The Startup Ideas Podcast - WebMCP clearly explained (and how to make $$)

Starting point is 00:00:00 Think about this. AI agents are about to become a new kind of visitor on the internet, and I think billions of dollars of wealth will change hands in the process. Because millions of AI agents are going to be browsing and buying on the internet, but the internet isn't designed for them to do that. And that creates a really interesting question, which is, what does an agent-ready website or app look like? And how do you actually build one? Well, there's this new little thing by Google called, web MCP, and I'm honestly surprised more people aren't talking about it. The simple idea is that websites can expose cleaner actions for AI agents. So instead of an AI agent, trying to figure out a page like a human being, the site can basically tell an agent, here's how to search, here's how to book, here's how to buy. So today I have my friend Vinnie come on and just clearly explain

Starting point is 00:00:52 what web MCP is, how it differs from MCPs, why it matters, where the opportunities lie, and I give you two startup ideas that I think could cash flow that you can start today. Enjoy the episode. I hope it gets the creative juice is flowing. And if you haven't liked comments and subscribe, what are you doing? This stuff should be paid. We make it free. I just want to see you win.

Starting point is 00:01:14 Let's go and I'll see you at the end of the episode. Thank the MRR gods. Vinny is back on the pod to explain something very, very, very powerful WebMCP. Vince, you actually brought it to my attention. Vinnie, by the end of this episode, what are people going to learn? They're going to learn how to use WebMCP, which is kind of exactly like it sounds. It's MCP, but in the UI, in the browser. And this is like a brand new experimental feature from Google and Microsoft. So I really think this is a first mover advantage here. So if you can learn this stuff, you're really bringing, you're getting the skills that you need to bring stuff to market that is emerging. that's going to probably be the next big trend. So I think this is a really cool thing.

Starting point is 00:02:09 So Vince is going to actually explain to you in a clear way what WebMCP is, how to use it. And at the end of the episode, I'm going to share with you two startup ideas, cash flowing startup ideas that I think could use WebMCP that you can build today. That'll be at the end. Vince, you start them off. All right, cool. So let me start from like a problem I had myself. when I heard about WebMCP, the first thing that came to mind is this shop that I've got here. It's an espresso machines and gear shop. So I was buying an espresso machine and it has like a lot of specifications.

Starting point is 00:02:45 It's got a lot of gear and accessories and figuring out which one's right for me and also which accessories fit was kind of a pain. So it was a lot of back and forth, Googling, also putting stuff into an agent or chat. GBT or whatever and like does this fit will this work so I thought oh cool this is a great use case for something like this where we could build a shop with web MCP tools in it that your agent can talk to so you can bring any agent that you're using which is whether it's like clod code or chat gpti co-work whatever these whatever agents are out there maybe grokbot and you can talk to the website and ask it stuff so let me just show an example real quick. And so I want to upgrade from my current MaraX machine. And I want to get one that's able to make two flat whites a day easily and fits on my counter space, which is about 32 centimeters wide.

Starting point is 00:03:45 By the way, in simplest, like in the simplest form, correct me if I'm wrong, Vince, but WebMCP is basically just websites with agent buttons. So we're making websites agent readable and agent actionable. Is that right? Exactly. And it's not like this wasn't available before, but it was available in a lot like clunkier way. So you've got computer use or you've also got agents that can kind of browse and use a browser and go through websites, click buttons. But before they had to do it in a very not so efficient way. So they're getting the entire DOM, which means they're getting all the code from the website. And they're like, scanning through it and finding where to click, taking screenshots and stuff.

Starting point is 00:04:31 But look at that, what just happened right here on the screen. It manipulated the website, was able to find two machines, put them side by side, and then it highlighted really cool feature, highlighted what features or what specs about the machines fit my request that I just had. So it's like the port of filter fit, which is going to fit like an accessory you have, and then will it make the two flat white? like you asked, and does it have the right width? Will it fit on your counter spot? So like, that's really cool. And before, to do all that, well, first of all, it wouldn't be able to manipulate

Starting point is 00:05:07 a page in this way as easily. And also it would have to go through all this scanning, all this code reading, whereas this, now with WebMCP, there's a set of tools, like the ones that are just popping up on the screen here. And those are the ones that you make available to the agent. So it gives it a really efficient, concise list of things that it can do on the website, and it knows how to do that, and it can just do it easily. And you can do some really cool creative stuff with this. Yeah, and I think one of the reasons I think this actually matters is if you think about SEO, right, it was like, can Google understand the page? And then we had AEO, which is can AI cite the answer? And now we're entering this WebMCP era, which is basically, can Google,

Starting point is 00:05:55 And the agent finished the job. And what we're learning is that websites aren't really optimized. And I think that's why Google and Microsoft decided to build this thing, right? Because they don't want a broken internet to exist. And I think as like a founder, like why does this matter? It's because there's billions of dollars changing hands on these tools, right? Or on the internet. And if we can optimize some of these websites, you know, there's just, there's a lot of opportunity.

Starting point is 00:06:31 Yeah. And I think, I mean, maybe the people like you and I and the people listening, they're more agent-native or they're more this kind of early adopters of new technology. But we have to remember that the web has been around for a while. People are used to UIs and browsing websites and stuff like that. And people don't really want to, the majority of web users don't want to use just like a CLI tool or an agent in the terminal like this. They want to be able to interact with something visual. So this Web MCP really makes it the best path to do that. So we'll really allow for kind of normal, the average consumer, to adopt agent tools. So it's kind of like meeting them in the middle, right, with the new agent technology. and the stuff that they're used to.

Starting point is 00:07:22 Cool. What else do we need to know about it? So, like one of the big things is, like I said, is that it's not scraping and parsing the entire DOM. So it doesn't have to, it's more efficient in the way it works. But the other thing that's really cool is that these tools, you can make them conditional. So like you see here in this list, we have tools that are always available versus logged in.

Starting point is 00:07:45 This is all happening through the browser session, right? So I've logged in. You can see I'm up here. I've logged in. These are the tools available. And if I log out, well, then we'll see there's only three tools available now, right? So this is really cool because one of the big hurdles to doing other kind of agent-native approaches like an MCP server or an API is that you have to deal with tokens, credentials, people sending their, their identity through. you know, API keys and tokens and things like that, whereas this you just log in, right?

Starting point is 00:08:23 So I can log in with my username and password. And it has access to the entire browser session there. So the fact that I'm logged in, the agent has to, doesn't have to do anything special. We don't have to connect an MCP server and authenticate and all that stuff like you have to do with, you know, a typical MCP server or with an API or things like that. So it's like the least technical approach to making an app or a website or a web store agent native, but with the most flexibility and the most features. So this conditional tool calling, all this stuff just happens because of the browser state.

Starting point is 00:09:02 And that's a really nice thing. Yeah, that's really cool. That's just like built in. I mean, it makes sense, right? Yep. It's using the whole browser session. So that's really nice. All right.

Starting point is 00:09:12 So I saw this tweet the other day and I've seen this going around a lot where people, are saying, I don't want to use your products agent. I want my agents to be efficiently enabled to use your product. And that's exactly what we're getting with WebMCP, and especially for things like web stores and stuff like that, for e-commerce or maybe even, you know, other examples that we'll look at later that you, that you have prepared for us. So this is a sentiment that's going around on Twitter a lot lately. And what does Dylan from Cloudflake? mean by this. So we have these like different ways an agent can meet an app, right? And let me go zoom in real quick because we have a lot of them. And they're going from headless. You see at the

Starting point is 00:09:57 bottom here it says headless. So basically running an app via your agent without ever seeing the app. That's what headless means, right? There's no app in the browser. There's no website. So you bring your own agent, you can hit a raw API. But there's like we said, there's lots of things. I have to deal with authentication and the agent has to hold an API key and things like that. But then you can bring your own agent and you can use MCP, which is a typical MCP server, but that's another layer on top of your app's backend. And then your agent can't manipulate the UI, right? Like we just saw.

Starting point is 00:10:35 So those are completely headless ways. Add a little bit more complexity to your app or to your product. Now we have computer use, which is where the, the bots or, the agents literally have a computer that they can use, whether it's your local computer or a computer in the cloud, and they're running through screenshoting stuff and clicking stuff on the computer, on the UI, like you would. And this isn't really the most efficient because of, like we said before, you've got to go through, you've got a screenshot, or the agent has to screenshot, all this stuff. So this is slow and fragile at the moment. We have browser MCP, which is similar

Starting point is 00:11:14 to computer use where like we were talking about where you get the whole code from the website and it has to kind of figure out through the HTML code and where it can click and what it can do. Again, that's a little bit more fragile. WebMCP kind of comes now to be somewhere in the middle of that or the things we just mentioned and an in-app agent, which is what the tweet was complaining about, right? Which is that like you build an app with a little agent sidebar and your the agent in that app is the only thing that can communicate

Starting point is 00:11:48 and deal with it, right? So what they're saying is they want something where they can bring their agent and they can use your product efficiently. And that's where I think, especially for traditional apps with a front end, storefront, whatever,

Starting point is 00:12:05 that they can then use, bring their agent and use your app efficiently. So that's what's really cool about it. Does that make sense? So if I'm getting this right, so WebMCP, you've got your agent, there's a set of tools or things that you can do basically on the website or the app. And it's sort of limited to that. Exactly. The in-app agent, you have the vendors agent and the app's agent, right?

Starting point is 00:12:38 and that just interfaces with the app's logic and data back and forth. And the app is like, isn't that like the best end state of where we think things are going, right? The in-app agent, you mean? Yeah. Well, I mean, you can look at it from two different perspectives, right? That it's nice to have like a little in-app agent there. But you're responsible then for configuring everything. you're responsible for the tokens, right?

Starting point is 00:13:09 You're responsible for studying it all up. Whereas with WebMCP, WebMCP is like this really nice middle ground where you don't have to pay for tokens. You're giving the app agent capabilities, but the user is bringing the app, the agent that they used, and any agent will work with it. So if we're moving quickly to this world where everyone has an agent that they use at work,

Starting point is 00:13:32 that they use for personal preference, you know, you don't have to then separate them from that. That's the agent they know, and they can just use it with your tool via WebMCP. So it's kind of like the nicest, because it makes it easy for you, and it makes it most comfortable for them. Yeah.

Starting point is 00:13:51 I guess the in-app agent was sort of like a transitionary piece of software that needed to exist for a certain amount of time, but we're like kind of growing out of it. The idea of, you know, you use Intercom Finn, right? Which is basically like I think a customer support agent. That experience is really, really good. But now in a world where it's pretty clear that over the next 24, 36 months,

Starting point is 00:14:19 everyone, consumers are going to have their own agents. It just makes sense that these agents are going to need places to interface with the Internet at large. And the Internet is just not designed for those agents. agents yet. So, you know, that's why WebMCP, you know, exists. Yeah, I think there is a place for in-app agents, right? Like if in Notion, for example, they offer both. But like, if there's something that is like Cloudflare dashboard, they have an in-app agent. And this is, so it's like very, it's gated to only use their, um, their tools and not to change preferences too much so that you don't do anything kind of hazardous. I can see where in-up agents,

Starting point is 00:15:04 might make sense, but we're moving into this world where it's like, we're doing so much with agents and we want them to keep their context. So you want one tool that kind of has a context and knowledge and memory of everything that you do. And to be able to bring that one agent with all that context to different tools is really powerful. And I think that's where you're going to see people. They're going to kind of get stuck on one agent, one kind of agent products, whether that be Grockbot or Claude, co-work or chat GPT work or whatever. And then they're just going to kind of access the internet and interface with different things on the internet through those agents. So that's where I think WebMCP is really powerful. It's also just like

Starting point is 00:15:44 a great companion for when you're shopping with for complicated things. You can use an agent if you need to kind of help you out, figure out which specs with things right. And you don't have to build in this big bloated, you know, in-app agent into your, into your app to get some better customer experience for you. Can you show an example? Another example? Yeah, sure. So I could say like, yeah, so you suggested the Lalit Bianca. What are some really nice accessories I should add that will fit the Lalit Bianca? So this is something that like bugged me a ton was I found the machine. Now I need a grinder.

Starting point is 00:16:26 Now I need the right porta filter. I want a new basket. They're all different sizes. They all use one might use a measurement, one might use a name or some kind of like brand name. And parsing through all that information as a customer can get really tedious. So just to have an agent that can do that and then let you know is perfect because it has all the context here from the store. It knows what I've ordered. It has different tools here. Like if I scroll over here, it can see, get my gear is one tool. So it knows what I've ordered from here. It can get the cart. It can add coupons. So it can apply a coupon on checkout and all this stuff.

Starting point is 00:17:04 So here it's just saying the first thing you should buy is this little lit water softener two pack. So I'll just say, cool, add that to my cart. That sounds great. Right. So the nice thing is, is that I could tell it, highlight that on the UI. Or I can just say, you know, whatever. I like the sound of that. Go and add it to the cart.

Starting point is 00:17:23 And it'll do all that stuff in real time. And I have visual confirmation of that. I can see the things, right? And I think that's pretty cool. That's really cool. I mean, come on. That's really cool. Yeah.

Starting point is 00:17:36 So, and even tell me here, you have, softener is in, you have a barista 10 coupon, 10% adding it for the rest. So. I don't know if it's just because I'm a coffee aficionado, you know? Yeah. I love coffee. You know, you can see my coffee set up back there. Yeah. That this is just hitting for me so hard.

Starting point is 00:17:58 But like, to me, it's like, it makes so much sense. Like we're as customers, we're going to be using agents. We're going to want them to go and do stuff for us. We want to make sure that the things that they do, it actually is able to complete the tasks. And to see this work just so seamlessly, just is awesome. Yeah. Yeah, it's really cool. I think so this is a great example, but you can think about anything where it kind of gets complicated.

Starting point is 00:18:23 You know, technical stuff. If you're buying hardware and I don't know, you need to, you're doing like hard. hardware hacking and you need to write the right battery for the development board or car stuff and you need the right parts like all this stuff that can be really complicated like that this is a great use case i mean i'm just curious you know where mcp in your mind where mcp fits which i think is like another way of saying you know how do you like in which sectors niches should you be implementing web mcp because there's going to be like some some industries are going to be last to implement it, right?

Starting point is 00:19:05 And I think from a founder perspective, I care about the ones that have the highest priority. So which are those and why do they matter? Yeah. Yeah, I think like, like I said, these compatibility driven commerce, right? So it's like all those things, camera systems,

Starting point is 00:19:21 any kind of web store that's complicated. That would be a great thing. You could do it in SaaS admin. So like any kind of admin board or console or analytics or marketing tools, you can add these things in to make it easier for your marketing team to deal with these things and set up features and stuff like that. And even like regulated industries, like the nice thing is, is that you can't use API keys. So you don't want to go to, you don't really want to use MCP servers or APIs.

Starting point is 00:19:51 You maybe want to set up a few. And that's a great thing about WebMCP. It's like a few tools that are using the browser state that can, can be read only so they don't actually touch anything. So let's imagine like insurance or banking self-service. Like you could set up some web MCP tools that are read only or only touch very non-sensitive settings. And then that will help the user figure out where they need to go to enter something or where they need to go to get the right information or can even summarize and pull the right information for them via their agent. So stuff like that. I even think.

Starting point is 00:20:30 internal tools, so building internal tools at a company. So if you're like a entrepreneur, you don't necessarily need to use MCP right away for your customers. And there might be a really great solution for you in-house, like your internal tools, stuff for yourself, for your employees, that's where I think there's a lot of potential. I love it. All right. Let's go to some business ideas that people can implement. I'm curious your take on it too. So, Let me share my screen. Yeah. All right, Vinnie.

Starting point is 00:21:04 So WebMCP is here. I wanted to come up with two business ideas that I think are cash flowing, that I think are anyone non-technical can set up or somewhat non-technical. You just don't need to be like a hardcore engineer. The first idea is a WebMCP conversion agency. And I bet you there's probably some domains that are available like WebMCP. agency or something like that that someone can go and grab right now. So the big idea is how can you make boring business websites agent ready?

Starting point is 00:21:37 So you have these boring businesses like law firms and home service and HVAC, you know, med spas, dentists. And you just go and create these, you know, V1 tools for them. So that can be like requesting a quote or booking consults. And you're just selling them on this idea that there's going to be a new class of citizens on the internet. These are going to be these AI visitors. And these AI visitors, these agents, are going to book jobs, right? So you sell them on maybe a $2,000 to $10,000 setup. And then you charge, you know, could be $500 a month or a few hundred dollars a month, $700,

Starting point is 00:22:17 depending on who it is. Is it a big law firm, a small law firm, on monitoring and evals? And then you build this retainer that, you know, is building cash for every single month. and you just update it and that sort of thing. What do you think of this idea? Yeah, I actually think it's a really good idea. I see stuff on the web floating around where they're showing that things like you said, dentists or law firms are starting to use agents at their offices.

Starting point is 00:22:47 And so I think they'll get it. And it's good that you bring this up because I think it's worth mentioning that WebMCP is an experimental feature in Chrome right now. So it launched in February. joint effort from Microsoft and Google, but it's still in early days. But that's why I think there's real potential here because we talked about this package. I showed the diagram of, you know, from Headless to bound to UI, the different types of agent-native approaches you can use to connect to apps, right? So you could do this as a package where it's not only Web MCP but other tools, maybe a MCP server for them,

Starting point is 00:23:28 And you tell them that like, look, this is new emerging stuff, or you're learning this because it is this new emerging stuff so that when it does get adopted, you're the first mover. You have the first mover advantage. And so you can sell it as a package. Like we'll put in MCP. We'll put in web MCP. We'll create, you know, whatever, all this stuff that aligns with their vision. You can make whatever they want available and you can show them the different approaches. You have one that uses a UI. one that is headless and you're the expert there and that's the way that you make money. Yeah, maybe it's just agent conversion agency. And then WebMCP is just one of the technologies that you use. Yep. Or like we were saying before, internal tools. It doesn't only have to be about customer facing stuff. You can show them how they can make processes in their office better for them and their

Starting point is 00:24:23 employees using stuff like WebMCP. Totally. Cool. Startup idea number two. I'm calling it the agent mystery shopper. So it's testing whether agents can complete the most important user journeys. So if you think about a user journey, maybe it's a Shopify store, you know, you want to buy a hoodie. If it's a consulting company, it's booking a console.

Starting point is 00:24:48 If it's an insurance company, you're filing a claim. Requesting quotes, you're ordering items, all these different things that you want the agent to do. So the agent is going to run a bunch of tests and you, the customer, get a report back. Where did the agent get stuck? Were there bad descriptions? Where they're missing tools? Conversion risk. And then you're basically selling, you know, agent reports.

Starting point is 00:25:12 Maybe it's monthly, weekly, and you turn the repeated fixes into software. So I like this business because it's pretty simple. It's almost like, you know, just a report, like a credit report. almost on, you know, how your website and your apps are around implementing things like WebMCP. And then you can charge a few hundred dollars a month or $100 a month for something like this. What do you think about this idea? Yeah, I think also a cool thing. It makes a lot of sense. And the cool thing is, is that you can even pair approaches like this to your agent conversion agency, right? So it's like you can go and run these agent tests and evals and you can be,

Starting point is 00:25:56 and you can show them that, hey, this might work better if you used WebMCP on top of it. You're only using X, why not check out this as well? So you can kind of combine the approaches, I think, here. So that's a really cool idea. I like it. Yeah, so I just also wanted to mention that I showed a couple examples here, but there are 16 different tools that you can use with your agent via WebMCP on this example, espresso machines and gear store.

Starting point is 00:26:25 and you can do that right now. You can go to Kremah and co client Fly.Dev. We'll put the URL in the show notes and stuff. And you can test it out there. One thing you should know is that, like we said, is experimental stuff early days. So go into Chrome flags, enable WebMCP support, and go into Chrome Inspect and allow remote debugging.

Starting point is 00:26:46 So these are the things that you'll want to enable so that you can start using these experimental features. And if you want to build your own app, like this, you can also click on the GitHub repo button down there next to the WebMCP Tools Pill and you can clone this repo and you can play around with it, make your own version, whatever, it's free for you to start playing around with cool business ideas. So those are two important things I think are worth mentioning and yeah, it would be cool to see what other people do with this repo, what other people do with these ideas. So people should definitely share those

Starting point is 00:27:23 online with us. I think that would be awesome to see. Yeah, we'd love to see it. And I think, you know, there's, I can just see the comment of someone saying, well, it's an experimental, it's, you know, it's so experimental. Like, it's so, it's way too early. Like, why even bother playing with this stuff? And I, you know, you know me, I love playing when things are experimental. Because when, you know, the arbitrage exists when, when it's experimental. Like, I love trying things, getting my hands dirty and seeing what I can build now because when it's going to be super popular,

Starting point is 00:28:00 that's when that window starts closing. Yep, definitely. Think about when ChatGPT came out or when GPT, I think it was GPT 2 and 3 came around and people were playing with it and going, it was just two years before Chad GPT launched and people were going, ah, this is crap. We're way off, we're years away from AI being like a serious

Starting point is 00:28:21 thing, a serious tool, and then boom, the explosion happened just a year or two later. So like shouldn't pass this stuff off as, you know, just like experimental stuff that's not going to get implemented. I mean, we're talking about Google here. They have pretty good foresight. This is something that they've been talking about for a few years now. They made the proposal two years ago. They launched it at the beginning of this year. So I wouldn't, I wouldn't fade this personally. And I'm excited about it. So that's enough for me. All right. Thanks for coming on, Vince. Sharing the sauce and I'll see you next time.

Starting point is 00:28:56 Yep.

The Startup Ideas Podcast - WebMCP clearly explained (and how to make $$)

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.