What is an MCP? Model Context Protocol and MCP servers

Hey, my name is Mahan. I'm the CMO of Nairon. My name is Luka and I'm the CEO of Nairon. Luka, in last video we spoke about tool calls, how they give the AI an ability to go out and actually do things on the internet. In today's video, let's talk about MCPs, what they are, and how they actually help agents uh connect to the internet and actually do things a lot easier.

For sure. So the use of an MCP is really just to give uh an AI agent the ability to perform an action on any external uh tools. So for example, Gmail, Slack, any other tools that you're using. Um the AI agent needs to be able to connect. In order to make a connection, you need two things. You need authentication. So do you have the right to use this tool? And then the actual tool calls itself. Do you know how to instruct the tool in order to use it? The problem that existed before the MCP was created was that essentially the LLM would need to figure out how all of these tools work and they all work in different ways and they all authenticate in different ways and they have different tools and there was no universal way of interacting with tools for an LLM. We had APIs which we covered which is essentially the ability for one machine to speak to another machine through a uniform pattern. But the LMS, they don't work like that. And the tools themselves require a lot more than just what an API does. And so the analogy that we like to give is that a MCP is like a USB. Prior to the USBs, we had the same issue. Every device that needed to connect to your computer had a different input. Your mouse had a different input. Your monitor had a different input. Your chargers had a different input. And so we created the USB to be able to essentially connect any device to your computer so that that device knows how to communicate with your computer in a uniform manner. And so you can see over here uh just to take the point home u the math that the MCP solves is essentially if you had four different AI models needed to integrate with five different tools like Gmail, Slack, your CRM, etc. you would essentially need four * 5 20 different connectors and you'd have to custom build all of the connectors. And so what ended up happening is that uh Antropic actually came out with this model context protocol which is MCP to essentially provide a middle layer that the tools themselves provide so that any LLM will be able to use that tool. And so you ended up having just to build nine connectors instead of 20 connectors. And by the way, a lot of the AI agents that we build end up having 15, 20, 30 different MCPs. And you can just imagine that would be 100 plus connectors that you would have to build or you could just standardize to 15 MCPS. And so today, the way that the AI agents interact with these tools are through the MCPS themselves.

Let's get into the plumbing a little bit. Uh how does the AI agent actually know when or how to use the MCP itself? So whenever you start a session, the first thing that's added into the initial system prompt is essentially like a document of how to use each MCP down to very small details of these are the list of tools and this is essentially the description of how to use each tool. Um and so again this needs to be provided by the actual tool providers. So Gmail has a Gmail MCP that Google built out. They have a document and then the AI agent just reads that document. they understand how to use it and they whenever they feel like the user is asking them to use a specific tool or asking them to do a task that requires that tool the AI agent will essentially just call that tool uh run it and achieve whatever outcome that tool provides. So a lot of the times we work with companies that want to build out their own tools and so this is sort of the other way around. If you have a tool, for example, you are a CRM or you provide something like a Slack, um, you would need to build out an MCP that you would expose and then the AI agents would be able to integrate with that MC with that MCP. Some of the examples could be something like, hey, here are a set of actions I'd like you to do. Um, you can do search contacts, you can do send email, you can do get calendar. AI agents a lot of the times will just understand how to use those tools because you built out the docs that way.

It's very important to explain to the AI agent how these tools work together in your system. So for example, if you wanted to send somebody an email, the AI agent first requires in order to send them out an email requires the email address of the person you're sending.

They don't just magically know that. And so in your docs, you kind of have to say first use the search tool to find the person's email, then use the send email tool in order to send out the actual email. And so that's a lot of the consulting that we do and thinking about all the cases that MCP can be used. Um, and then we go ahead and we build the MCP.

And so what's been the overall impact of MCPs on how people can now use AI to get more stuff done? So it went from essentially having a chatbot which could just give you answers to now you have an assistant that can do actions for you. A lot of the very complex AI systems that we have right now and that we build out.

They all essentially use this actionoriented way of interacting with an LLM. Okay. And that's been a summary of what MCPs are. Guys, if you enjoyed the video, make sure to subscribe. You can also subscribe to our newsletter links down in description below. Also find us on LinkedIn. We're very active on a day-to-day basis. And uh we'll see you in the next video.