Introduction to MCP and Its Role in the Future of AI Agents

[00:00 - 00:17] Okay. Okay. So let's go. So today, we are going to discuss MCP. I believe this is like future for the AI agents. And we will look how it will change the way. Basically, we build an AI apps and why exactly we need to learn it and preferably as soon as possible.

[00:18 - 00:34] But before we ever start with the MCP, we need to cover one small topic, not so small, maybe, but about the AI agents. So are you familiar with the agents? Have you worked on when we heard about this? If no, yeah, basically, that's great.

[00:35 - 00:59] So just in case I would read, you're going to use a couple of details, in short , AI agents, those are AI applications that need to perform work autonomous lives. And what word autonomously is basically the key here, right? Because if we give it some task and give it some instructions, this is all that we need basically to do from our side.

[01:00 - 01:26] And after that, we expect that AI would search the information, would go through information, try to change it maybe in some way and do everything from beginning to start autonomously and would give it just a result. And it's important to know about AI agents, at least briefly, because MCP is really great in giants a lot specifically when you're building AI agents up and especially complex AI agents up.

[01:27 - 01:39] Okay. And then how about MCP? Have you heard about MCP? Or at least something maybe, if no, that's completely fine, but just interesting. Oh, okay, that's great. Okay. So we're on the same page.

[01:40 - 01:53] And I would start from the basic basics anyway, and we would go then in more deeper stuff. In short, we are going to call today three main topics, basically what is MCP?

[01:54 - 02:01] Why exactly you need this? And you need to implement it as soon as possible, potentially. And exactly how to use it.

[02:02 - 02:08] We would go over the really basic MCP application. We would write it from scratch.

[02:09 - 02:23] And it would give you the most basic level of the understanding, like the core level from which you can build up even more complex apps. So it would be small, but really powerful and good application for learning, basically.

[02:24 - 02:32] Okay. But before we move forward on, like a couple of words about myself. My name is Maxine. I'm basically from Ukraine.

[02:33 - 02:36] That's my name might sound a little bit strange to you. You guys, you can call me Max.

[02:37 - 02:42] No problem with that. And I actually have started initially as the full-stack developer.

[02:43 - 02:47] My initial language used to be a JavaScript. It used to be still JavaScript.

[02:48 - 02:58] But then as actually AI started to emerge, I started actually learn more about it and switching to the Python side. And started to learn and build more location.

[02:59 - 03:01] And it's basically my second nature by now. Okay.

[03:02 - 03:18] I've authorized virus A based articles. I'm starting for the deep-seak and finishing this MCP, for example, and builds multiple different AI applications, including candidate evaluation, content generation, see optimization, different chat apps, so on, basically.

[03:19 - 03:36] So we'll be with no one I'm talking about. And before we even start to explain MCP, just in case I also want to address the difference between the classical implementation that working right now with, right?

[03:37 - 03:49] And and like then we moved to a towards the MCP right now, all the agents from the more basic one to more most complex one are actually built. Usually we just direct APIs, right?

[03:50 - 03:53] We are directly managing APIs. We are writing with ourselves.

[03:54 - 04:08] And this is like brings, first of all, like issues in terms of the supporting the code, right? And also how easily we can connect or disconnect the different services to our AI application.

[04:09 - 04:13] So it's just kind of rigid way of building apps. It's still working.

[04:14 - 04:20] It's completely fine. But it's not really that much flexible and actually easy to implement and easy to work is right.

[04:21 - 04:29] And here's where the MCP actually. So NCP once again, model context protocol in like first and the foremost.

[04:30 - 04:34] And it's actually the protocol. Yeah, it's a protocol.

[04:35 - 04:43] And you can think of it as first of all, this protocol is built by anthropic. So on Tropic, it's pretty a large company.

[04:44 - 04:50] So it's not something that's couple of guys decided to go out on the weekends and basically share with the world. Right.

[04:51 - 04:54] It's actually picked up by the big company. And it is open source.

[04:55 - 05:09] Not only that, since it's a big company, it will be supported for a long time and the more and more like popular different applications and companies start to actually implement it inside. So it's not just a buzzword.

[05:10 - 05:16] It's here to stay. And I really hope that it will soon become the form for a lot of applications.

[05:17 - 05:27] Then also is worth mentioning that it's still emerging and new technology. So first of all, whatever we are going to call here is all the basics will be there.

[05:28 - 05:36] OK, there might be a single exchange changes in the future. It might be like small documentation changes, but the core principles and idea behind it.

[05:37 - 05:39] MCP, they will be the same. So don't worry about this.

[05:40 - 05:46] After we go through everything, even the docs will change, you would understand everything. So that's completely fine.

[05:47 - 05:53] But else of since this is new technology. Right now, it might be sometimes not really stable.

[05:54 - 05:57] I would show you later on. And there are some hiccups here and there.

[05:58 - 06:07] Why sometimes it's not like the greatest, but guys are constantly working on it and constantly trying to push new stuff. So it like it's just matter of time when it just gets better and better.

[06:08 - 06:17] The next point is actually what the problem solves. So it just may be you short about the MVC model view controller.

[06:18 - 06:24] This is the way how we organize like back and upside. And MCP, it's kind of the same.

[06:25 - 06:43] Even though it's a protocol, it's basically the unified way for us to agree on how we actually build the AI apps and how we connect external services or external processing power to our AI apps. So this way would all would be on the same page and it would be super easy for us to build and share information and stuff like that.

[06:44 - 06:50] Right. And before that, once again, it was the API implementations, it's just like a wild west, right?

[06:51 - 06:53] You can build whatever you want the way you want. So it starts to maintain code.

[06:54 - 07:10] For example, even me, when I see my own goal that I have written, I don't know, six months and year before I can be confused. So it's good to know that there would be some pattern which would help you actually to get the idea about that and start to work and visit faster, basically.

[07:11 - 07:14] And what is it? More details.

[07:15 - 07:18] You can think of it as the type of C. I really love this analogy.

[07:19 - 07:31] Basically, we are providing the one unified method of connecting our client and the backend side. So you can have like multiple MCP servers, right?

[07:32 - 07:41] And they all connect in one simple way that is predictable and which is easily understand by us. OK.

[07:42 - 07:51] We can think of the MCP client and because no worries, just a couple of slides that explain difference between them. Not nothing to just to be scary about.

[07:52 - 08:03] And then we have multiple MCP servers, which just like use this, which are connected via this specific, like protocol, this jungle, like in one easy to go or manner, no matter the application. All right.

[08:04 - 08:32] And one of the also benefits, not only we can easily interchange, like and play different MCP servers, but we give ability to our application to actually dynamically choose which MCP server to work. OK, I'm going to explain this process a little bit more later on, but for you, just to journal, understand the concept idea why it's so beautiful and great.

[08:33 - 08:53] AI clients has access to all the descriptions of these different MCP servers. And as we are giving some instructions, it dynamically can figure out what is the service is the best for the job and even if it needs one, right, to make a job to do a job.

[08:54 - 09:05] And then it just chooses automatically without our any like permissions and it starts to perform the task. And with the API solution, it was like much more rigid.

[09:06 - 09:13] We had to record it or write additional a lot of logic to actually make it happen automatically. For us, it's just plug and play.

[09:14 - 09:19] So just to wrap it up, what why it's so beautiful. First of all, it's flexible.

[09:20 - 09:31] You can switch servers like like that super easily. Not only that, once you read in the server or some have also have created server, you can super easily share it.

[09:32 - 09:43] So you can just go to the web, download someone's server and just with providing a couple of API keys, connect to your application, like super fast, super easy. Once again, yeah, we can choose this server automatically.

[09:44 - 10:02] And also it helps us to abstract a lot of complexity of actually how to like connect service and stuff like that, like inside of MCP servers. So the core logic would be there, but we have started in that way so we can like easily share it and connect to whatever.

[10:03 - 10:18] And yes, about sharing and since we are able to share those servers, I want to share with you right away the link. This is official document official GitLab repo in this case.

[10:19 - 10:29] From the MCP server, I dropped it in the chat. This is actually already a bunch of different servers that you can download it and use it right away.

[10:30 - 10:41] Okay, and we can see some of the big names there. And you as easily can just drop your servers for someone, he would be able super easily to pick this up and to follow.

[10:42 - 10:58] And when we are talking about not service, but actually about client side, because usually for the client side, we're going to use some established application like clothes, cars or something like that, right? It's start to be more popular as well. So let me draw another link for you.

[10:59 - 11:09] This was interesting in the future. We'll get back to this second link. And you can see right now how many companies start to actually integrate the M CP in their infrastructure.

[11:10 - 11:23] Okay, because we need once again, MCP both on the client and on the server side to work. And guys, like one or two months before this list used to be like 50% of its size.

[11:24 - 11:33] So it's grew twice as much just in couple of months. And I recently heard that OpenAI actually plans to implement MCP into their services.

[11:34 - 11:46] And that would be huge because they just additionally skyrocket everything even further more, basically. So in short, this is how it is on really like top level.

[11:47 - 12:20] And right now, by now you should be able to start to understand the difference basically between the just classical approach and the MCP one, because with the MCP, we have something like a middleman, which helps us to connect whatever API we have with whatever client, a long choice we have, which is available for us right now. Super easy to connect disconnect and also when it's connected to the LLM, it helps it gives the LLM right away dynamic knowledge about all the service, which is connected to it.

[12:21 - 12:25] It just performs faster, better and it just is much more easier to manage. So it's just beautiful.

[12:26 - 12:27] (chuckles)