Goddag allesammen og velkommen til AlwaysAsia. I dag kommer det til at foregå lidt anderledes, end vi plejer, because we'll do it in English today. I invited my good friend Massimo from Microsoft, and we'll be talking about... Model Context Protocol. MCP Server. The MCP Server. The USB-C of AI, right? Yeah, yeah, yeah. I can't wait. I can't wait. So today we're going to go in and talk a bit about MCP, what it is, some about the components. We'll do some demo of some Microsoft MCP servers. Maybe also talk about that you can create your own MCP server, right? Right. Okay, okay, cool. Should we dive in? Yeah, we can start. Are you ready? Yes, yes. Okay, cool. The USB-C of AI. Not USB-A? No, yeah, just USB-C. How happy are you that, you know, nowadays you use USB-C for charging your devices? I can't live without it. I'm missing my bag full of dongles for every interface I used to have, right? Now I only have one because I still have some USB devices, but USB-C for everything. Now MCP for everything when it comes to having an interface for your LLM, right? Exactly. And this is the same happiness that you have of having this only USB-C for your devices is the same for AI application. It is mind-blowing. And also the Modic Context Protocol. We will talk about how it, what it is, how it works and so on, but it's also a little bit how people is changing the way they interact with software. It's not anymore SaaS users, it's AI users. How do I consume this intelligence all around? How do I keep the context of the AI conversation that I'm having, right? This is basically what we will be discussing, right? Yeah, okay. So let me start off with a very simplified architecture on today's AI application. So when you have an AI application, you have a prompt, you have some memories, someone, the user is asking something through the prompt, you process it, you keep this memory of conversation. And then usually the AI applications is doing something in different applications that you need to integrate with. Let's say Outlook, Outlook calendar, create an invite for me. So what you would normally do in your code base, you maybe have a plugin, this plugin needs to be configured, you need to keep the API keys and so on. And then you do this type of integration. Then you have another plugin, another plugin, another plugin. Yes, yes. And this has become a little bit of a problem when you scale AI applications. What it really happens is that we are basically in the middle of so many integrations. It's a mess. And we are really trying to figure out which cables belongs to what and managing basically all type of integration that we might have. Let me just switch for a second to this one actually, because this explains it a little bit even better. It's a nice slide. I took it from an article online. It's extremely interesting. So you need to think that each of these applications that you need to integrate with is some sort of a lock, right? Then you need to have the right to access this lock, what you will have normally for an API, an API key. That explains it all, right? So each integration, one API key, each API key that opens one particular lock. But even more, right? Some might use keys, some might use some other key cards, some might use some other technology, right? Exactly. Where you would have an API key, you would have a token, you would have OAuth, you would have different technologies that access well, right? It's even worse. Yes, it's not only a key. It's even worse than we see now, right? So this authentication flow is something that you need to manage. Yeah. What about changing an endpoint? Same problem. You need to manage also that. Or you have rate limits for each API. You need also to be careful on this part here. And of course, the development cost of all of these increases all the time, right? Because this multiplication of AI application with the type of integration that you are doing. So it's a big cost. And of course, another thing is that you don't have a centralized logging system, monitoring system. And you as AI application owner, you need to manage the context of the conversation of your LLM, right? So it's not really straightforward. So there is a need definitely, according to also what we are discussing, to have a standard, right? And what happens when you want to have a new standard? You get even more standards, right? Exactly. A standard on top of all the other standards, right? Exactly. So the typical situation, I find it very, very funny because, you know, these companies that say, okay, we need a new standard. And it is for going forward, we're going to use this, right? That is going to fix everything. You turn out to just add an extra standard on top of the existing ones. Exactly. But I believe something different happened this time. Antropic came with this MCP. And I believe all the tech companies started loving it. and it was basically adopted by all the big companies, Google, Microsoft as well. And we are really building on the top of it. And this makes life easier for us as well as developers, right? As someone who wants to monetize even MCP and say, okay, how can I do it? How do I do it? And Microsoft delivers all kinds of tools already, MCP servers, service service. We will do that as well, right? Yes. So MCP is more or less the official standard now for doing these types of integrations with L&Ms. It has been adopted by Big Tech, Microsoft as well, right? Exactly. So I don't know if there's an official, official type, but more or less it's becoming the official way. It's the official way, more or less. Yes. And I will show you also what Microsoft is delivering, the list of the MCP servers, what can be relevant for you. And I believe there is always this twofold, right? As a user, I want to consume these MCP servers. I want to use this intelligence outside. And as a professional, I want to be able to build my own MCP servers based on the software that I'm producing, right? Yeah. So I want to, as a user, I want to be able to get my information with the prompts, get that information from you because I utilize your tool. Yeah. And you as a developer, as I going from a SaaS company to more agent-based type thing, you want to deliver your information and your business logic to the user in more ways than just your browser and your interface. Yeah. And you're a new, completely new interface of interaction and easier way also of integrating, right? I don't want to keep all these API keys. I don't want to understand each of these integration, what they are really doing in depth, because there is some sort of a real cut between what I'm doing for my business as a client and the server people, right? The people that are providing me the service, they are also having their own life, right? Yeah, yeah, yeah. And we need to always be in sync. And if we are not in sync, something is not going to work really in the right way. At the same time, also contextualizing the conversation that we are having with the DLLM is very important. Yeah. So MCP was introduced and there are these, I will say, three main building blocks. So the first one is, of course, we have simplified architecture and integration. The USPC type. USPC type. The universal cable. You don't have to have tools that access or these 47 different integration. There is a reduced development time. And I will show you why there is a reduced development time. faster time to value, simplified scaling, because you don't actually do tailored integration for each application that you have, right? And there is also another interesting point from an enterprise perspective. Me as an enterprise, I want to be also sure that these MCP servers are secure, safe, and I can easily connect with them and I can trust those servers, right? And Microsoft is also delivering this capability of having your own, I would say, MCP registries code. You can find it also online if you want to know a little bit more about it. But the concept is, I am whitelisting in my company the MCP servers that where I can, you know, which I can consume safely, right? And I believe this is very powerful. And also, of course, from security and governance, unified authentication layers. Yeah, finally. And error handling and logging that is also unified, which is super cool. And then from model performance and operational efficiency, I just want to mention, I believe, the most interesting one is the coherent multi -step reasoning and conversation continuity. I believe that that is the point of having MCP server or servers. Yeah. So, as we have this conversation, you remember what we talked about yesterday, right? It's the same here. It'll remember, it'll be coherent, it'll remember, and it'll do reasoning across the different conversations we were having. We would know that as people, right? And the MCP here would understand that as well, right? It's the same concept. Okay. Yeah. And also, MCP is just, the power when you expand to multiple MCP servers, then your conversation becomes even more relevant for what you're working on and specialized on what you really need, right? So, that's again, a very lovely feature. I believe he says it all. Very trendy, right? Yeah. It's the same interface. It's the same way we're in terms. Yes. It's just one single entry point for all your MCP servers. In this example, of course, I put Outlook, Outlook Calendar, OneDrive. Maybe you have a simple integration that you want to synchronize your files and have some meetings. And maybe in the meetings, you want to attach the latest presentation that you had in the previous meeting. And everything, this is done with one simple connection with the MCP server, right? And then, normally in the MCP architecture, you have an MCP client. Yeah. So, this client, you can use Cloud, you can use ChatGPT, you can use all other type of clients that you normally find, but you can also build your own client. And I did that, and I want to show it to you with semantic kernel. I believe it's super powerful. And I will also show you how the LLM actually is orchestrating the requests coming from the user. And this is good for running your LLM locally on your own machine, instead of using one of the big LLMs out there, right? Exactly. There is a lot of open source stuff. There's a lot of different stuff on the market that you could just bring down, run on your local machine. That's a very nice point. Right? Because actually the MCP server, there are two types. The one that is locally running, that could be the one that you have on your machine. And it uses the library standard IO. Sounds like very C++ all the time. Exactly. But I like the LLM. At least there is something familiar. Exactly. And then you have the remote one. What they have in common is some sort of a configuration that you need to do. And you need to specify, is that a local or is that a remote? And then of course you need to, if it's a remote, do the authentication authorization process. But let's go one step down and see a little bit what's the life cycle. Oh yeah. Let's first talk about the musketeers of MCP. So we have the MCP server. Yep. What can we do with that? So we can do basically three things. Yeah. Resources. So we can ask the MCP server, give me information about the database schema. Yeah. Or you know this type of, even the old word of the API, right? Yeah, exactly. Right. Those are read only data access type of requests. So I can access any resource. Yes. And ask for information about them. Right. Very low level layer. Then you have the tools. The tools is the real, I would say, the power of the MCP is to basically be able to do actions. Do something for me, right? Exactly. Give me the database schema. Oh, there's something wrong. Change it so it XYZ does something else, right? Change the schema, right? Change the char links of whatever. And then it does it for me. For instance, in this example, create a GitHub issue for me. Yeah. What should I do? I should integrate with the API, do the API, get the API key, find out what's the function name called, and then do the code. be careful on the parameter I'm sending. It's incomplete payload, incomplete payload. What is required, not required. And here is just natural language asking for creating me a GitHub issue. And it does that. So that's pretty powerful. The prompt is a collection of reusable templates. So it could be that your MCP server gives you the list of capabilities through a list of templates and say those are the templates that I want to use. So let's see how the interaction between client and server works now. Yeah. So we have basically these four phases. The initialization phase, which is basically me as a client. Yeah. I'm there. Help me out. Yeah. Are you there? Yeah. That's your endpoint. That's your transportation layer. Are you there? Oh, yes, I am. Cool, cool, cool. After that, there is the discovery phase. The discovery phase is the aspect that I love the most. Because the MCP server is going to answer you with the list of capabilities that the server has. And that also means that you have a list of these capabilities and these capabilities are recorded into your client right away. You don't have to specify anything. That is pretty awesome, right? So how can I interact with this MCP server? By just this discovery phase, I'll ask it how I can interact with it and it'll tell me. And okay, now I can. It's a simple. So that's, it is natural language. If I was in a hotel going to the reception, they would ask, how can I help you? And I say, I have no idea. How can you help me? And they will tell me, I can get you a room. I can do it, I can do it. It's like you get the paper with the list of services that they can possibly give you. And then you say, oh, wow. Are you really good? Yeah, yeah. In the normal world, you should have registered each one of them. If I understand. You know what I'm really doing. So this is, to be honest, it's one of the biggest and most powerful aspects of the MCP, which I really, really love, right? An MCP server can be updated on the fly. In the next call, when I ask which are the services that you can do. It'll be a new one. It'll be a new one. There will be more. Yeah, but I don't have to configure anything. No, no, no, no. Yeah, so that's powerful. Then the invocation part is basically selecting the right service that I need to call. And then you have the execution. And the execution, you get something back. This example is the just pull request from GitHub. So, I think Microsoft is working in all of this world of the MCP servers. We are delivering a number of MCP servers. I have to be honest, I didn't try all of them. I tried some of them. Yes, I've tried some of them as well. And it's a total eye-opener of interacting with the services you use day to day, right? Yeah, completely. Completely. So, I've played around with the Azure MCP server and the Azure DevOps MCP server. And just getting that information so quick. So, doing that discovery, how can I help you? And then doing the next parts of that phase, figuring out, getting that information so fast and so timely. Instead of me writing my own logic to get that and doing my own integration, right? To get that logic. Exactly. I believe one of the powerful things also with the Azure MCP, we tried it together as well in our demo. I believe there is a big difference on going on an open AI-created model and ask for a BSEP template. You need to ask for it and say, I have this in mind, please give me the template out of the box. You get the template, right? Yeah. It's not contextualized on your own subscription, on your own resources. And this is where it gets… It's a game changer. Very awesome, right? It's the context of the co-pilot and the MCP server working together, giving me that context. So, I ask for that BSEP template. It'll know what I'm working with. It'll know the context of my project. Know, oh, you probably would need these type of resources, right? Yes, exactly. There is the AI Foundry MCP server, Clarity, Dataverse, Files, Markitdown, amazingly. DevBox. I want to try this, the Fabric Real-Time Intelligence MCP server. I believe this is going to be something unbelievable. Microsoft Learn, as you can see also from my shirt. Yes, exactly. And again, it shows just the potential of where you can get that information, how you can get it, right? Yes, exactly. So, getting that information where you are, right? Yeah. Actually, you could, in principle, also have it running locally, these MCP servers. So, you don't have necessarily to go there and connect remotely. And then you can basically work with that SQL. I believe SQL is also, there was a need for that. Playwright. We will do a demo on Playwright. We'll do a demo, yes. Very exciting demo. I can't wait for it because it's very good information in that demo, right? Yeah, it's very, very nice. Very good, very good. Well, it's very important, those informations. Yeah, yeah, exactly. And what else do we have? We have the Kubernetes service server as well. Kubernetes, I will say, is not completely straightforward, right? To do operation with it and the management of it and so on. I believe the MCP server just really makes your life easier on managing this type of resources in Azure. and then the 365 agent toolkit as well. There is a link here, I believe. We'll ship that out. We'll make sure. Exactly. Or if you just go in, ask your LLM, your favorite search engine, whatever, Microsoft MCP. Oh, yes. It's a GitHub. Yes. You can just find it on GitHub. Exactly. You can just do that work. Because it's fun. Because it's fun. Yes, exactly. Let's go through the first demo, which is the Playwright. So, it's an MCP server for browsing the internet. And you build these LLM prompts. You interact with web pages. You ask the LLM to maybe do other actions on these pages that you have been searching for. And it's pretty powerful. If you think about a word of a development team. And in this development team, there is always the testing phase, right? So, if I could use a Playwright MCP server and say, please test the pages that I've been developing. I already know what my use case would look like, right? In plain text. I already know which steps I need to go through. Throw it in as a prompt with some context. Show me the result. What is the result? Exactly. Oh, I couldn't because I get this 404. Ah, okay. Something like that, right? Yeah. You can even ask, do these test use cases, as you said. And then say, build a test report according to this template that you built on the site. And then you get the... Maybe even give me suggestions of how I could fix the problems I'm having, right? That's also another good part, yes. I mean, you can really go... So, you can just go further and go further and go further and experiment, right? Let's do a little demo. I built this demo using the Playwright. Let me go... Yes. Okay. I just want to show you a couple of things. Yes. Let's go in. So, the first phase here is the initialization phase. This is where basically I'm saying to the server, I would like to communicate with you. Are you there? Yeah. And of course, here is a standard IO, as mentioned, the transport layer. But they could have had the... The HTTP... Yeah. SSE actually, client transport. Let me see if I can actually hover it. Yeah. As you can see, it provides over HTTP and it used the service and events or streamable HTTP protocol. So, you can have these two options. Of course, for me, I just use the local one to make things a little bit easier. And that's basically what you do. You just configure it and you're good to go. The discoverability phase, this is where the beauty comes, right? List tool async. That's it. That's all I need to do is asking, give me all your capabilities, please. Once you get all your capabilities, when I run it, I will show you which capability it has. And you're just outputting the capabilities. Yes. Here, I'm outputting the capabilities here in the console here. And after that, I am using the semantic kernel plugin registration phase, I will say, where I register a capability for each plugin that I have. I just called the add from functions. It's everything that basically comes out from the server. Yeah. So, it's like registering the entire server. Okay. And you give it as an intelligent capability to your client. Okay. This is completely mind-blowing in my view. So, yes. After that, I configure my OpenAI execution. I asked something. Do you want to see what I asked? Yeah, yeah. Let's go for it. What do you ask? I don't know. Let's find out. No, let's find out live. Drum roll. Let's see. Here we go. It's doing something. In the output here, you should see in a second. Yeah. Those are the list of capabilities. Capabilities. Yeah. If the server adds new capability, it's automatically there. You'll just get it. It'll be there. Hey! Oh, yeah. That's me. Oh, yeah. I know this guy. Yeah, you're this guy. So, I had actually a prompt. Let's see. I had a prompt which was asking about you. Okay, yeah. And I actually wanted to find a little bit more about you and what you're doing and your interest. and so on. So, as you can see, and you need to confirm if it's you or not now. Okay? This is also a good head start for a first date. Right? Yeah. To go out and get some information. Who am I actually? Who am I seeing here, right? Or someone that you meet for the first time. You know, you say, are you really you? Yeah. Can I trust you? Let me see. What does my... Maybe also use it for something more serious. More important. Okay, okay. Hiring someone. But, you know, we know each other but I didn't know for instance... Of course, I knew that you are open-minded and enjoy sharing stories. Do you? Yeah, yeah, yeah. It's true. It's true. But also other information like gardening. Is that true? Yeah, it's true. Yeah, yeah, yeah. Oh, wow. Okay. You have mentions expertise, interests. Of course, you love Microsoft AI and Copilot. Who doesn't? Yeah, which one would love. And also LinkedIn profile and all these things. So, it went out, found some information about me, right? And you just gave it that context, right? Yeah. So, it opened the browser, did the search of your name. He... The server basically has this capability of doing that. And then it took the most important information between sites and did a kind of report for me. That's pretty interesting. And again, as a good use case, the UI testing is a pretty neat use case for this type of thing, right? Yeah, yeah, yeah, exactly. So, I did also another one for fellow mind. I believe I have it as a video. So, we do it very, very fast. Let's see. And I'm asking about the inner cycle recognition. You know about it? Yeah, yeah, yeah. It's the 1%. We're in the top 1% in a circle, right? Even better. Fellow mind. Yeah, even better. I got another information which was also very interesting to know. and the information. Yeah. It's asking for it. Going on the first link. Summarizing the article for me. And... Oh. Yeah. It's a video. You cannot interact with it. Yeah, sorry. Okay, okay. Okay. Yeah, yeah. One interesting information that I got was, I believe it was the 14 years in a row that you won it or something. You got this recognition. Yeah. And I believe it's super interesting to see just the main highlights of that, right? So, again, if you have a new business partner, you're interacting, you're getting some, or you have a salesperson calling, whatever that you used to, go in, get that information. Yeah, yeah, exactly. It's pretty cool. For instance, or as we say, as a tester, I believe it's also... I think the testing part is a pretty neat use case. Now, the one that I believe is also kind of interesting to try was the MCP server. So, I've shown basically how you build the client, how you call the MCP server. Maybe at the end, I can also show how an MCP server looks like, if we have time. But this last one is like, I'm a consumer. I'm working with the WeSell Studio Code. I need to write templates. Yes. I need to build some infrastructure, right? It takes some time to do that, right? Yeah, yeah. Well, we couldn't do a live demo. It will take too long. It will take too long. But I built a super big, I would say, kind of architecture with the AI Foundry, project and hub, AI search, blob storage, private network, you know, anything that makes these templates lean. I added that I wanted to do best practice of it. And of course, then I'm checking other prompts like, am I compliant with 2701? If I'm not, what should I do? And automatically, it generates me the changes that I need to do to the already generated template. It created everything for you, right? It created everything for me. Maybe I want to, instead of going into the video, I want to show you how it looks by showing in the WeSell Studio Code, right? Okay. Let's jump into the... Yes. So, this was the prompt that I just showed you. The beautiful part of it is it created for me the... Everything. It did all the changes. Even the project. It created the project. The project. We have a complete project. So, how many lines of code did you write? Zero, right? One prompt. One prompt. No, no code lines. One prompt. No, just one prompt. That's mind blowing. That is pretty awesome. Yeah. Yeah. And of course, you would still need as a developer, as an infrastructure, a platform engineer, whatever, data engineer, know the context, know what it... You know what you're working with still, right? But there's so much typing here that you don't have to do. No. It's so fast. Yeah. It's really fast. Not fast enough for a live demo, but in your everyday work, this is very, very fast. It's very fast. I mean, think about how many hours you would have spent on understanding the architecture, getting the requirements and so on. There's another thing actually interesting. You could have even asked for, I want to be compliant with my existing security policy, right? Yeah. And it'll know that context. It will know that context. Yeah. Don't be too heavy. It's amazing. It's amazing. So, let's say now, I am in a... I'm operating in a compliant environment. My company is using 2701. I need to be compliant with that. And the beauty with that, I just keep a little bit down here on the problems that I did. I said, okay, I need to be compliant with the 2701. Well, you need to change this. And that's your policy of the 2701. Okay. So, extremely, extremely powerful. Yeah. And then the next thing you can do now, because we're closing into going out of time, right? So, the next thing you could do now is going into... Now you're in development, and then going into operation with your MCP server and your LLM going to ask, am I compliant? And then having these running, reoccurring... You can always ask basically all these type of questions. That is pretty neat. Yeah. So, we're closing in. That was a speed presentation. Very speedy presentation, but very fun, I hope. Awesome. Yeah. I hope they learned something new online as well. I'm pretty sure they did. Let's go. Thanks for now. Take care out there. Thank you very much. Bye. Bye.