Subscribe
Share
Share
Embed
Founder Reality with George Pu. Real talk from a technical founder building AI-powered businesses in the trenches. No highlight reel, no startup theater – just honest insights from someone who codes, ships, and scales.
Every week, George breaks down the messy, unfiltered decisions behind building a bootstrap software company. From saying yes to projects you don't know how to build, to navigating AI hype vs. reality, to the mental models that actually matter for technical founders.
Whether you're a developer thinking about starting a company, a founder scaling your first product, or a technical leader building AI features, this show gives you the frameworks and hard-won lessons you won't find in the startup content circus.
George Pu is a software engineer turned founder building multiple AI-powered businesses. He's bootstrapped companies, shipped products that matter, and learned the hard way what works and what's just noise.
Follow along as he builds in public and shares what's really happening behind the scenes.
New episodes every Monday, Wednesday, and Friday.
George Pu (00:00)
Before we start this episode, I just wanted to share a little bit of personal thank you to all the listeners who have listened throughout those past couple of weeks until episode 22. I'm really grateful because for most majority of the podcast, think over 90 % of them, they don't make it past episode 20. So episode 20 is a huge milestone. So I thank everyone for listening in and a few updates about the formats moving forward. It's always difficult for content creators in general to identify.
What's about themselves versus how to help their audiences. Right. And we see in the real world about what, you know, content creators using their influences to do bad things. And I think for the founder reality podcasts, we're always going to stick with helping listeners to achieve more things. And our model in the future is going to be building wealth and freedom without permission. Right. And that's such a profound statement. think.
a mission that I personally believe in, I will be sharing that more with you guys, audience. So in the future, you'll be hearing more about my personal digestion about latest trends and news that are applicable to founders, entrepreneurs, anyone, right? And also some personal frameworks that I've been actually been using, instead of just theories, some frameworks I've used that can help you get further, you know, in life, in career, whatever. So I'm hoping this podcast can be the vehicle for all that. So without further ado, we'll spend today's episode, talk a little bit more about
A recent phenomenon that's picked up, I think has scared away quite a lot of people because of name, right? And the name is MCP servers, MCP servers. So what is MCP servers? We'll talk about that in a short second. But if you have been consuming tech news for the past couple of weeks, you can probably see that there's a shift happening in AI infrastructure in general. I think most vendors are missing this trend because everyone's asking about what's the best model, right?
While everyone's arguing, you know, is Google Gemini the best model is Chachibe the best model is Anthropic. The Spar money I think in tech are actually building connectors, you know, that work with all of them, right? This week, past week on GitHub, OpenAI and Microsoft just made a huge slot of announcements that made this approach inevitable. And I'm talking about MCP servers and MCP momentum. And if you're still building, you know, custom integrations for every different AI tool for your platform, you're about to feel very stupid.
First, as I mentioned, promise let me explain first what is MCP because the name obviously MCP and server itself sounds way scarier. I think that what it should be. So MCP stands for model context protocol. All right. And yes, the server name is very panicky, but don't panic. MCP is not about, you know, buying servers or managing infrastructure. Think of MCP in a very simple term as like a translator that helps your AI talks to different other business tools. Right.
So for example, there's a very easy way to understand this. Think about like, you know, when you use Chatchabit T and ask, Chatchabit T, what's the weather, right? It can't actually check the weather because it's just a large language model, right? So it doesn't have connections to weather data. So usually what happens is that Chatchabit T has to connect to a search functionality in order to search for the weather data, right? And that's essentially what MCP server is. MCP server is basically giving Chatchabit T a phone number to call the weather service.
But there's also a magic in this at all. Think about this, like once you set up that phone number connection, right? Any AI can use it. So it's not just chat GPT. It's actually invented by, you know, Anthropic, which is the parent company of Clot AI. So, but now it's being adopted by all the different AI services out there. So you can use it with chat GPT, can use it with Clot, you can use it with Gemini, you can use it with Grok, I think, and you can use it with any custom AI assistant. So they can all get.
the weather data. for example, right from the same connection within the same MCP servers, you just build a bridge once by building that call number to the weather service and everyone can cross reference it. And let's pause for a second because I want to show how this can be helpful for not just tech companies, but for any different type of companies out there, because with these sort of news, there's always this genuine confusion about this like MCP server, you know, it's a server it's for tech companies for bootstraps drives or SaaS. It's actually not.
So it's not just actually for tech companies. Let's say, for example, let's, let's think about a traditional service that we all know, like a dental practice, example, right. Dental practice. So if you are the owner of a dental practice, you probably want your AI assistant to check client schedules. Right. But right now, if you go to chat, you say, oh, access my client schedule. It doesn't actually have access to your booking system or your schedule, your database. Right. But if you can build an MCP connector, that's basically like giving your AI assistant the keys.
right to look up for appointments and reminders and check availability, whatever basically you are allowed to do, whatever you add to the MCP server, chat, GPT, Claude, whatever tool they're using is able to reference that actually make better about it. Right. And that's also think about a second example. Like if you're an e-commerce business, instead of manually checking inventory every time on Shopify or Wix or whatever platform that you use about, you know, asking the customers like, is this in stock? Right. Using MCP servers.
You can actually connect MCP server with chat GBT and your AI customer service is it be able to, it's ⁓ or you can call it agent, right? It's able to actually check your inventory system directly and give answers in real time, right? To your customers. That is basically, I think what most people are talking about AI agents these days. Personally, I still don't understand about the name AI agent quite a lot, but I'm assuming what it is. It's basically just like an MLM plus an MCP server. If that's the way of like AI agent, maybe that's the
Definition if you have a better definition definitely comment for the podcast below to let everyone know But I personally feel like the name MCP server and the server just means that the connection runs on a background, right? So you set up once it just keeps working just like for example like Wi-Fi, right? You don't really think about the router. He just use the internet So that's the same thing with MCP once you set it up very easily and no you don't have to write a code for it It just uses the internet and it just goes to your AI system that easily. That's basically why it is
So now that you know what MCP server actually is, let's talk about what happened this week and basically why I changed everything in the landscape of MCP and AI in general. So first, think GitHub, that wishes everyone understand what GitHub is. So GitHub launched a thing, a tool called MCP registry. So what it is is basically an app store for AI connectors, right? So let's talk about a little bit more about what companies are doing now, including like for simple direct.
So what we do is we basically go through random repos. example, Notion has their own MCP service on their documentation. So we'll go there. We'll go to Notion. We set up with that service and connect that through Clot or ChatchBT. And then we'll go to Zapier. We go to Intercom and go to those different tools. So as a private business, it's actually pretty, a lot of hassle to actually connect with different sorts of MCP service. You actually have to build multiple MCP service and connect with multiple, which basically makes it very difficult for you to do something, anything with MCP.
And that's why the barriers of entry are quite high. It's basically only reserved for companies that are a little bit more founded, a little bit more, you know, successful, have more funding resources, right. And it's not applicable to, for example, we just mentioned dentists and e-commerce companies. However, with this GitHub launch, think GitHub MCP registry launch, think that opens the door for everyone to be able to use this product without having to be strong quote unquote technical, right? So you now have a curated directory, right? And,
GitHub already launched with a lot of big names that everyone has probably heard about in the space. So Figma, Office Design Tool, Postman, HashiCorp, and many, many others. So this is not just some experimental side project, right? It's obviously GitHub, which the parent company is Microsoft's way of saying, this is infrastructure and this is the future of how companies and projects are supposed to do it. And just because name comes from GitHub, it carries significant weight and it's probably able to change the way that people use MCP service in the future.
There's also a second news about MCP. OpenAI also just announced that their responses API, which is basically the API that takes responses, that connects with other services, now supports remote MCP servers as well. So the translation of that is basically that your chat GPT agents can now call your MCP service over HTTP instead of you having to in the...
embed everything locally. So in the past, you have to embed things locally. There will be a local address for to call for Chatchabudy to call, but now you can do it anywhere. For example, like for my website, SimpleDirect, we will go to getsimpledirect.com slash mcp slash tools I need. And Chatchabudy can just visit that page, authenticate and start to use that server successfully. Right? Isn't that amazing? You don't have to always keep your laptop running for your AI agents to be able to use MCP servers. That's a huge change for.
chat GBT side on opening. I now supports the responses API now supports that. And also Microsoft, which has, you know, a juicy ratio with open L. Lee is also going all in, you know, with MCP servers. So this week they just shipped 10 different MCP servers for developer workflows. have integrated MCP into like the fabric, which is like Microsoft fabric, which is basically their enterprise database center. And they're also building MCP support directly into the windows operating system. You know, although I guess I'm not quite sure about how that will work with a windows operating system.
But I think the takeaway is that if Microsoft starts treating something as like an infrastructure and adding it to tools like windows, right? You know, probably that's past this, experiments phase and now it's getting a lot more serious. So as a founder or as a business owner, now it's probably the time for you to look into MCP servers, understand about it, right? But before we dive in further, I just also want to talk a little bit more about the MCP itself. So I actually thought about for a quick second about solving the same problem as well. I thought about, okay, simple direct, can I just solve the same problem because.
MCPs are so fragmented and you know, tools are everywhere. So what GitHub just launched this week, what they did is basically they just nuked entire categories of startup that are relying on MCP service because there are companies like Glamma, MCP.iso, Composio, and many, many others. They were building basically paid MCP directories and marketplaces. And for those companies, what's a business model is that they build the infrastructure for you.
So you use their platform to publish your own MCP servers and use their platform to connect with other MCP servers, right? They were like essentially marketplaces that allows any AI companies to build their own, you know, plugin. And for me personally, to be using some other developers products. So that's pretty cool. However, this week they, you know, they just got nuked because these companies, they are not Microsoft. They don't have hundreds of billions of dollars of cash on their balance sheet. So obviously they have the charge for discovery. They have a charge for premium listings. They have a charge for analytics.
And then I have to charge for usage as well. that's, and that's normal, right? For someone to be charged in that. But what GitHub does is that GitHub is like, Oh, here's the same thing. It's free. It's where your code already lives and your code doesn't have to leave the repo. And this is a classic platform move, right? It's basically launching something that takes something emerging and basically bundle it into becoming the default, you know, that's the crazy move. So if you're a developer, I definitely think twice about building something in this space right now, just because how competitive, ultra competitive it is. And it can be.
You know, even I was thinking about building something in a space before abandoning the idea because I thought, okay, someone's got to catch up. And wow, it's been two weeks and Microsoft just nuked that entire idea away. Definitely fast changing. Everything's moving pretty fast. And also wanted to add about, this is not really about like which AI is better, right? So there are many debates about which AI is better. I don't think actually we will actually care, you know, very soon about which AI is better. This is actually about infrastructure becoming a commodity, right?
It doesn't matter if it will open as doing it, Claus doing it, Microsoft is doing it or Google is doing it. The companies I think figured out MCPs while they will basically have AI agents or like AI systems that work across every platform, right? Which means that it's coming to everyone around the world. And in the future, MCP is going to be probably a name that everyone's going to remember similar to ChatGBT because the same thing is going to happen. Whereas MCP is going to become part of life, part of how you and I use AI in the future.
You know, so it's very profound and let's talk about, you know, a few other companies of like implications. Obviously we know that, you know, the MCP rushes have a very bad week. think let's see if a GitHub MCP rush should becomes the default. think that's a little bit hard for something I just launched. So we'll see if they're actually going to take away businesses from like Glami, MCP, ISO and Composio and a few others. they going, are it's GitHub going to nuke them forever or in the future is going to be a few other platforms out there. So.
That doesn't implicate that much for you as a founder. think it's more just about like if you're picking one of those like other platforms, just be careful. Don't use something that eventually can be shut down. Right? So if you're building a business, you know, never build where the platform can just roll in and give it away for free. You know, so now I think these MCP server companies have to justify why they're charging money. So it's a really, like I said before, it's really challenging a competitive space. And if you have to
just justify for being charging even like any money. That's crazy. Right. So let's move on and think about, know, basically how this AI integration MCP momentum is showing and what does that mean? Right. So I think this means that, for example, like for us at SimpleDirect, we were building separate integrations for a customer support system. We have to use the MCP server to be able to talk to Stripe to pull billing data, right. Which of our customers have to have subscriptions. So if a customer actually comes in and ask a question.
We have our MCP server talking to Stripe and basically pulling the data from Stripe about whether this customer has an active subscription, if they cancel, were they overcharged, a few other things that's for the AI agent itself to decide. And then obviously we also talk about like Zapier, we'll start to like different automation tools or that they're actually multiple. But with MCP, I just build one custom integration using Cloud Code with Stripe connector and every other person from our team can actually start using it.
So it's not just like time-saving. think it's architectural advantage, right? And also I spent, I guess, maybe a day and a half or two days, right? Like building this tool, I think in reality, probably that counts for like about five to six hours over the course of two days to build this like custom MCP tools and actually understanding what is MCP really means. And I was glad I didn't get discouraged by the name MCP server because in the old days, I would just probably not even look at it.
So now I don't think, I think the good news about this GitHub registry means that you actually don't need to reinvent the wheel about what exists, right? If you need to connect to Notion, this GitHub registry already has that, right? If you have, to connect to Stripe, you want to connect to Figma, you want to connect to anything out there, this GitHub tool already has it. And now it's launching, I was checking, look, I think it has around 30 tools out there. Be confident that it's going to add probably like many, many tools every day, every week. So by the time, you know, like you start building something you have, it's only to grow bigger and bigger and it's
probably a good idea to stick to GitHub slash Microsoft for this, because they're going to be here doing for the long term. And I think that stability is probably very important. So let's also talk about like how you can future prove your AHA stack, which I think is pretty important. So I personally think your business logic in the future is going to lie in the MCP connectors, right? So, and think about MCP as your core infrastructure is a really good way to looking at it.
Because think about like ⁓ models itself, it's very swappable, right? If you have built something, anything with AI, you know that models are swappable. So we were using Chatchpd 4.0 and then we swapped to Chatchpd 5.0. We were using, for example, Cloth 3.5 and now it's Cloth 4.0, right? And it initially is going to be 4.5 and 5. So the model itself gets swapped pretty fast. But with MCP, the thing is you only had to build it once, right? If let's say GPT 6 or 7 comes out, you just need to switch the model, keep the connectors, right?
If clock gets better reasoning, the same connector will always work. So you never have to keep swapping and stopping the code. just build one MCP code and all the models that come out in the future, can swap it and the performance will probably just be the same. And you're decouped from essentially model when you're locking in. So you can change from chat GPT to claw to Gemini to grok, right? You can change to any different sort of AI providers.
And I think as a founder, as a business owner, that's really like a liberation as well. That means that you're not locked in for anybody. So that's a good news. So basically my advice is build your own MCP infrastructure, or maybe just use the existing ones with GitHub and just keep that flexibility and start using that with your daily workflows. So also obviously there's also a lot of risks. I think it might be worth mentioning for MCP servers, right? So first of all, think security is a huge thing. Security don't think of it as optional for MCP servers.
There are some recent research that shows MCP servers can be exploited basically for credential theft and remote code execution. And if you, if you, if you really understand how MCP code like server work, right there, they're not the most secure servers on the planet. Right. So let's just put it that way. So obviously it's going to be very easy for anyone to, if you don't even have authentication and for anyone to actually visit your server ⁓ and be able to read everything that's in there. So I'm not saying that you should avoid MCP, definitely use it.
audio connectors, right? Use read only permissions first, because some AI, they can actually go into your MCP, go into your MCP server and make edit permissions, right? And that's potentially dangerous. if you, if you do not actually have a copy of whatever you put into MCP servers and don't trust the connector just because it's in the registry, right? Definitely do your own research. I personally will say do not put confidential information in the MCP server just yet before you understand security.
I'm actually trying something out myself as well. So for me, I'm putting MCP server just like our help center articles and something that's like, you know, already available online. So nothing crazy, nothing serious. If I were to put the customer information in there, I will be super careful. I still have to do more research about whether or not that's a good idea. Probably not. And MCP service again, could be exploited. Even if it's just like a 1 % or half percent risk, it's not worth taking it. So obviously don't just trust the registry. You have to do your own research and make sure that.
At this time, don't pull confidential information in the registry just yet. No, in MCP service yet. And also there's also some issues with, you know, performance or reliability, right? ⁓ So it's not going to be a hundred percent success for the MCP service because it's remotely calling MCP service over the internet. So there's obviously going to be more latency. If you use chat2pt, if you use cloud, sometimes you use the thinking mode and you, it has to search the internet. Sometimes you might notice that it's going to take a little bit more time, right? So I think that's normal.
So this expect the same thing for your own MCP servers. If your MCP goes down or does that mean your AI will stop working? So you have to be very careful about that. So if your MCP server goes down, which, you know, it tends to, it has to at some point, right? Does that mean your whole application system falls? Right? Is there any backup? Is there anything that you have backing up the whole system? You need caching, timeouts and fallback logic. That's, know, very technical obviously, but I'm just thinking, I'm just basically saying you have to have a second, you know, backup option.
If the MCP server falls, if you're just using it yourself manually, there's not nothing automated. ⁓ that that's probably fine. This is just like basically real engineering. So I think this week's action, ⁓ basically I, I personally very intrigued about this MCP server as well. So I'm actually going to look into it more as well and be sharing my experience running MCP servers. So for you, probably, you know, stop everything you can, ⁓ you know, and try to build one MCP connector this week, probably use GitHub or something that's already on the market.
and try to connect the MCP server with your cloud, Chatchapiti. So with cloud, I know that you can already actually connect with Google Drive, connect with Notion, different other ones. So if you're using one of those, that's already MCP server, right? So it's not, it's not the fastest experience, but it gives you an idea about what you can build with MCP. So once you're comfortable with that, probably try to also launch something, you know, yourself, you know, with your own MCP server. are many, there are many tools out there that can help you build your MCP servers.
Just go to one of those out there search like MCP server creation. There are many platforms that allow you to do for free. Try it, right? Connect that on the custom, custom integration page on ChaiJPT and also cloud or whatever you're using and give it a try. You know, I will recommend I would get up MCP registry first. That's the easiest, right? Or something as your, your, your AI already has. So I would just read only permissions, audit it and definitely test it before you write. And also connect two different AI agents to the same connector. So for step by step, I would say.
try to check the MCP registration first, registry first. So someone probably already has built something that you already want, right? And if you're using Chess GPT, you're using Claude, there might already be built in Google Drive and other tools that already set up. So you can actually try it and see how MCP is like. Second of all, I think if you're building your own custom, ⁓ know, custom MCP servers, make sure it starts with read only permissions. So you actually can actually audit and test before making it, allowing it to write something, right?
I mean, something can actually be dangerous if you don't know what it means at first. So they can actually change your Google Drive files. can actually like edit things, you know, like that might be something that you're not comfortable with. So make sure you're starting with just like. We only permissions as super, super important. And also just connecting two different AI agents to the same connector and prove the reusability. Right. So two different agents to the same system. Right. So for example, for Claude, you can have a search functionality, kind of a Google Drive functionality and Figma.
And you can have two MCP servers open at same time. And you can give it a command about doing both at the same time to see whether you can actually do two at the same time in the same prompt. That's also super important testing your agents because in our world of like AI agents in general, using AI in general, it's very often for an AI to be able to do two, three, or even four things in the same session, in the same query. So that's definitely very important.
And also, obviously, document everything, document the schema, document your MCP servers, document everything because your future you will thank the present you for doing that. So, you know, if you do this by next week, you'll probably be able to ask HWD and Claude the same questions about your business data, about something that's really relevant to your business. And you can actually get consistent answers from the same source, right? So you're already maybe two times or five times faster and more relevant and more expert on this on this topic.
Then other people were just using the generic and touch BT. And that's when you know, your MCP is working. So the infrastructure obviously shift is starting with a very boring, you know, and sorry, we're very scary. People thought MCP is a scary word, but most important while everyone's debating, which model is the best, right? GitHub just commoditized the whole MCP registry industry. And I think it's a super great chance for any founders, builders, solopreneurs or anyone else to be building something with this new system of MCP.
So I'll be doing that myself as well and posting my thinking. So if you obviously want to learn more, obviously follow me on Twitter. I'm the George Poo on Twitter and also post weekly, you know, insights and also other blog posts on founderreality.com. And this is the Founder Reality podcast. Stay tuned for the upcoming changes that we're bringing more value, you know, to the podcast. And we're going to be sharing more and more tips that's going to be relevant for your business and for yourself. So stay tuned and see you soon.