Curious Companion: WTF is MCP?

LISTEN TO THE EPISODE

In this episode I break down MCP (Model Context Protocol): what it is, why Anthropic created it, and why it matters for anyone using AI tools. I cover MCP servers, the two-party connection, the API relationship, and the three main concerns to keep in mind when using MCPs. Fair warning: this one is mildly techy, far less than past episodes, but I do believe MCP is a term we’re going to start hearing a lot more, so we’re getting curious and talking about it now.

WTF Is MCP?

WTF is MCP? So glad you asked.

If you’ve been in the AI ecosystem for a bit (and even if you haven’t, because these companies are shoving it down our throats), there’s a good chance you’ve seen the term “MCP”. This episode was actually inspired by Kit, my email marketing software, sending out eleventy billion emails announcing the launch of their Kit MCP (which should actually be referred to as a Kit MCP server, but we’ll get into that shortly).

Fair warning: this episode is going to be mildly techy, but far less than others I’ve done in the past. But it felt important to dive into wtf MCP is because

We’re curious people
I think we’re going to start seeing and hearing about it a lot more

So let’s dive in.

MCP stands for Model Context Protocol. The name actually explains what it is, which is helpful, but you have to understand what each word actually means for it to make sense. In this case:

Model = the LLM
Context = everything you want the model to be able to see and use and interact with (files, data, information, etc.), aka external tools
Protocol = a published set of standardized rules

Thus: MCP = Model Context Protocol = a published set of standardized rules for how LLMs can gain access to external tools.

Why MCP Was Developed

Before we get into the nitty gritty, it’s helpful to understand why this set of rules was developed in the first place.

To present things simply, MCP allows your AI model of choice (Claude, ChatGPT, etc.) to connect with software and take actions inside of that software.

Example: using the Kit MCP to have Claude write and schedule (or send) an email to just a select portion of your email list that it had identified as being the most engaged. (Note, you would be prompting this from the Claude interface.)

MCP (again, MCP is just published standardized rules) was created so that you could connect ANY AI (Claude, ChatGPT, etc) with any external tool (Google Drive, Slack, WordPress, etc.) that allowed for that connection. Before MCP, if you wanted to integrate AI into an external tool that supported integration, a developer would have to build a custom connection from scratch for each AI model.

So, to recap: the whole goal with developing and publishing MCP was so that AI would be able to act (aka do things) directly INSIDE of external tools like Kit. No copy-pasting or switching between browser tabs. The AI would be INSIDE the software.

Fun Fact: Anthropic Created MCP

MCP was created by Anthropic and released in November of 2024.

For those of you like me who wondered why Anthropic would create this universal connector that could be used by other AI models, it comes down to encouraging adoption. By creating a connection standard that could be used with any AI model, Anthropic was encouraging software companies to build the corresponding MCP servers that would allow for AI integration, irrespective of model loyalty.

The universal standard meant that any Claude user would get access to every integration, even the ones built by companies that preferred a competitor AI model. More integrations means more value for Claude users, which ultimately means more reasons to choose Claude.

Anthropic published the spec and open-sourced it in November 2024. And then in December 2025, Anthropic donated MCP to the Agentic AI Foundation, a nonprofit under the Linux Foundation, turning MCP into what’s known as neutral industry infrastructure.

The Industry Standard

MCP became the industry standard (remember, the P stands for protocol) because developers understood how much more efficient it would be to use one approach for everything as opposed to building a custom integration for every AI model separately.

Worth noting, OpenAI had their own thing before MCP, called plugins. They adopted MCP in March 2025. I think this is a fun angle to look at because it speaks to how adoption happens and how hands can get forced. Essentially the MCP ecosystem built up enough gravity that staying outside it would have been detrimental to OpenAI. Developers were building MCP servers expecting them to work across all major AI tools. If ChatGPT didn’t support MCP, it meant ChatGPT couldn’t use a growing ecosystem of integrations that Claude could.

MCP Servers Explained

Let me do a quick recap before we move into the tail end of this topic:

MCP = Model Context Protocol = a published set of standardized rules for how LLMs can gain access to external tools. These rules allow any LLM to work within any external tool that allows the connection. That second part, “that allows the connection,” is what we’re going to cover now.

MCP is the protocol. It’s the rules for how the AI model and the external tool communicate. This involves two parties: the AI model and the external tool. The external tool has to ALLOW for that connection. You can’t just go putting AI models in everything without permission.

What grants the permission is known as an MCP server.

Two new terms for your bag of tricks:

MCP client: the part of the AI model that manages the connection to MCP servers
MCP server: a program that lives inside an external tool or software that allows an AI to communicate with and operate inside of it

In this case, “server” is being used in the classic computing sense, not the physical-hardware sense. A server is any program that sits and waits to receive requests and then responds to them.

So when companies like Kit say they have launched the Kit MCP, what they’re actually saying is they created an MCP server specific to Kit that enables any AI model (specifically the MCP client of that AI model) to interact with Kit and be used “inside” of Kit.

Worth noting: what the AI model is able to do within Kit (or any external tool) is limited to the specific functionality that Kit has determined it wants to allow. The company controls what is exposed and what the AI can and can’t do.

The API Connection

The last term I want to discuss (I know, deep breath, I believe in you) is one I’ve spoken about in the past: API.

API stands for application programming interface and it is how two software systems talk to each other. If you want two different tools to work together, for example, having content from a Google Doc get turned into a post on your WordPress website, you’d need to write a script that has instructions to access Google’s API and WordPress’s API.

What MCP does is allow AI to replace you in that scenario. You tell Claude what you want, and Claude’s MCP client communicates with Google’s MCP server and WordPress’s MCP server, and then the actual exchange of information happens via those same APIs as before.

Hopefully this solidifies the role and value of MCP servers, or MCPs as they’re more commonly referred to. It is literally both a set of rules (MCP), and a program (MCP server), that allows AI to act inside of other software and external tools.

Three Main Concerns With MCPs

One last thing before I wrap this up: the three main concerns I believe exist when it comes to utilizing MCPs.

MCPs give AI the ability to actually execute tasks (delete things, send things, move things), so be mindful and pay attention to permissions.
Costs and usage: Having Claude perform actions inside of external tools absolutely uses tokens, and depending on what you’re doing, it can add up.
Prompt injections: Yes, I know, another new term. But an important concept to understand. A prompt injection is when someone hides instructions inside content like an email, a document, or a webpage that are meant for the AI, not for you. The AI follows these instructions without you realizing it, and bad things could happen.

Imagine you ask Claude to read an email and draft a reply. But that email was sent by a whack person with bad intentions who had embedded hidden instructions in it, invisible to you but readable by Claude, that instructed Claude to also forward your contact list to an external address. Claude would follow the instruction because it looks like a legitimate command. Meanwhile, you never knew that command was there in the first place.

This is called prompt injection, and it’s just worth knowing it exists. It’s not new and it’s not a reason to avoid MCPs, but it is a reason to pay attention to what you’re asking Claude to do and knowing where that content is coming from.

How I Used AI This Week

Each episode I share a quick example of how I used AI that week.

This week I used AI, specifically Claude, to learn about MCPs and create this episode.

I really do love using AI as a learning tool, particularly because you can ask it as many questions as you want and it doesn’t get tired of you. You can go back and forth as many times as you need to understand something and it won’t get annoyed with you. You can ask it to share sources with you and then go and read those sources (strongly suggest you do this!). And you can do all of this with everyday language. That’s absolutely incredible to me.

I’m gonna plug my Wispr Flow affiliate link again here because when I’m using AI in this way and trying to learn new topics, I will often opt to use voice dictation as opposed to typing. It’s just way faster and easier for me. If you don’t know, Wispr Flow is an AI-powered voice dictation software. I’ve been using it since last December and did a whole episode on it (episode 45), so check that out if you’re interested in learning more.

If you use my affiliate link to signup, you’ll get a free month of Pro, and when you hit 2,000 words, I get a free month too! A few of you have already done that and hooked me up which is super dope, so, THANK YOU.

Da Wrap-up

MCP (Model Context Protocol) is the published set of standardized rules that allows any LLM to connect with and act inside external tools, and MCP servers are the programs companies build to make that connection possible. It’s a concept that’s only going to become more relevant as AI usage grows and more platforms start integrating it, and I like keeping you ahead of the eight ball.

I realize this episode was heavy on the what, and basically completely devoid of the how. That was intentional. I just want you to know what it is and that it exists. If you’re ready for the next level and want to actually connect an MCP server, your best bet is…to ask your favorite LLM how to do it. What a time to be alive.

As always, endlessly grateful for you and your curiosity.

Catch you next Thursday.

Maestro out.