What is MCP?

Maybe like me, you have overheard people sounding terribly clever, saying stuff like “Well I just connected the MCP server” or “First we need to ensure the MCP client is available” and thought: HUH??

This must be suuuuper complicated and technical. 

But actually, it’s really not. 

MCP is a very simple and I personally think, a very elegant topic, that once you have learned it, makes absolute sense. 

Let’s dive straight in with: 

MCP stands for Model Context Protocol and it is a set of rules and standards which define how AI applications can communicate and connect with external software.

Thanks to MCP we can do stuff like: 

Let’s look at it in more detail:

MCP is made up of three bits: 1. Model 2. Context 3. Protocol

Perhaps you hear “protocol” and it brings to mind other protocols you may have encountered, like

The word “protocol” simply means a set of agreed rules that defines how two or more parties communicate.

In this case, how the AI model communicates with external software.

Still finding the concept tricky to grasp?

I find it helps to think of INTERNATIONAL DIALING CODES.

At some point - don’t ask me when - the whole world decided “if I want to call Germany, I will add a +49 to the phone number”

The whole world was like “Yep, that’s a smart idea and makes everyone’s life easier”

We all agreed going forward, we add +49 when calling Germany. 

Alles gut. 

This is a protocol, a set of rules and standards that we internationally agree upon and it provides an elegant solution to a real problem.

To summarize this chapter: 

MCP is not a piece of software or a tool or anything particularly scary. It is just an agreed upon set of rules and standards. 

In this case describing how AI models talk to external software

A fair question.

Why do we need this set of rules and standards in the first place? 

To answer that, let’s look at a real life situation:

I am using Claude to help me respond to emails at work. Cool stuff.

Now I want Claude to go one step further and automatically get the email from my account and help me respond in one go. 

To do this, I will need Claude to be able to connect with my Gmail account.

But when I try, I get this error message:

The problem here is: AI models (like Claude, ChatGPT) have no native connection to other external software. 

Think of them like little bubbles of smartness. 

Very clever. But with no information outside of their bubble about your life; they only can see what you tell them to see.

This limits how useful an AI model can be.

How can I build a cool AI app that organises my calendar for me, if the AI can’t talk to my calendar app? 

How can I use AI to auto-reply to emails if the model can’t access my email. 

We need to connect tools to get the most out of them.

Preferably simply and quickly

Before MCP, it was possible to connect AI models with other tools. But boy was it painful! 

You would have to write the code for this yourself, for EVERY combination of connections. 

Imagine, having to develop integrations yourself for every combination of Gmail/Outlook/Claude/Notion/Excel…. Argghhhh

Even if you were a coding genius, all it would take is for Gmail to change a bit of their API policy and bam, your connection would break.

So the smart folks at Anthropic proposed a solution that the whole AI world got behind.

MCP.

MCP replaced all of this mess with one protocol that any AI model can use to connect to any tool.* 

(any tool with an MCP server - see below) 

Build it once, works everywhere. 

Still with me? Ok lets go: 

We know that MCP simplifies the connection between AI and other tools. 

But how exactly does it do this? 

Well, first we need know that there are three main components of MCP:

Stay with me, because we are going to get a little bit technical…

MCP Host: This is the AI application or model that you are trying to connect to something. E.g. Claude, ChatGPT, Gemini. The MCP client runs here.

MCP Client: This is the helpful connector that provides a means of communication between the MCP Host and an MCP server. 

The good news is that MCP clients have already been built into common AI applications, such as Claude or Gemini.

You can write your own MCP client to communicate with an MCP server, if this isn’t provided for you.

Important to know:

MCP Server: The external tool or service that enables us to connect to a specific external tool.

Still struggling to understand Host, Server and Client? 

Imagine an old lady (MCP Host) speaking with a travel agent (MCP Client) and wants to book a stay at a foreign hotel in German (MCP Server). 

The old lady asks for some help in communication with the foreign hotel since she doesn’t speak Spanish and doesn't know the local customs. 

The travel agent however is great at communicating with foreign hotels. It’s their job after all. 

The travel agent calls the foreign hotel and gets an automated message back asking:

“Our phone services include the following: [book a trip], [cancel a trip], [book an all inclusive package holiday with jetskiing included]”

The travel agent goes back to the old lady and asks what she would like to do.  

The old lady says “book the jetski package”. 

The travel agent communicates with the foreign hotel and asks it to book the jetski package for the old lady. 

OK maybe this example was a bit out there, but I hope the basic roles of the different components are a bit clearer: 

At a very high level this is what the different components are doing:

This was my post on MCP, I hope it helped! 

I will be doing a more technical deep dive on MCP servers in the future, including how to build your own. 

Subscribe for more fabulous content and art work!

The most important bits from this post: