YouTube Deep Summary

Hey, my name is Greg Boggas. I'm a software developer. I typically make technical tutorials here on this channel, but wanted to do something a little bit different because a lot of my non-developer friends have been asking me about MCP lately. The easiest way to explain MCP is to show you a quick demo. So, here is Claude. I'm going to ask what time is it. Uh Claude doesn't know what time it is. Large language models don't know the current time. They might have the date because it's injected into the system prompt. It just simply does not know the time. But if I come click down here on settings, you can see that I have an integration setup. And integration is Claude's name for a remote MCP server. And inside of there is this tool called current date time that I'm going to turn on. And now when I ask the same question, what time is it. Something different happens. Now you see that Claude realizes that it has access to this tool. It makes a request to the tool. It gets a response back. That response includes the current date time. And so now Claude is able to answer a question that it was not able to do before. This datetime integration or MCP server has given Claude new abilities. I can show you what the code looks like and this is even if you don't know how to write code or read code, this will probably make sense. There is this function called current date time and then down below here is a description or what we call a dock string. And this describes to Claude how to use the tool. And then the code here is effectively three lines. It just uh uses Python to get the current time and then it just if there's an error because the time zone is not right or something, it will kick back an error. But really, it's just these three lines of Python that are powering this tool. So in many ways, this is sort of the simplest MCP server that you could build. But of course, we could replace all of this code with something much more complex. One way I like to think of MCP is let's say you're playing Minecraft. You're walking around with bare hands. You open your inventory and now you have all of these tools and items that you can equip yourself with that give you abilities that you didn't have before. Another more technical analogy would be to think of MCP as APIs for LLMs. And in fact, MCP is how we can wrap APIs and give LLM the ability to use APIs. Let me show you a slightly more complex example. I have my YouTube channel here. I've been focusing on this a lot this year and I have equipped Claude with two new MCP servers. The first is file system access. So, it has the ability to read and write files in a specific restricted directory on my local hard drive. The second is the YouTube data API. I'm going to write a prompt here that asks Claude to do some research on my recent videos and then to do research more broadly on YouTube and to suggest content for me to make uh going forward. And then I want it to take that report and to write it to a markdown file so that I can view it in my Obsidian vault. But it is using these various methods of the API to go do research on YouTube. We are using that to augment our generation. We are creating a report and then we are writing that report out to an external source. In this case, that external source is my local hard drive, but that external source could be Google Drive. It could be an external database. It could kind of be anything. MCP and remote MCP servers in particular has the potential to be the foundation for what will eventually become the equivalent of the app store for chat GBT and other large language models. So you can imagine a world where you're used to signing into a SAS dashboard for instance and you'll no longer want to sign into that SAS dashboard. You're going to want to be able just to chat with the data from chat GPT and MCP promises to be the way in which ChatGpt can communicate with that service. Now, let's talk a little bit about some bad news. One, there are some serious security concerns. My favorite title of any blog post that I've read in the last year is the S in MCP stands for security. It's very difficult to have confidence that the code that is powering that MCP server is secure. Perhaps the developer of that tool could use their MCP server and change the code out from underneath you to exfiltrate data from a private data source. Uh you can also theoretically, not theoretically, but you can actually use MCP servers to do prompt injection. A lot we're going to need to figure out here before MCP servers are considered safe. Next bit of bad news, you kind of need to understand that there's two different flavors of MCP servers. local MCP servers and remote MCP servers. Uh the term server can be a little confusing here because typically when we think of a server, we think of something that's living in the cloud. A local MCP server is actually a script that's running on your local machine. These are the installation instructions for the YouTube data API that I just showed you. In order to use that MCP server, I have to have npm installed. I need to be able to open up a terminal and run some commands. I need to be comfortable editing a JSON file. These things are not particularly difficult, but they can be very intimidating if you're not a developer and you've never done these before. Then there are remote MCP servers. These are a server that instead of having to do all that complex code stuff, you can just drop in a URL to access it. So this is the datetime server that I showed at the top of the video. We'll just push local MCP servers to the side for now and we'll say if a non-developer is going to use an MCP server, it's going to be a remote MCP server. And this is a good time to talk about MCP clients. So, there's two things you need in order to use MCP. You need an MCP server and then you need an MCP client that can talk to that server. Enthropic just rolled out support for remote MCP servers in the web version of Claude. But as of when I'm recording this, which is basically June 1st, 2025, you need to be on the $100 a month Claude Max plan in order to use these custom integrations. So your clients today that have MCP support are Claude and then a lot of coding idees. So Cursor, Windsurf, Augment Code. As of today, chat GBT does not support MCP. And this is the biggest hindrance to widespread MCP adoption is that quite frankly, people just can't really use it unless they're a developer or unless they're paying $100 a month to Claude. Does look like MCP is going to have some real staying power. It's adoption has really been impressive. And the fact that OpenAI says that they will support it in the client. And the fact that they recently actually launched support for MCP in the API, I think is a very strong signal that MCP is here to stay. So if you're a non-developer, what should you be doing with MCP today. As of today, who knows when this video will be out of date. If you are a non-developer and you want to start using MCP servers, the easiest way to do that is to sign up for the $100 a month Claude Max plan and then to go look at the list of remote MCP servers. There is a tiny tiny fraction of remote MCP servers compared to local MCP servers. But I suspect that whatever it is you do for your job or whatever you do it is you do for fun, there's probably some remote MCP server out there that you could use that would give you some pretty interesting productivity boosts as you play with these things in cloud. Once OpenAI supports remote MCP servers in chat GBT, that's going to be the thing that opens up the floodgates. That's going to be the catalyst that will eventually lead to something like the app store. So, I think it's okay to just kind of hang out until that happens and let the industry figure some things out. But once you see that announcement, once you see the announcement that ChatGpt now supports remote MCP servers, that's going to be your signal that a whole bunch of exciting stuff is about to happen with LLMs.