Design Your First MCP Server: Connect Claude to Any System

Charlie Cowan: Well, hello, let me, share my screen. There we go… Welcome to today's Wednesday webinar. probably about the 20th or 30th in our series, but if this is the first of the Wednesday webinars from Kowalah that you've joined, we get together every Wednesday at the same time, and we have an hour to spend together, and… talk about one, topic, or feature, or way of working with AI, that we're certainly experiencing, either for ourselves or with our clients, and that we think is useful for you to learn about and to be able to share, within your own, organizations. It's such a fast-moving space that we're in, that just taking a little bit of time to keep on top of what is going on is so valuable. And, we will go through our normal, sort of three-step process.

Charlie Cowan: So, to start with, we'll talk about our main topic of the day, which I'm going to introduce in a moment, which is MCP servers. And this might take us, I don't know, 20, 30 minutes or so that we're talking about this main topic. Then we'll take a look at AI in the news. Like I say, there is so much going on, so many new releases, moving parts, that it's difficult to stay on top of it, and so we will digest all of the new releases of the last 7 days, and walk you through what matters for your own business. And then finally, there is an opportunity to ask any questions that you might have. Now, these questions can be about the topics that we are discussing today, or it could be anything. Maybe you've got some advice that you need about building a new skill for your HR team, whatever it might be.

Charlie Cowan: But don't feel that you need to wait until the end of the session to ask your questions. If you've got questions as we go through, you've probably used Zoom before. You've got both a chat interface, and I'm just gonna move this up onto my screen so that I can see it all the time. So there is a chat where you can chat, and then others can see, and then there's also a Q&A bar, so let me open that as well, and then I'll keep an eye on both of them. And so, don't feel that you need to wait, just pepper your questions in, and I'll try and keep an eye on both of those. And if not, my lovely human colleague, Caitlin, who's on the call as well, will keep an eye and keep me honest if I miss your questions.

Charlie Cowan: So with that, let's dive in to the topic of the day, which is MCP servers. And this is probably one of the most technical topics that we've covered in our Wednesday webinars. We spend a lot of time talking about prompts and skills and business-focused topics. And this is a bit more technical. But I'm going to talk about it in a way that I hope is useful in a business context, because as a business user, a knowledge worker, a business leader, understanding what these are and how you can use them is going to be so valuable to what you're doing, both in your business and then for your own customers as well. And so you can then go and have an educated conversation with your technical teams about, why you want one, and how it's going to be implemented. So, let's go back to basics.

Charlie Cowan: This is a safe zone where everyone can ask the most silly questions that they want. So, firstly, Charlie, what is MCP? Well, I don't know how technical you are. I started my career working in email security, and you may know that SMTP is the protocol for how email travel around the Internet. Simple Mail Transfer Protocol. So you might not have heard of that one, but you probably have heard HTTP, which is the way that web traffic moves around the internet, and you'll know that from your browser, http, colon slash slash. Well, here's a new one to add to your dictionary. MCP, and it stands for Model Context Protocol. And it is the protocol, the standard, for how AI tools, like a Claude, like a ChatGPT, like a Gemini, how they talk to other systems. So there's a few words in that.

Charlie Cowan: So, model, this is the AI model, the Claude, the ChatGPT, the Gemini, so that's the model. The protocol is, you know, this is the standard, but the important word is this one in the middle, context. How do we get context into the model so that it is better able to help us with our question? So, if I was to ask, for example, you know, what meetings have I got tomorrow into Claude. Well, for Claude to be able to answer that question, Claude needs access to my calendar to be able to then answer that question and go, Charlie, you've got this meeting, and you've got that meeting tomorrow. So it's the context. How do we bring that into the models? Now, a little bit of history, one slide, don't worry too much. But the Model Context Protocol, MCP, was created by Anthropic. Anthropic is the company that created Claude.

Charlie Cowan: But they've turned it into an open standard, because they recognise that if everyone follows the MCP standard, then it's going to be much easier for them and for their customers as well. So, if you want to geek out on this, or to share it with others in your technical team, then ModelContextProtocol.io has got you covered, and it explains in great detail about what the protocol is, about the standard, about how you create these things. And so this is a great place to go if you want to get more into the detail of how these things work. So now let's go into, like, what is it, and how does it work? And then we can have a think about what that might mean to you and to your business. So there are really 3 main aspects of Model Context Protocol. And the first is a thing called MCP Client.

Charlie Cowan: And what the client is, is the AI model that you're having a chat in. So it could be Claude, it could be ChatGPT, it could be Gemini, it could be Copilot, it could be Grok, it's the place where you're having the chat. And that might be in Claude or ChatGPT, or it could be you've built something on your website that is communicating with the Anthropic API. But think about this as the place where the work is happening, either as a human is having a conversation with the AI, or maybe you've got a system or an agent that's having a conversation with the AI. But this is the MCP client. That then interacts with what is called the MCP server. And this is where the guts of the MCP protocol works. And it does two things, mainly. Firstly, it lists a registry of tools. Like, what is it that this MCP server can do?

Charlie Cowan: And I'm going to give you some examples of this in a minute. But what can it see, and what can it do? And then the other part of it is actually executing it. So it's deciding which tools to call, and when, and what to do with that information, so it's kind of the brains of the MCP server. And then the third aspect of the MCP chain is your source systems. So, you can think about these as the things that already exist that hold your data. So you might have a customer database that involves all of your customer information, and their products that they use, and their usage, so that might be a useful source of context. You might have external APIs, so this could be out to something like your Google Docs, or to your calendar, or to your call recordings. So, third-party systems that you want to connect your MCP server to.

Charlie Cowan: You might have files and docs, that could be in SharePoint, it could be in Google Drive, could be something else that you think would be useful context. And then you might have other internal tools, so you might have a CRM system, you might have an accounting system, something like that. All of these are the useful context for your MCP server. Now, when I think about what would make a good source system, I always ask myself, like, okay, if I was the human, and I was going to be asked to answer this question, what systems or information would I need as a human to do that? Well, you know, if you were gonna ask me to plan for a meeting with a client next week, I would need access to my calendar to work out what availability I've got.

Charlie Cowan: I would need access to my CRM system to go and find all the details about that customer. I would need access to our product system to be able to go and work out what their usage of our product was. I'd want access to all of my Granola call recordings, or transcripts, and my email. So, as a human, I would need access to all of these systems, so that's what we need to give Claude or ChatGPT access to via this MCP server. So, having created one, I'm going to show you a little bit about how do we actually use an MCP server? So let me dive across into my Claude here. And I'm gonna go down, if I can click here, to my connectors, and my organization settings, and I'm gonna go to connectors down here. So, a connector is another way of saying MCP server.

Charlie Cowan: MCP is the protocol, kind of the motorway, and the connector is, like, what are we connecting it to? And what you're gonna find is that pretty much every tech company has worked out they need to provide these connectors. They need to provide these connectors, because if they don't, their customers are going to start using something else. I want to be able to chat in my Claude about my Canva designs, about all of our DocuSign contracts that are out with customers, about all of our code for the products that we're building, email, calendar, drive, Granola transcripts. You know, these are the things that we want to be able to interact with. And one of those examples is the Kowalah platform. You know, I want to be able, as a member of the Kowalah team, to be able to talk to the Kowalah platform.

Charlie Cowan: So, if I just have a look at what this is, see the details… for an MCP that you create, there are some different ways of connecting these, so I'm going to keep it fairly simple. One of the most common is that you're going to come up with a web URL that you're going to point this connector to, and so in our case, for our internal employees, it's employeemcp.kowalah.com. And when our users connect to this, they're going to authenticate, and it's going to give them access to everything that's in the Kowalah platform. Exactly the same way here, with these pre-built ones that come from tools like Canva, or DocuSign, or Google, they're pre-built by the vendor. As a workspace administrator, you're gonna connect to it. So this is from the company perspective.

Charlie Cowan: Now let me just switch over to my personal settings, and you'll get a view a little bit more about these connectors and what they've got access to. So, let me go to the Kowalah one here. So, this again is for our own internal employees, but if I click on Configure here, you're gonna start to see that we've exposed all these different tools to allow our employees to have conversations about the Kowalah platform in Claude. And we've got a number of different types of tools. So, we've got read-only tools. And we've exposed a lot of the different entities or objects that are in the Kowalah platform. So we've got decisions, we've got expert requests, we've got initiatives, milestones, projects, all of these kinds of things. And so each of these tools tells Claude, you can get a decision, which is an object in there. Oh, you can get information about a project.

Charlie Cowan: You can list what decisions have been made. You can list what initiatives exist. So this is a lot of reading information from this external system. And then you can also have more powerful write and delete tools. So, in this scenario, we've only given 4 to the MCP at the moment, which is to create a new expert request, to create a new task, or to update an expert request, or update a task. So it's useful just to have a think about the types of tools that you want to expose, depending on the types of activities that you want the human to be able to do when they're chatting in Claude. So, we've talked about setting up the connectors, we've talked about the different types of approval that you might be able to have. And then we've talked about the different types of tools that you might provide to that MCP server.

Charlie Cowan: Now, I'm just going to take a little bit of a gamble here. This is a bit of a gamble, because I literally built it 20 minutes before we started our webinar. And I want to try and add in a new MCP server, so you can see what the process looks like when you're connecting a new one. So, I'm going to go back to my organization settings here, and I should caveat, there is a very high chance that this doesn't work, but this is what I love doing on these webinars. It's much better to show you what's actually happening. So, that's the Kowalah one there. I'm gonna go to add a new one, a new custom one, and a new web one. And I'm going to call this Kowalah Client. And this is a new one for our clients to be able to access.

Charlie Cowan: And, let me just ask my Claude here, what is the end point for me to add into Claude to connect. Just ask my other little Claude what I need to put in here. Okay, here we go. And so this is in here. So the previous one that I showed that's already in there is for our internal team, our employees, and then we're just launching one, which is to give our clients access to their own data. And so, what we've created is this remote URL. I'm going to call it Kowalah Client in here. I'm going to click Add. And that has added a custom connector that's down here, Kowalah Client. And now, as an individual user, I can now go to my connectors, and it's gonna show me that this is not connected yet. And I'm gonna click Connect. And… it couldn't reach it.

Charlie Cowan: I told you there was a very high chance that wouldn't work. But what I will do after this is just fix that, and then that will connect it. And then, what I'm able to do, and I'm not going to have a chat with it right now, but I'll say, hey Claude, can you use the Kowalah connector? Don't show me any data, but can you just tell me what tools you've got access to? I was trying to add the other one so that it wouldn't show any information about client information, but hopefully Claude is going to behave and just show me the tools that you've got access to. Let me load the Kowalah tools to check what's available. Here's what I've got access via the Kowalah connector. So, 20 tools in total, broken down into two categories. So, the read tools, getting project, organization, user, team member details, and then listing tools.

Charlie Cowan: and then one write tool to update expert requests. So it's mostly read-only, the only write capability is updating expert requests. Everything else is query and lookup. So you can imagine how I would carry on this conversation, as I do when I'm using the Kowalah application via Claude, is just saying, right, list the open projects that we've got, can you update that? Can you create a new expert request for that client? And it's just such a great interface. Why is this a great interface, rather than going directly into the Kowalah platform? Well, my Claude has also got access to Gmail, to my Granola Notes, to my Google Drive. And so the Kowalah platform is super powerful, but it's just one spoke in my world, and Claude is the hub, and this is why I find it so useful to bring this information in here.

Charlie Cowan: So hopefully that gives you a little bit of a feel for how you'd use an MCP server, and why it's so valuable to get this data from your other systems directly into Claude or ChatGPT. I'm keeping an eye on the questions. If anyone has a question, feel free to pop them in there, no problem at all. So we've talked about read tools, great, that's just getting information, and then write tools. My general guidance is start focusing on read information, getting stuff into Claude, or into ChatGPT. This gives you the confidence that the data that's coming in is correct. But very quickly, you're going to gravitate towards wanting to write back into these systems as well. And this is where, bit by bit, you're gonna open up the ability for Claude to write. And I'll give you one example.

Charlie Cowan: We use a pre-built HubSpot connector, and I use the HubSpot connector to update and manage our pipeline in HubSpot, but from Claude. Claude is the one that is writing, creating new opportunities, updating next steps, and I can tell you my CRM has never been as tidy as it is now in the 25 years that I've worked in sales, because Claude's the one that is guiding me through that and putting the information in there. Great question here from Andy. You've used 75% of your weekly limit. Do you often hit that limit, and what do you do then? Do you have any tips for when we hit the usage limit? This is a great question, and I have to say, if someone was to ask me what keeps me up at night, it's this, at the moment, causing me a lot of stress.

Charlie Cowan: It's the reverse of your phone running out of battery and watching it go down. So let me just go to my settings, and usage, and I'll show you what I'm looking at. The way that it works in Claude is that you've got two different usage limits. So one is a 5-hour limit, and every 5 hours this renews. And when you hit that, it says, you know, you can't play anymore. It's a bit like the barman telling you you've had too many drinks and you've got to go home. So it's like, right, no more chat for you, you've got to wait until 6pm. You can see here, 1 hour 38 until that limit. And then, you've got the weekly limit, which, as it sounds, goes for a week, and this goes until Friday at 12 o'clock, and this is one that I'm nearly at.

Charlie Cowan: Now, I fairly regularly hit my 5-hour limit, because I'm hammering Claude Code in multiple instances at each time, and I very rarely hit the weekly limits. The reason why my heart rate is quite high, and that I'm thinking about how long it is until Friday, 12 o'clock, is because we had the Easter bank holiday. And I hammered Claude building loads of cool new stuff that I can't wait to tell you about on future webinars. So, I hit Claude for 4 days very, very hard, and that's why I'm nearly at my limit. But I think I can get to Friday at 12 o'clock, so I'm just sort of managing my usage a little bit.

Charlie Cowan: When we get to the AI in the news, I'm going to provide a little bit more context as to why this is so important, and why you're hitting limits in Claude, but maybe you're not hitting limits in ChatGPT or other tools, because it's very specific to Anthropic at the moment. So, sort of just wrapping up a little bit about how to build your MCP server if you are non-technical. Hopefully, what I've shown you so far is not scaring you technically, but actually, let's think more about this from the business perspective. So, firstly, I would say start planning. You don't need to be an IT specialist or an AI specialist for this — have a whiteboard, and have a think about this. Like, who's the user? Who's going to use this thing? Is this someone in our company, and we want our employees to be able to use this thing? Is this our customers?

Charlie Cowan: And we want our customers to be able to access this thing? Is it prospects? Is it suppliers? So, who's going to be in their own ChatGPT, or their own Claude, and using this thing? And then secondly, what are they going to be asking? Like, if one of our customers comes to their own Claude and starts asking about our product, our service, what is it they're going to be asking? And what do they normally pick up the phone, or what do they normally email us about? What systems do they use? So, if it was a human doing this, what systems would they be going to, to get that information? Is it from email? Is it our products? Is it our accounting system, our ERP? Do they want to just get information? Have they got questions, and they just want the answer? Or actually, do they want to update it?

Charlie Cowan: So in the Kowalah scenario, not only do our customers want to get information about their project, but they want to update it. They want to create new tasks, they want to create new requests. And then also, where do they want to ask? So, I've talked about using these in Claude, or in ChatGPT, because that's easy to demo, but maybe they want to go to your website, and they want to ask these questions. Maybe they want to ask it in WhatsApp. Maybe they want to ask in Teams or Slack. So just having a think about who's the person, what are they trying to ask, and where are they? And you can do that on a whiteboard or with a flip chart. Next, you start to move into a bit more of the technical side of things.

Charlie Cowan: Okay, right, we now know what systems and what data and whether we want people to be able to read or to write. And so now we should start thinking about, well, where is that data today? And so let's use the example of a CRM system, and you want your employees to be able to access your proprietary CRM system that you've built. Well, where is that data? You know, is it clean? If it's junk, and it's all of a mess in this other system, then Claude is going to get junk and a mess, and it's going to be answering off bad data, and so people are going to be confused and say, well, this is rubbish, but it's because the data wasn't very clean to start with. Is it accessible?

Charlie Cowan: Like, can you get to it via an API? Your IT team will help you with this, but, like, how do we give our MCP server access to this data? I would suggest starting very, very small. Start with basic tools that you're going to create. I want to read this information and that information, but don't come up with some long list of tools across lots of different systems and reading and writing. Start small, get confident, and then start to build from there. And once you've got that, then you're thinking, right, well, now I need to build this thing, and either I'm not an IT expert, or even our IT team are not MCP specialists, so how are we going to build this thing? Well, don't worry, Claude has got your back, and I'm going to show you on another screen in a minute, but let Claude do the work.

Charlie Cowan: You don't need to learn this. It's useful for you to know, because you can structure on the left side the planning and the data, but the actual building, Claude's gonna nail it. So, your IT team are going to use Claude Code, is how we would suggest doing it, because Claude is going to then have access to the source system. But you're going to use a thing called the MCP Builder Skill, and I'm going to just show you that in Claude. So let me go Customize, Skills. And I want to go to, here we go, MCP Builder. So, Anthropic have built a skill to teach Claude Code how to build MCP servers. So you don't need to learn, you can cut out the middleman. And so, just so that you know this exists, it's called the MCP Builder Skill. It is pre-enabled in all of your, your users, you know, Claude instances.

Charlie Cowan: And it basically describes to Claude about building an MCP — understanding the design. Here's the protocol documentation, so that Claude knows all of that. Here's how it should work, the different stack. You don't need to even know what this understands, because Claude understands it, and you just need to be able to talk. So yes, you can have a chat with it in your Claude, like this, but I would advise that your IT person that's going to build this opens up a Claude Code instance and talks about the MCP Builder skill in that, because then Claude Code is going to have access to your source systems and it'll be able to do a much better job for you. And of course, the Kowalah team are here to help you with any of that if you get confused.

Charlie Cowan: Now, I just talked a minute ago, as we start to wrap up the MCP topic, about thinking about use cases, and that it's really easy to think about use cases for your own employees, because you can give access to your systems. But where this becomes really valuable is to then start thinking about people outside of your organization — your customers, your prospects, candidates that might be potential employees, it could be investors if you're a public company, it could be suppliers if you've got a big supply chain. Start thinking, these are all people too, and they're in their systems, having a chat with ChatGPT or with Claude — how do we help them to get access to the right data, after they've been authenticated? And one example that I really want to show you, which I love, is a connector from a podcast that I listen to called Lenny's Podcast.

Charlie Cowan: Now, let me just go over here, and I'm going to go to Lennysdata.com. Now, Lenny's podcast, if you're into anything to do with AI, or product, or growth, Lenny's podcast is amazing. Episodes every week, hour and a half, interviewing some of the leading people in product, and especially AI — hundreds, I think maybe even 300 episodes, and it's what I listen to when I go on a run at the weekend, absolutely love it. What Lenny has done is he has taken the transcripts of every single podcast, word for word, and put it into a data repository, including, in fact, it says here, 349 newsletter posts and 293 podcast transcripts. Amazing. And so you can just get to the data as is, but he's created this MCP. Fantastic. So, all I had to do, having logged in here.

Charlie Cowan: was then go and connect up Lenny's podcast in exactly the same way — he's given me mcp.lennysdata.com, and then I authenticated with my subscription. And what tools has Lenny given? List content, read content, read excerpt, search content. So, let me go and have a chat with Lenny's podcast here. There was an episode last week, I think, with the head of growth at Anthropic, so let's just see. Hey, can you use the Lenny's podcast connector to take a look at the most recent Lenny's podcast? I think there was one with Amal, who's the head of growth at Anthropic, so can you just take a quick look at that, and then if you got access to that, what were some of the key themes that Amal covered? This is a risky one, because it was a week ago, so let's see how often Lenny updates it. Otherwise, I'll have to go back a few.

Charlie Cowan: So, Claude wants to use search content, always allow. Found it, published 3 days ago, let me pull out the key themes from the episode. Read the excerpt. Reading the excerpt of the podcast? It was an hour and a half. Come on, Claude. Well, I'll let that build it out. What I found… oh, here we go. Got a solid picture of the episode. Here's a summary, Charlie. Published fifth. Anthropic's $1 billion to $19 billion growth run: how Claude became the fastest growing AI product in history. Okay, and I listened to this on my run at the weekend. I can tell you this is exactly what the podcast talked about. So I talked earlier about, but what's the value of having an MCP server? What's the value of bringing all this content into Claude?

Charlie Cowan: You know, why don't we just get people to go off to the actual source system, whether it's our CRM system? Well, it's because we can do stuff like this. Hey, this is amazing. Based on the Kowalah strategy document that you've got in Google Drive, how do you think this could be relevant, or what lessons can we learn from it? So because Claude's got access to other documents, it knows more about me and what we're working on and different things, and so I can take that whole one-and-a-half-hour podcast, which Lenny's MCP knows about, and now Claude is going to have a look at Google Drive, where I've got a strategy document, and I'm able to have a really interactive conversation with Claude about bringing all of these things together. So this is super useful.

Charlie Cowan: I'm gonna let that carry on, but let's wrap up what we're talking about with MCPs. So, planning, data, use the MCP Builder Skill with your IT team, or of course, the Kowalah team, very happy to help you out as well, to get the thing built. This is not a month-long project — literally the one that I was hoping I was going to show you earlier, I just started earlier this week, so this is a quick turnaround once you've got the plan and the data ready. So, we've talked about MCPs there. I hope that if you're non-technical, that has not scared you, made you run for the hills, but I hope it has given you some ideas that you can start to think about for usage in your own organization.

Charlie Cowan: Now let's switch gears and head into AI in the News, our recurring segment each week, where we look at some of the top releases that are going on. Just as a reminder, I have a recurring search that happens on a Wednesday morning, and it looks for any updates, model releases, feature releases, from OpenAI, from Google, for Gemini, from Microsoft, for Copilot, from Grok, and also from Anthropic. And I like to think that I'm very fair at looking at all of those releases. I think about what is useful for our business-focused audience, and that's what I cover. It just so happens that a lot of the releases are coming from the two leading labs, OpenAI and Anthropic, and so that's what we're going to talk a little bit about today.

Charlie Cowan: The first one — this is a little bit of an Easter egg, things have just come out of Easter, but maybe it's a little bit of a gimmick — but the reason why I wanted to show it is because I think it's very relevant to the direction of travel that OpenAI and Anthropic are going on. So, last week, 4 days ago, ChatGPT is now available in your car via Apple CarPlay. So, if you've got the right iPhone model, you can now have a chat with ChatGPT while you're driving. So, on the one hand, that's good, quite fun. But on the other, what's the business implication of that? Well, maybe there's a little bit of some exec having a coaching conversation with ChatGPT.

Charlie Cowan: But I think the reality is, this is really going for the consumer play, and this is what we're seeing a lot of with ChatGPT. It's not to say that they're not focused on the enterprise, because they very much are, with their coding model Codex, and a few things like that. But they are also very much playing the consumer competition to compete with Google and Gemini. Google is, at its heart, an advertising business, and it is aiming at consumers to get eyeballs on search or on their new AI mode in search. And that's where Gemini is focused — to get on as many devices as possible. And that's the game that OpenAI have got to play. So things like this are just a good little indication of where their focus is.

Charlie Cowan: Like I said, it's not that they're not focused on enterprise, but if one-third, or a half, or two-thirds of your focus is on driving a billion weekly active users into consumers, then it just means that you're not 100% focused on the enterprise. So, what does that mean on the Anthropic side? A few things to talk about here. This was the 6th, and it is the 8th today, so this is 2 days ago — Anthropic announced a massive new partnership with Google and Broadcom to build out tons and tons of compute over the next couple of years. I think this comes online in 2027, so it's really a commitment to have loads and loads of compute that is going to support their growing needs in the coming years. And to your question, Andy, about usage — this is the problem.

Charlie Cowan: Anthropic is limited by the amount of compute and the amount of usage that they are seeing from their users, and they're having to put some plans in place in order to be able to service that need. So what does that mean in numbers? Well, this is their revenue numbers, and it's just crazy. The start of this chart is January 2025, so we're talking about just over one year ago — this is not a five-year chart, this is a year ago. So, one year ago, January last year, they were at $1 billion of revenue. At Christmas of this year, so this is just 4 months ago, they were at $9 billion of revenue.

Charlie Cowan: At the end of February, so when Amal was doing his podcast with Lenny the other day, he talked about being at $19 billion of revenue at the end of February. We are now at the start of April, so one month later, and in that Google press release, they said we are at $30 billion of revenue. So numbers are crazy — in one month, they've gone from $19 billion to $30 billion of revenue. They've added $11 billion of annualized revenue in one month. And this line is going up, up, up.

Charlie Cowan: So, when we talk about why Anthropic is limiting the usage on each account, it's because people are throwing money at them, and they are throwing prompts at them to use that money, and the subscriptions that you get are so well subsidized, that they're having to limit the usage that we're seeing. So, who knows where this goes? I can only imagine that this chart goes more vertical, not less, because when I'm speaking to our clients, the number of people that have heard of Claude is still minimal to near zero. The actual adoption in terms of users compared to a ChatGPT or Gemini is still minimal in terms of number of people. But the money they are transacting from the users they've got, because they're all enterprise, is absolutely huge. So, who knows where this goes in terms of the size of the business.

Charlie Cowan: But the predictions are this will overtake companies like NVIDIA that are right at the top very, very soon. Another announcement that came out from Anthropic — I think this was yesterday, one day after the announcement of the $30 billion revenue and the stuff going on with Google and Broadcom — is a thing called Project Glasswing. And there are two threads to this. Project Glasswing is a multi-company initiative that brings together Amazon, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux, Microsoft, Nvidia, Palo Alto Networks. So, big companies are involved in this. And what is the initiative?

Charlie Cowan: The initiative is that because the AI models are getting so good, they are able to find vulnerabilities in existing software that humans would never find. So what does that mean? Well, let's take Microsoft — Microsoft Office, Microsoft Outlook, Microsoft SharePoint. Here is software that exists across all of these different companies, and humans have been writing penetration testing and cybersecurity tests to try and find vulnerabilities, and as far as a human is concerned, it's as safe as safe could be. What we're finding is that these new models are able to go and find zero-day exploits — by zero-day, we mean something that a human was not aware of, the first time it's been heard of. And Anthropic is finding these at a rapid rate using their models.

Charlie Cowan: And so, of course, that could be used negatively, but what Project Glasswing is about is looking at it from a positive side — let's use these models to find the zero-day exploits, and as a collective, solve these things before bad actors find out about them. So that's what Project Glasswing is, driving forward the cybersecurity agenda in this new world. Now, the second part of Project Glasswing that was announced is a new model that Anthropic is launching, called Mythos, that is only in preview and being available as part of Project Glasswing and the partners that are in this.

Charlie Cowan: Benchmarks, benchmarks, benchmarks — and all you ever really need to know with a benchmark is, is the new number bigger than the old number? And yes, the new number is bigger than the old number. So, up until now, Opus 4.6 is the largest, most powerful Anthropic model, right there at the top of the benchmarks with the Pro models from ChatGPT 5.4 and Gemini 3 Pro. Opus 4.6 is right up at the top there. Mythos Preview is demolishing all of those. This is agentic coding — new number bigger than the old number. Reasoning — new number bigger than the old number. Humanity's Last Exam — new number bigger than the old number. Agentic search and using computers — new number bigger than the old number.

Charlie Cowan: And so that's all you really need to know — this new model is a significant advance on the current top models. What you should also know is it's very expensive and very big, and therefore it's only being used in the preview world for Project Glasswing so far. They've talked in their blog post about how they might license this — maybe just for large enterprises, for certain use cases. But why is this useful for you to know? The models you're using today are the worst they will ever be. Models like Mythos are just around the corner, and they're going to take what is already amazing up to a completely new level. So, be aware, get ready, and buckle up, because there are some fantastic new models with huge power coming down the road.

Charlie Cowan: So I think that was all I had for AI in the news today — Project Glasswing, new model better than the old model. If you've got any other questions, feel free to ask them. Oh, there is one here — Paul. Hi Paul, hope you're well, good to see you. Do you think the leak of Claude Code source code will close the gap between more consumer-aligned vendors and Anthropic? That is a really good question — for anyone that doesn't know, there was a Claude Code source leak, reported by The Guardian among others.

Charlie Cowan: This happened maybe about a week ago, just before the Easter holidays. This was human error, not a problem with Claude Code itself — but in the process of deploying Claude Code out to the bundle that people could download, someone inadvertently left all of the source code in there. Someone was able to extract it, and they could go, oh, we've got Claude Code, we've got Claude Code.

Charlie Cowan: I actually think that whilst this was a bit embarrassing for Anthropic, it's not too much of a problem, because what you're looking at is where they were a couple of weeks ago, but they're shipping so fast, and they've got such a product roadmap, that if you spend too much time looking backwards at what they were doing, yes, you might understand a little bit about how they structure it, but I don't think — if I was Anthropic, other than being a little bit embarrassed — I'd be too worried about it, because the longer someone spends looking at where Claude was two or three weeks ago, the faster Anthropic's gonna be accelerating away. So that's my general feel for it — the Claude Code team ships at such a crazy rate.

Charlie Cowan: I mean, they're literally shipping hundreds and hundreds of PRs a day, so I don't think they'll be too worried about it in the long run. And Andy has a comment on that — I guess that also creates a significant competitive advantage, commercial advantage, to those big businesses listed, that they can have access to new AI tools before the main marketplace does. I think this goes back to the Glasswing thing — companies like the Googles, the Microsofts, they are getting on board with the fact that Anthropic is pioneering here, and they need to be on board with figuring out how to work closely with Anthropic, because Anthropic's got the enterprise pretty much nailed, and they need to be on board with providing them the infrastructure and getting their models embedded. Thanks guys, great questions and comments there.

Charlie Cowan: So let me wrap up with what we're going to talk about next week — the Kowalah Concierge plugin. This is a little plugin that we built to help executives at our clients, or at our prospects, get started with Claude Co-work really, really quickly. You don't need to wait till next week — you can access the plugin now, just go to the Kowalah website, go to Resources, and Concierge Plugin, and download it, and load it up into your co-work. But what I'm going to spend a bit of time on next week is just walking you through how to install that plugin, and what each of the 6 skills in it do. And there's kind of a meta topic here, which is…

Charlie Cowan: When I give you the plugin to download off our website, and you install it into your Claude, I'm basically taking what's in my brain and transferring it into your brain via your Claude, so that you can know everything that I know about setting up co-work. And what I hope you'll get as we talk about that next week is, well, what's in your brain, and how can you get that into either your employees' brains, or your customers' brains, or your competitors' brains, or your suppliers' brains — using this concept of a plugin as a kind of information exchange. So, hopefully we'll come away from that with two or three useful learning points. With that, thank you very much, for another fun Kowalah Wednesday webinar.

Charlie Cowan: I hope you come away from this thinking about one or two MCP server ideas that you might want to build, and as always, just ping me, either directly or via LinkedIn or X, if you've got any questions. Always happy to unblock anything that you're stuck on. Have a great week, and thank you very much!