Ep. 51: WTF is MCP?

[00:00:04] Hello, welcome to Prompting Curiosity, a podcast for the AI curious. No coding background required. I'm your host, Dr. Shantae Cofield, also known as the Maestro, and I created this show to explore what these AI tools actually are. Really, though, are the files in the computer, how to use them, and what they might mean for how we think, work, create, and move through life. Whether you're skeptical, intrigued, or already experimenting, you're in the right place. All that I ask is that you stay curious. All right, let's get into it. [00:00:38] Hello, hello, hello, my curious people. And M, welcome to episode 51 of Ah, prompting Curiosity. I'm your grateful host, the Maestro, and today we're talking about mcp. M. [00:00:49] What the is mcp? So glad you asked. So, if you have been in the AI ecosystem for a bit, there's a good chance you've heard the term mcp. [00:00:59] Um, this episode was actually inspired by Kit, formerly known as ConvertKit, which is the email marketing software that I use. Um, they recently launched their kit mcp, which should actually be referred to as a Kit MCP server, but we will get into that shortly. So they launched that. [00:01:17] I was like, yo, we're seeing MCP everywhere. I've heard, been hearing about it for quite some time now, and I don't really understand what it is. So let me do it a little bit of a deep dive and then share my findings with my people. So here we are. Fair, uh, warning. This episode is going to be mildly, just mildly techy, uh, far less than others that I've done in the past. But it did feel important to me to dive into, you know, WTF MCP is, one, because we're curious people and two, because I think we're going to start seeing and hearing about it a lot more. So let's get into it. MCP stands for Model Context Protocol. The name in this case actually explains what it is, which is super helpful. But you have to understand what each word actually means for it to make sense. So in this case, we have Model Context protocol. MCP model equals the LLM that you're using, the AI that you're using. [00:02:10] Context. That is everything you want that AI model to be able to see and use and interact with right files, data, information, etc. Aka what we're going to call to. We're going to call to our. What we're going to collectively refer to as external tools. [00:02:26] So in this case, I was talking about the Kit, uh, mcp. In that case, Kit would be the external tool. [00:02:32] And the last term here, word here Protocol, very simple protocol is a published set of standardized rules. Thus MCP equals or stands for model context protocol, which means a published set of standardized rules for how LLMs can gain access to external tools. That's it, right? A published set of standardized rules for how LLMs can gain access to external tools. [00:03:05] Where it gets a little bit confusing is that people will refer to this, use the term mcp, but they're actually referring to MCP servers, which we're going to get into a, into a bit. But MCP as it stands by itself, it is literally just the set of standardized rules that, uh, that indicate how LLMs can gain access to external tools, right? So before we get into the nitty gritty, you're even more into the nitty gritty. I think it's helpful to understand why this set of rules or MCP was developed in the first place. So to present things as simply as I can. These rules, right? MCP, model context protocol, they allow your AI model of choice, be it Claude, ChatGPT, what to connect with software and take actions inside of that software. That's literally it, right? Mcp, those rules that's going to allow. Because right now we have to copy and paste stuff, right? So you're like, oh, I want this to like, look at my email, or I want it to look at something. So you have to like copy whatever the things are inside of your email or inside of Kit, and then paste it over into Claude and then have it analyze things. And if you want to make an email, then, uh, it can, you know, it can help you write that email. Don't do that though. I did a episode on that. But let's say you're doing that. You're listening to me. You have to copy it from Claude and then paste it into convertkit or Kit, right? [00:04:29] MCP allows you to get around this and have AI be directly integrated into your software of choice. [00:04:41] And by software of choice, I mean any software that allows for this, right? So the example again, uh, is I'm going to keep using kit, right? Using the kit mcp, you could have Claude write and schedule or send an email to just a select portion of your email list that it has identified as being the most engaged. And you would do all of this from inside of Claude. [00:05:05] Um, so MCP, again, the published standardized rules, MCP was created so that you could connect any AI with any external tool, right? So any AI, Claude, ChatGPT, whatever, with any external tool, Google Drive, Slack, WordPress, Kit, um, any external tool that allows for that connection before this right before MCP was created, 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, right? Uh, and there's lots of AI models. [00:05:48] So to recap here, the whole goal with developing and publishing that standardized set of rules, right, MCP was so that AI would be able to act, AKA do things inside of external tools, like kit. Um, fun fact, a little fun fact here. MCP was actually created by Anthropic. I had no idea until I did this episode. MZP was created by Anthropic, uh, and released in November of 2024. Like I said, I've been hearing about this thing for quite some time. And I was like, this is too rich for my blood. [00:06:20] Uh, but you stick around long enough and suddenly you're like, well, time to learn this. So why did they create it? You know, why, why was it anthropic? I should say it's literally because they had the most direct incentive, right? They are smaller than OpenAI, they are smaller than the other, the other players. [00:06:36] Um, so in creating something like that, their inherent value would increase dramatically if they could connect their AI inside of or directly to the tools that people are already using. [00:06:52] And other, you know, that's, that's incredibly valuable for them. You know, the things that you can think about. If any of you are like, uh, I don't, again, I don't know who, who my user, the listener base is here. But you know, for folks, any of you that use Canva, like, imagine, and they, you know, it's integrated now, but imagine being able to use Claude directly inside of canva. And so you could be in Claude and be like, hey, uh, go to my, you know, canva account and create this, this thing. Um, that's what MCP allows for. [00:07:24] So it was very smart of them to do that and very smart of them to build a standard the whole industry could adopt as opposed to like, you know, proprietary integrations. And one, one by one, it's just like, no, everyone can use this, any AI users, but we start, we started this. So MCP became the industry standard. Remember, the P stands for protocol. So it's a standardized protocol because the developers understood how much more efficient it would be to just use this standardized approach as opposed to building a custom integration for every AI model separately, right? So Anthropic published the spec and then they open sourced it in November of 2024. Um, and then in December the following year, 2025, Anthropic donated MCP to the Agentic AI, um, which is a nonprofit under the. Linux. Linux, Linux, however you pronounce that foundation. Um, and that turned MCP into what is referred to as neutral industry infrastructure. So, like, nobody owns it. Um, worth noting because I was like, did, did OpenAI, like, want to use this? Like their, their competitors made this thing and initially they didn't use it. They adopted it fairly quickly. I think it was like four or five months. But OpenAI had their own thing before MCP, and that was called plugins, right? Uh, and that was how you got your, you. Excuse me, that's how you got Chat GPT inside of software that allowed for it, right? Via plugins. [00:08:47] They adopted, they switched that scrapped it. They, they adopted MCP in March of 2025. And I think this is a fun angle to look at because it speaks to how adoption happens, right? And how hands can get forced. So basically the, the ecosystem built up enough gravity quickly enough that staying that for, for OpenAI, staying outside of it had a real cost, right? Developers are building, and we're going to get into what these, uh, developers were building. MCP servers, and we're going to get into that in a little bit. But developers were building MCP servers and they were expecting them to work with all of the major AI tools, right? So you see these major AI players, they are adopting it. [00:09:26] OpenAI is holding out and suddenly people are like, what the. Right? So if ChatGPT didn't support MCP, it meant that ChatGPT couldn't use this growing ecosystem of integrations that Claude and other, uh, you know, AI models could use. And so they adopted it. So just to get back to the boring stuff to recap real quick before we move into the tail end of the, you know, the back half of this episode, uh, MCP means or stands for Model Context Protocol. It is a published set of standardized rules for how LLMs can gain access to external tools. [00:10:02] We can loosely say, hey, this is the rules for how you're going to integrate an, uh, AI into this software, right? There are set rules for how this happens and how they communicate with each other. [00:10:14] These rules allow any LLM to work within any, any external tool that allows for it, right? And that second part we're going to cover right now, right? So the external tool that you want to integrate the LLM into has to allow for it, all right? MCP is the protocol. It's the rules. It's the rules for how the AI model and the external tool kit I hate. It's kind of tough using that name because it feels like it, like kind of goes into it, like hides within what we're talking about. But so maybe I'll say Convert Kit. [00:10:47] Um, so MCP is the protocol, right? It's the rules for how the AI model for Claude and the external tool in this case ConvertKit will communicate that involves two parties, the AI model and the external tool. [00:11:04] Perhaps you remember from earlier in this episode when I said that the external tool, and by earlier I mean a few seconds ago when I said that the external tool has to allow for that connection. [00:11:15] You can't just go put in AI models and everything without permission. Right? It's not, that's not how it works. It is a two way street here. [00:11:22] So the permission that has to be granted from the external tool, so from ConvertKit, from Slack, from whatever that is known as an MCP server. [00:11:34] Right? So we have mcp. That's the set of rules. [00:11:37] It governs everything. [00:11:38] And now I'm going to introduce two new terms. The first new term is MCP client. That is simply the part of the AI that manages the connection to the MCP server. So AI that side of things that you can just kind of have m them be synonymous with MCP client on the receiving end. [00:12:00] Right. Uh, how ConvertKit allows for this to happen, that's going to be the MCP server. So whether it's ConvertKit, Slack, WordPress, Squarespace, whatever the external tool is that you're trying to integrate into the AI model into, uh, so that you can use AI inside of it, that is going to be the MCP server. In this case, server is being used in the classic computing sense, not like a physical hardware sense. So a server is simply any program that sits and waits to receive requests and then it responds to them. That's it. [00:12:36] So when companies like kit and ConvertKit say that they've launched the kit MCP, what they're actually saying, this is what I was saying earlier in the episode, what they're actually saying is that they created the MCP server and Server is just a program that is specific to Kit that enables any AI models to interact with Kit and be used inside of Kit. Right. Uh, but they call it, you know, the Kit mcp. That's it. They don't put that word server on the end. But for me it is helpful to, to think about it like that. Or you, when you hear people talking about MCPS and plural, making it plural, they're talking about MCP servers, which is the external tool side of things. Right. So we Have. How is this AI talking to the software? [00:13:19] On the AI side, it's called an MCP client on the external software side. So, you know, Kit, Slack, whatever it is called an MCP server. And the protocol that they use, right, the, the rules, um, that they follow, that is mcp. [00:13:41] Okay. [00:13:42] Of note, what that AI model is able to do within Kit, or any external tool for that, for that matter, is limited to the specific functionality that Kit has determined that it wants to allow. So it's not like, oh, well, now it's, it's, you know, connected. And so Claude can just do whatever inside of Kit and they can send things and delete things and go look in files. And that's not the case. It can only do what Kit has decided it wants to allow Claude to be able to do. Okay, um, the last term that I want to introduce you to, you folks are doing great. You're doing great. I know your eyes are glossing over a little bit, but you're doing great. The last term I want to introduce you to is one that I've spoken about in the past, and that's API. Right? An API. API stands for Application Programming Interface. And that is how two software systems talk to each other. It has been around forever. If you want two different tools to work together to talk to each other. For example, having content that's on a Google Doc get turned into a post on your WordPress website like I do, you'd need to access Google's API and WordPress's API. Okay, so maybe you're listening and being like, wait a minute. That sounds like what you just said about mcp. [00:14:57] Very similar, right? What MCP does is that it allows AI to replace you. [00:15:05] So you tell Claude that you want what you want, right? So you're like, I want to have these things connected and I want to, I want this, I want this, um, to be inside of this, Whatever. [00:15:17] Claude's MCP client would communicate with Google's McP server and WordPress's McP server, and then the actual exchange of information would happen via those same APIs as before. [00:15:33] Right. I'm highlighting this process and maybe got a little confusing there, but I'm highlighting this process because I think it really does solidify the role and the value of McP servers, or MCPs as they're more commonly, you know, referred to in that it just allows for AI to communicate and work within and take action within software. [00:15:55] When you just have software talking to software, then that simply uses APIs, and you need a programmer to say like what to do. And it can only do that same thing each time. You'd have to have multiple different things that you'd want to call to get different actions to happen. I've been using a ton of APIs with the automations and stuff that I'm, that I use. [00:16:18] If I want Claude to be inside of any software and then do stuff, that is when you would use mcp. [00:16:27] And then if I'm having, you know, I still want things to talk to each other, I still want actions to occur that is still going to be APIs. Okay, I know it's a lot of acronyms, I know. But the main thing to take away is that MCP allows for AI to work directly inside of software. Okay, I, I, uh, I do believe, and this is one of the reasons I did this episode, that I do believe we're going to continue to see MCP usage can, you know, grow and continue to grow as more people continue using AI and thus more companies want to integrate that AI so that they get more customers. It's all about money, right? It's all about money. And if people can make money, they're going to try and make that money. And so I do believe we're going to see more MCP servers. Um, and you know, again, people will call it like the Squarespace MCP or the WordPress MCP, which does already exist. You're, um, going to see more of those popping up and more official ones popping up. Uh, so you can use these AI tools directly inside of your software. Do you need to be using that? I cannot tell you. [00:17:33] I have yet to have a direct use case for this. [00:17:36] Um, and you know, I watched kits, uh, release and I was like, all right, that's fine. Um, for people that really like data, it can be helpful. Um, and you know, basically you have it look at things and you're like, what was the most engaged? And it can look at links that were clicked and things like that. I'm not a data hound, so like, I don't really care, but, um, if and when I find a better use case, I will let you know. So my goal with this episode and with all the episodes, honestly, is just to introduce you to things. I want to introduce you to this term and explain a bit about the tech. Um, I intentionally chose not to dive into how to connect MCP servers because you can go and ask Claude or your favorite LLM how to do that. Right. So one last thing before I wrap up. Um, and that is the three main concerns when it comes to Using utilizing mcps. So, number one, clearly MCPS give AI the ability to actually execute tasks. And this is the same issue we had when CLAUDE code rolled out. And it's like, oh, it can actually do stuff on your computer that in mind, right? It can delete things, it can send things, it can move things, it can erase things. So be mindful of that and pay attention to, you know, the permissions that you're granting and the things that you're saying it can do. [00:18:50] Second is costs and usage. So having Claude perform actions inside of an external tool. Absolutely. Uses tokens and uses more than if you're just asking it a simple question. Um, and depending on what you're doing, that could add up. So just be mindful of that and be mindful of, you know, just check your usage. [00:19:06] Um, and then lastly, a term called prompt injections. I know another new term. Okay, I know, but you can handle it. We're almost done. [00:19:14] Uh, this is an important concept to understand. So a prompt injection is when someone hides instructions inside of content, like an email or a document or a web page. And those instructions are meant for AI, not you. You can't see them. But AI reading the whole page and it can see them, and then it follows those instructions without you realizing it. And bad things could happen. So, for example, imagine you asked Claude to read an email and draft a reply, but the email was sent by some nefarious person who had embedded hidden instructions in it. Right? The instructions invisible to you, but readable by Claude. And those instructions tell Claude to also forward your contact list to an external address. Something like that. Claude, then just follow the instructions because it looks like a legitimate. Wow. Uh, it looks like a legitimate command. [00:20:05] Meanwhile, you never knew that the command was there in the first place. So again, that is called prompt injection. It has it's worth knowing it exists, Right? It's not new. It has been around. It's always been a concern. It's always been a thing. Um, but it's not. And it's not a reason to avoid mcps, but it is a reason to pay attention to what you're asking how to do. [00:20:24] Uh, and just knowing where the content is coming from that you're going to be interacting with. Okay, all right, that is enough tech talk. [00:20:32] Let's wrap this up and let me share how I used AI this week. So each episode, in case you don't know, I share a quick example of how I used AI that week. This week, I used AI, in this case claude, to learn about MCP and MCPS and create this episode. [00:20:49] Right. I really do love using AI as a learning tool. Um, especially because you can ask it as many questions as you want and it doesn't get tired of you. Right? You can go back and forth as many times as you need in order to understand something and it won't get annoyed if you're like, wait, I don't understand, wait, I don't understand, wait, can you give me that? Uh, in like a different scenario, give me an example of that. [00:21:09] You can ask it to share sources with you, which I always do, and then I go and read those sources. [00:21:15] My favorite question for Claude is, are you sure about that? I stay asking. All right. And then at the end of the day, like you can go after understanding, right? Using your natural language. You can just talk to it. That's absolutely incredible to me. You don't even have to, you don't even have to write. You can speak and it will help you learn. That is incredible to me. So I'm actually going to plug uh, my Whisper Flow link again because when I am doing things like this and I'm trying to learn new topics, I will often opt to use voice dictation as opposed to typing because it's way faster and I don't use the voice dictation that's already inside of it. If I'm on the phone sometimes I will, but I'm on the computer doing this stuff and the native, the native dictation sucks. [00:21:58] Uh, so I use uh, a software called Whisper Flow. Um, as I said, I have a, uh, I have an affiliate link for that. [00:22:10] You m. Know for me using voice dictation as opposed to typing, it's just way faster. So if you don't know. Whisper Flow is an AI powered voice dictation software. I've been using it since like last December. I um, did a whole episode on it. It was episode 45. You can check that out. I will link that in the show notes if um, you're interested in learning more. But like I said, I have an affiliate link that will get you a free month of Pro. I think it's like 11amonth or 15amonth, something like that. Uh, it's less than 20amonth. Um, but it will get you a free month of pro if you sign up. And then if you do, when you hit those 2000 words, I get a free month. And a few of you have hit that. This is really cool. I'm like, wow, that means you folks are listening and getting it from somewhere. I don't know if it's from this podcast or other places that I've shared it, but if it is from this podcast, thank you. It's really, really dope to me. So, uh, I will link that in the show notes. I can't tell you the link because it's like, not. You're not gonna remember it, but I'll link it in the show notes. So, uh, that, my friends, that is all for today. I know it was a little bit more techy than usual. Uh, I still think it's medium tech. Um, but my goal was just to introduce you to this, to this concept. So it's not. [00:23:16] So it's not super far into you. Right. [00:23:19] Hopefully you found the episode helpful. If you did consider sharing it with someone you know who's curious about AI, I am going to ask. This one is a more dense one and it might scare some people. But, uh, for folks that, you know, are a little bit more tech inclined and they want to get into the weeds a little bit, maybe share this episode. All right, all right. Don't forget, I have a companion newsletter and blog, the Curious Companion, that drops every Thursday. That is basically. And by basically I mean exactly the podcast episode in text format. So if you prefer to read, and maybe this one is a good one to go back and read and look through, uh, or you just want a written record of things, you can join the newsletter, you can check out the blog. It's always going to be there. Head to promociosity.comnewsletter or forward/ blog. Or you can just keep it simple and check out the link in the show notes. [00:24:06] As always, endlessly, endlessly. Let me give it to you one more time. Endlessly appreciative for every single one of you. Until we chat again next Thursday, stay curious.