these websites. And I've heard from a lot of people that their bandwidth is going way up lately. And it's because, not because people are visiting their website, but these hungry LLMs are scraping the top 50 websites when somebody asks a question, compiling that all in and, and pushing it out the other end.
17:24 Scott Tolinski? Yeah, yeah. Interesting stuff. Another thing that people are doing with the Sentry actually made one themselves. And Kramer was telling me a little bit about this. David Cramer was saying like, you know, the idea would be like from within a file to reference and be able to discuss like specific errors or look up, are there any? Century, you know, errors that are associated with this file, et cetera. Being able to like actually communicate with your error hunting platform directly from your codebase as like a cool use case for this. He actually just published a video too today that was talking about how they built the or how he built the MCP server. So probably worth your time to watch it. I think it's a pretty useful.
18:07 Wes Bos? Yeah, that's it's such a good use case of like, what would you normally do as a developer? Well, you flip over to your Sentry, take a look.
18:13 Scott Tolinski? You go back.
18:14 Wes Bos? Flip back to your, your, your codebase. You, you try to line those things up, but wouldn't it be nice to just say, hey, is there, fix all the issues? Like, it's funny that we've, we've literally joked about this in the past where like, the only century thing Sentry doesn't do is, is fix the code for you. And like, that kind of does that now.
18:35 Scott Tolinski? It's getting there. Yeah, that's for sure. I know, which is crazy. Yeah. Yeah. Which again, if you're having like production software and people are using it, you kind of want it to be fixed really, really fast. So if you could do that directly from your codebase or even like supervise something as it is fixing the issues and things like that, it could be helpful. I did give the YouTube MCP server a rip and I found it to be pretty bad, which is another thing that is like, you are kind of at the, the whim of the specific server. So I think some of them function and operate better than others. Like the YouTube one specifically says that it's implemented the following API calls. Now I've used the YouTube API pretty extensively.
19:14 Scott Tolinski? So I know what these API calls can do. And so even prompting it and then having the MCP server say that it's going off and doing the thing and then actually returning wrong or bad information, very possible, still very possible with these things. So that's not like they're always going to be a silver bullet, especially if you're at the whim of who wrote the server.
19:37 Wes Bos? Totally.
19:37 Scott Tolinski? But you could do things like compare my most four popular videos. Now I found it to be like, I would say, give me the, the five most recent syntax videos and it has our API key. It has access to our channel. It has access to that API functionality. And it would still give us like five random videos from our channel.
19:56 Wes Bos? Oh, interesting.
19:57 Scott Tolinski? That's not what I'm looking for, you know? I, I don't know why. I haven't dove into it yet, but again, these things are, at the end of the day, they're still LLMs and it is using this method, but it's not, not a perfect science.
20:09 Wes Bos? And and whatever MCP server for YouTube must have that ability to sort or whatever built in.
20:16 Scott Tolinski? Right.
20:17 Wes Bos? And it certainly is very early days for a lot of the stuff. A lot of the stuff feels janky or like, we'll talk about the the transport protocols in just a sec. So it certainly feels very early and you should definitely be very careful because I've seen lots of malicious MCPs, you know? You are letting a random website run random commands on your, you can obviously have to approve them before they run, but you don't know what they're actually doing. And there's been a lot of people who have been been hacked or had their, I don't know, crypto stolen and all that.
20:58 Scott Tolinski? They stole my ape. Yeah.
20:59 Wes Bos? Certainly.
21:00 Scott Tolinski? Get your ape stolen.
21:01 Wes Bos? Don't go willy nilly looking, just trying every single MCP server out there because it can be a little bit, little bit freaky. We should say though, that these are not just for coding, like Scott gave a good example of YouTube, but there's, there's other ones like what DaVinci has one?
21:19 Scott Tolinski? Yeah, that's, that's what I found to be really cool. We'll share, there's a, Smithery is a website that you can go to like look up some of these. And there's also some repos like awesome MCP servers and stuff. But yeah, I found the DaVinci MCP server, DaVinci Resolve is the video editing software that I really like to use. And you can do things like create timelines, add clips to timelines or whatever, but you could imagine that could be handy to say like, I don't know, I'm not, I'm not even aware of this MCP server could do this, but you could say, hey, give me five random clips from my media pool, cut them at random intervals, throw them on a timeline. And this like, do those types of things that you could probably do with scripting in DaVinci, but like, then you're getting into Lua
22:04 Scott Tolinski? and you're writing Lua scripts, which I've been doing and it's not just, it's not a lot of fun. So yeah, you can control applications like DaVinci with it.
22:13 Wes Bos? Even just like, how do I do this is often the question, like I'll jump into photo. Photoshop or Illustrator and there's all these like 3D video editing. There's all these like fringe programs where like I'm okay at them, but like I'm not an expert at them. And often I will just type into an AI of like, how do I add a motion blur in Photoshop? And it'll tell you like, click on this, click on this, click on this, and here's all the different options. And wouldn't it be one step further is it just doing that for you or at least opening the context for where you're at? Or sometimes you're like just trying to click on something, but you're not in the right mode.
22:57 Wes Bos? And the LLM would know to switch you to the correct mode before you do that.
23:02 Scott Tolinski? Yeah, word.
23:02 Wes Bos? There's absolutely lots of them out there. Postgressql is another really interesting one for coding because if you're building an application, the LLM doesn't know anything about the actual data in your database. It might know about the schema, you know, it might need about your config and your types. But with the Postgresql one, you can simply just say like, show me the 10 most recent shows in our database. Or what tables are in here? Or even like, I would, I would imagine you could do something like create a drizzle schema for this existing database.
23:40 Scott Tolinski? Totally, yes.
23:41 Wes Bos? Right?
23:42 Scott Tolinski? Drop all tables.
23:44 Wes Bos? Drop, drop all tables, exactly.
23:48 Wes Bos? But like, you think about the kind of cool tools that you could build with like internal tooling? You know, give me the last 100 sticker orders and show me the average amount spent per country.
24:05 Scott Tolinski? Yeah. That's what I was really hoping for out of the YouTube MCP server, which why I was so disappointed by it. Because I wanted to be able to analyze stuff better. And right now I'm building my own set of YouTube tools that are doing exactly that, but it's doing it in code. And that's fine. That's totally fine. But like, if I want a piece of specific analysis, like, you know, I have to write that code myself. And I actually have to have it in code. I have to test it. I have to make sure it works, all that stuff. Instead of just being like, Oh, give me the five most popular videos and tell me what they have in common, you know?
24:39 Wes Bos? Mm hmm. So we'll talk about actually setting up a server or using the different servers. There's, there's two, maybe three different kinds of transport protocols. One of them being standard I O, meaning that it's a script that the MCPs will call and then that script fires up and you can interact with it. And then it will, it'll spin down then when you're finished, right? The, the playwright one is like that. It'll spin up a playwright instance. You can interact with it, change pages, take screenshots. And then, and then once you're done, it will, it will spin down and close that instance.
25:19 Wes Bos? A long running MCP server is something that is always up, up and running, right? And it works on server sent events where you're simply just sending data back and forth the entire time. Then they are now adding support for, for streaming instead. Because like, if you wanted to deploy a MCP server to like a serverless function, Vercel is pretty, um, pretty vocal in this area as well. It's not really possible, right? You need a long running server. Wouldn't it be nice just to have a serverless function that you can hit? And then stream back and forth with it. And then once you're done, you can shut it down. So I think that that is probably going to be the more popular way to go about it.
26:04 Wes Bos? But right now, at least most of the tools for building a server, you either go with one or the other. And, and usually the standard I O ones are, they're often just N P X scripts.
26:15 Scott Tolinski? Yep.
26:16 Wes Bos? Sometimes there are Docker containers that you'll run locally. Depends on, of course, the, the different tooling that is necessary. Writing your own, um, there's a couple like tools out there. So the one I use is called Fast MCP and it's sort of like, The jQuery of writing MCP server, just a bunch of like handy little functions, or maybe the low dash of it. And that is written, that was written in Python, but somebody has like ported the exact API to TypeScript, which is really nice. And then the, like Claude themselves has the MCP TypeScript library, which is more of the lower level APIs. You could build one with that.
27:00 Wes Bos? Cloudflare is starting to do a bunch of work in this space as well. I was at a conference last week where Angie Jones gave a talk. She works at this company called Block and they're working on this thing called Goose, which
27:13 Scott Tolinski? Oh, yeah.
27:14 Wes Bos? is another way to sort of get up and running with that as well. So there's, I think it feels very early in all of this stuff. So if you're dipping into it, it's, it'll probably feel a little bit rusty here and there, but I think everyone's trying to figure all this stuff out. It's only been, I think, November of last year is when Claude announced it. So we're only like six months into this. So I think that this will really change, but the standard has been adopted by all of the editors. It's been adopted by OpenAI. It's obviously by Claude. So it's, the standard is here.
27:55 Wes Bos? And what that standard looks like in six months, a year from now will probably look a little different.
28:00 Scott Tolinski? Yeah, it'll probably have evolved a little bit. I do want to say, you were mentioning that the fast MCP, one thing you'll notice with a lot of these like starters and tools is that data validation and general like
28:12 Wes Bos? Yes.
28:13 Scott Tolinski? data types are important. So most of them will be using ZOD or archetype or Valibob. I see most of them use ZOD, to be honest.
28:21 Wes Bos? Yes.
28:21 Scott Tolinski? Whether that is for the inputs or outputs or whatever, you're giving the LLM that much more context by telling it what the data should look like. So very widely used a tool to do data validation.
28:35 Wes Bos? This, yeah, because you need to make sure that the data coming in is exactly of the type that it expects, right? If you're asking an LLM to be very precise, you need something like that. So Fast MCP uses standard schema. Remember, we did the episode explaining what standard schema is, where basically somebody who needs to add validation to their library can use standard schema and then people can use any standard schema compatible validation library. And this is like one of the first like new instances I've seen of something using it. So Fast MCP will allow you to use ZOD or Valley Bot or what's the other one?
29:16 Scott Tolinski? Archtype.
29:17 Wes Bos? Archtype. Those are the, obviously, the big three, but anything that is standard schema compatible.
29:23 Scott Tolinski? That's syntax episode 873, if you want to get that lesson.
29:26 Wes Bos? Yeah. Other parts of MCP. So we've pretty much been talking about tools in MCP, but there are other parts of it. There are resources, which has not been implemented by any of the editors yet. So in my example of the downloading the icons, a resource would be something that is returned from that. And you could say a resource would be an image. You could say a resource is like a video or a PDF or it can be anything, right? Anything you can then feed back to the LLM for additional context. And in my case, the resource would be returning an SVG image. And it didn't work.
30:08 Wes Bos? And that's because none of the editors support MCP resources yet. So I had to go a different way where I just returned the SVG code. And then I had, I had the AI just cat it into, into a file. You could also just use node APIs, fs.writefile, something like that. So that's something that will probably be better supported in the next two, three, four months, I bet.
30:36 Scott Tolinski? Yeah, interesting. Yeah.
30:37 Wes Bos? There's also MCP prompts, which are like reusable prompts, templates, workflows that people who are using your servers can use. Again, those are not supported by anybody. And then there's also routes, which I found this out the hard way where I downloaded the icon and it literally put it on forward slash because it doesn't know where it's, it doesn't know where it's running. So if you can get, if you can provide it a route, like the currently opened folder, then you'd be in a bit better shape.
31:09 Scott Tolinski? Interesting. All of these are, I'm linking to the model context protocol.io docs for each of these as well. If you want to dive into them, we'll have these links available to you in the show notes.
31:21 Wes Bos? We should also say there's, there's also like Python, Java, Kotlin and C sharp SDKs for MCP. So you don't have to write it in JavaScript, although you probably should because it's the best.
31:32 Scott Tolinski? Yes, right. We are not biased on that one. I'm also going to link the awesome MCP server GitHub. There's a lot, you know what? I find this personally to be a little bit overwhelming. It's like, I hear about this new thing. I'm going to go try it out. I go to the awesome MCP's repo and there's like 10,000 things on this already. And it makes me feel like I'm late to this. Oh, crazy. So there's a lot of people building on this right now. And who knows if all of these are fully featured or working correctly or whatever. But if you want to be inspired or even just see what you can possibly experiment with this, if this sounds interesting to you and you want to give some of these a rip, this
32:12 Scott Tolinski? awesome MCP's GitHub repo or the link that I had earlier on in the show, the smithery are good places to at least hunt for some of these things to see, get inspired maybe.
32:24 Wes Bos? Yes. Again, be very careful. Don't just install a bunch of random ones, but
32:28 Scott Tolinski? Yeah, they can mess you up.
32:29 Wes Bos? Some, some fun stuff in here, you know, like I'm, I'm really excited about this because I have to, like one example is I have to go to like eight different bank accounts four times a year and download all my transactions.
32:39 Scott Tolinski? I'm not giving this access to my bank accounts. I'm not doing that.
32:42 Wes Bos? Yeah, that's probably not a good idea. But like, no, I'm thinking like, Could I build one with Playwright?
32:49 Scott Tolinski? Yes.
32:49 Wes Bos? That is flexible enough that it could click the buttons for me?
32:53 Scott Tolinski? Yes.
32:53 Wes Bos? And download everything. I don't know if I can or not. It's like a 15, 20 minutes of clicking. Do I spend two days?
33:02 Scott Tolinski? Yes.
33:03 Wes Bos? Building that? Yes, I think I should. That makes sense.
33:06 Scott Tolinski? Go for it.
33:07 Wes Bos? Developer math. All right. I think that's it. Let us know your favorite MCP servers. That's kind of the big thing is I'm, I love hearing from everybody about what are, what are they using it for? Cause it's, it still feels so early.
33:19 Scott Tolinski? Yeah. Are you down with MCP? Let us know. We want to know again, what you're using it for. If you think it's stupid, if you think it's great, if you think it's going to delete your entire database, whatever. We want to hear all about it. What, have you, have you done anything cool with it? Just let us know. Leave a comment. We see your comments on Spotify. We leave your comments on YouTube is a great place to see that. All of the above. So let us know. And I don't have anything else about this. Wes, do you?
33:45 Wes Bos? That's it. Thanks everyone for tuning in. Catch you later. Peace.
33:48 Scott Tolinski? Peace.
Previous 892: The History of Syntax Next 894: Open Source Matters w/ Chad Whitacre Page updated November 3, 2026 · transcript and summary are machine-generated