Wonderful.

Okay.

So I'm James.

I just hide this thing. Okay. Can you still see it?

Okay. Thumbs up.

Sorry. Sorry.

Yeah. Yeah. Okay.

Wonderful. Thank you. Thank you for your patience, everyone. Thank you.

Introduction

So I'm going to be talking a little bit about mcp uh and i've got a question for everyone um can i get a uh a thumbs up if you've heard of mcp okay there's a few thumbs up okay that's good those of you okay it's quite a lot of thumbs up and can i get one of those like party blasters if you've used mcp in any way okay there's a few in there okay cool cool that's good that's good so people have heard that And there's a few party blasters in there as well. OK, cool.

So MCP has been quite a big thing in AI recently. And lots of people talk about it like it's this very simple thing. And in many ways it is, but in many ways it's not such a simple thing as well.

And so what I wanted to do in this talk was just show a simple way for anyone to start to experience the magic and the benefits of MCP.

About the Speaker

Before I go into that, I'll just talk a little bit about who I am. So I'm James. I'm currently Director of AI Strategy at Awin.

Awin's a big marketing platform. I work on AI feature development. I work on AI enablement and responsible AI adoption company.

Before that, I worked at Adobe working on machine learning products there. And over about the past three years or so, I've spent many, many hours in my evenings and weekends learning to engineer and build around AI models.

And I've experimented a lot, built lots of small applications. And now I build big features in our platform with our engineering teams as well.

Why MCP Matters Now

And why am I talking about MCP? Well, I think MCP is part of this triumvirate, there's a word for you, this big three of things that have happened since the start of this year in AI.

The Big Three: Reasoning Models, Agents, and MCP

Those big three for me is the arrival of reasoning models, so models that can reason and plan their way through sequential tasks. And that's really leveling up what AI models can do.

And those models would be like OpenAI's O1, O3, or Gemini 2.5. or anthropic's new sonic model.

And then lots of people have been talking about this thing of agentic AI, so creating agents, virtual teams of agents powered by AI models to do tasks to get them to do things.

And the third of these big three is the model context protocol. 1And really, this combination of these three is powering a new level of capability and value in what AI can deliver for large organizations and everyday people as well.

Rising Capabilities: Models That Work Longer

And I think this is also these innovations or this trajectory of advancement in AI is also reflected in this research by META, which I've got pictured here, which basically shows that AI models ability to work on tasks. So the length of time that they can work on tasks is doubling every seven months or so.

And we can just see that trajectory picking up since 2024, but particularly since we've got

What MCP Enables

into 2025 with the arrival of these reasoning models and why this is so important in terms of model context protocol is that the model context protocol or MCPs is a way for these more powerful AI models and agentic systems to be able to talk to other systems or platforms like, for example, Google Drive, Gmail, Slack, Notion, and many others. And the MCP enables these more capable models to get information, to use tools, and to take action.

From Ad-Hoc Connectors to a Standard

And the thing is, people have been doing this for a while with these AI models, talking to systems. But before, when they were doing that, for each AI model that they were working with, they kind of had to build their own connectors to talk to these systems. And that meant there was lots of duplication of effort. And it meant you had to build different connectors for different models.

Whereas the MCP has simplified and standardized an approach in a way for these AI models to talk to these different platforms to do stuff. And that's really cool and really helpful. And the whole idea of it is simplification and standardization.

And in many ways it does. But in many ways, as I started to look into MCP, I started to feel like, actually, this really isn't that simple at all.

Demystifying MCP Components

And lots of people who talk about MCP will show you pictures like this. And this is a lovely picture because you have in the middle of the diagram, you have this USB-C connector and that you envisage as being the MCP. And in many ways, that is kind of what an MCP is.

But then you see, and you look at the diagram a bit more, it starts talking about these things like MCP servers, remote servers, local data sources, MCP clients, MCP hosts. And as I started to first look into MCP, I started to think, well, actually, this is all kind of framed in a very technical terminology, which is very much directed towards engineers and developers at the starting point as well. And in this deck, I kind of did a bit of a breakdown of some of the terminology that we have in here because I found it, at least in the early phases, quite a bit of a barrier to kind of get my head around what these things are.

Clients, Hosts, and Servers—Plainly Explained

So you have the MCP client, and that is very simply the app or the interface that you type into. So that could be a chat GPT app on the internet, or it could be Claude, or it could be a coding platform.

Then you have the MCP host, and that is basically the AI model that has the intelligence to talk to enact as an intermediary to go and talk to the systems and do all the tasks and things and again that would be a model like anthropic sonnet or chat gpt uh 4.1

And then you have this thing called the MCP server. And that is basically the platform that you're talking to and a bit of code that they make available that makes it easy for the model to talk to that platform. So there's an MCP server for Gmail. And that enables AI models to talk to Gmail or to Google Drive. And as we'll see in a second, there's one for Zapier as well.

So that's the top level. And don't worry if you forget about these things. I'm not going to really come back to these too much.

Navigating the MCP Server Ecosystem

One of the other things I want to talk about here

So Many Servers: Why It Can Feel Complex

Actually, I'm going to talk about MCP servers. Again, when you initially hear about MCP, the idea is simplification and standardization. But then when you go and you kind of look at all the different servers that are available, and I'll show you here.

These connections on things like Google Drive or Slack or other platforms like that. Third party servers. So these are all the different places. that are offering these MCP connections for AI models to talk to.

So here we go. You start looking at the list. You go, okay, I'm going to pick one. I'm going to pick one.

I'm going to pick one. And you're scrolling and you're scrolling. And that's one set. And then there's community servers.

And you find yourself scrolling. And all of these different servers have their own kind of way of doing things sometimes. And so that doesn't really simplify things too much for everyday people who are trying to understand this capability.

A Practical Starting Point: Zapier + Claude

1So from my perspective, the best place to start for everyone and where you get your biggest bang for your buck is using a platform called Zapier. So Zapier low-code platform. And historically, Zapier was all about connecting things in workflows, and it still does a lot of that.

But then they released something called an MCP connection, which serves as like a master entrance point to many, many, many different connections. So about 8,000 different apps. And that means that all you have to really do is connect your AI provider or AI model like Anthropic Claude into Zapier. And then Zapier will talk to all the other systems and kind of handle that back and forth for you.

And that just makes it really, really simple and easy to get started. And then I would also recommend that anyone who is starting this start with Anthropic Claude because Anthropic kind of created the MCP approach. And their models, in my opinion, are just a little bit better at handling these things.

Automate Real Tasks Quickly

And then what you can do is once you get started with this, you can start to automate tasks.

Could I just get a thumbs up if anyone's heard of Zapier? Okay, there's a few thumbs up. Yeah, a few thumbs up. Okay. Yep, yep. Okay, brilliant. You're all geniuses and you're all AI experts. It's wonderful.

Configuring Zapier as an MCP Server

And I'll just show you what this looks like in Zapier. So here you can see I've created the Zapier MCP server.

You can see I'm using Claude. That's the model that I'm using.

And then here, this is where I configure or I choose what tools I'm connecting into. So you can see here I've connected into Awin.

so the reporting that the company that I work for does. And then you can see I've connected one into Gmail, so I can send things that I get from through to my Gmail. And then I've also created one into Google Sheets as well.

And then adding new tools is just as simple as looking for stuff in here and then configuring it. So if you use Asana on Monday, you can make those connections as well.

Live Demo: From Notion Task to Report

So what I'm going to do is I'm going to show you just an example, a very, very quick example of one of these connections running and a little demo that I quite like. So in my Claude configuration, this is the one I'm going to show you through.

I have connected Zapier to Awin. which is in there. I'd also connected it into Gmail.

Wiring Up Notion, Zapier, and Awin

But then actually what I've also decided to do was I decided to connect Notion directly via Claude. And that's another nice thing that you can do. And the little example I'm going to show you is I created a task in Notion, an affiliate performance report task.

And I say, James, This is simulating work that is something that people ask me. I say, please generate a top line report of a win performance for the first week of May compared to the first week of April.

I need to know top line spend. It's a marketing report. And I set that up in Notion, just like somebody or a team might do if you're working within one of these platforms.

And then what I asked Claude to do is I said to Claude, Please check my notion for any actions that are required for aim and performance reporting and action them. So I will now show you the video of what happened. This is the little demo.

Running the Workflow

So here we go. So here you can see me submitting my query. That's the information that Claude is going to get.

And I say, go and check my Notion for any information relating to Awin or any actions. It says I found the information. And then now what it's doing is it's running some code to go and get information from Awin, so a performance report

And it's pulling that via the A1 API, which is connected via this MCP. And then what CLAW does is it goes away and it writes a bunch of code to then process and run that analysis using that data. And here you can see it's created the performance report and it's come away with that marketing report I requested.

And then what it also does is it then says it has sent that to my Notion. And here you can see that it has created a new page in Notion with that summary analysis report. And it's done a really nice job with that summary analysis report.

Why This Feels Like Real Progress

And I think for me, What I really like about all of that is it's a really simple, practical workflow. And for me, this felt like the real magic of AI right now.

I've been a bit of a purist when it comes to coding things and stuff like that. But actually doing the stuff in Zapier and having these connections, you really start to feel the AGI. And you can really get a sense of where these agentic capabilities are coming together, especially with these ever increasingly more intelligent models.

Oh yeah, I've talked through that, done the demo.

Key Takeaways

So my conclusions, MCP works. Don't be scared of it. It's very easy to do in Anthropic Cloud.

You can also do it in ChatGPT if you want to, and other places. But I'd say Cloud is a great place to begin.

Zapier is great to start with. It's free, and there's a whole bunch of stuff you can do there.

Start testing and learning. See what small stuff you can automate. And you'll probably be surprised at just some of the amazing stuff that these AI models can produce and do for you as well, and actually how reliable they can be.

Final Thoughts

They will trip over from time to time. But I think this is a great way to get your finger on the pulse of what MCP is all about and where Gentic AI is going as well.

And that's it.

Connect and Q&A

Oh, a couple of things. I'm on LinkedIn. Joe's Bentley AI. That's what I look like.

You can find me there. I have some other stuff as well. But like Joe said, feel free to connect on LinkedIn.

And yeah, I think that's that's me. That's terrific, James.

So stay on screen for us, because we've now got Q&A. We've got basically a full 10 minutes for Q&A, if you can hack that, and if there's enough questions. Totally.

Audience Questions

Is there any questions about model context protocols? Or maybe about Zapier more generally?

There's a shy group in the pre-networking.

Tools: Make, n8n, and Others

What do I think of make? What a lovely question. Thank you, Nick.

I think Make is good. Yeah, I've heard really good things about it. I think Make is good.

I haven't used it personally myself, and I see somebody else, Vicky, saying NA10, I think, is the other platform that lots of people are talking about now for automation of these workflows. Some people really love an ATN. I haven't really used it myself. I know lots of people love Make as well.

I think whatever you choose, I would always start off free if you can. And I would say whatever works for you. Whatever is the easiest flow for you to get into.

Zapier was one for me because Awin, we have a connection into Zapier. And what I was really keen to test was how these AI models could pull data from platforms. So Awin, we do lots of marketing performance stuff.

And about a year and a half ago, I had started connecting Awin into ChatGPT to pull some reports to see how it handled it. And when I first started doing that, the AI models just weren't that great at doing it. The amount of information they could hold at any particular time and process reliably wasn't that great.

So when the Zapier connection came in, it seemed like a really, really easy way for me to connect Awin up to some of the other platforms. And I really like it for that reason. And the models have definitely got better at handling stuff. And yeah, there's a lot of great things you can do there.

Where to Deploy MCP First

How do you suggest best identifying the area of the business workflow to deploy MCP?