WEBVTT 00:00.160 --> 00:04.520 And as always, we begin our new topic with the core concepts. 00:05.040 --> 00:14.240 So MCP has three pieces to it, three types of thing which makes up an MCP solution. 00:14.560 --> 00:17.840 So the first thing is called a host. 00:18.120 --> 00:19.840 Now these terms you're going to get used to them. 00:19.840 --> 00:20.720 You're going to hear them a lot. 00:20.760 --> 00:22.680 And some of them are overloaded. 00:22.720 --> 00:25.840 Terms that we use in other contexts like host can mean a lot of things. 00:25.840 --> 00:27.560 So just to be very clear. 00:27.720 --> 00:32.320 So a host means like a host environment, a host application. 00:32.440 --> 00:39.320 And it's something that the most common host that is referred to is the clawed desktop app, which is 00:39.320 --> 00:42.680 often used as the host for MCP. 00:43.200 --> 00:48.760 And the host represents the place where the application software is running. 00:48.760 --> 00:52.960 That's going to take advantage of these tools that we're going to equip it with. 00:53.120 --> 01:01.120 So, uh, Clawed Desktop is a typical example you can't use, at least as of now, the clawed chat, 01:01.160 --> 01:03.490 the web based chat, you have to use the desktop. 01:03.650 --> 01:10.050 But another example would be our own agent architectures, our own agent solutions that we will build 01:10.210 --> 01:12.010 in cursor and run it there. 01:12.050 --> 01:19.330 And since I mentioned cursor, cursor itself is also actually a host in that cursor can connect to MCP 01:19.370 --> 01:19.850 servers. 01:19.850 --> 01:22.850 So a host can also be the cursor environment. 01:22.890 --> 01:29.290 It can be a generic applications that we write like the one we did in Autogen last week, and like the 01:29.290 --> 01:32.490 one we're going to do in using open AI agents SDK. 01:32.890 --> 01:35.330 And it can be your desktop that is the host. 01:35.330 --> 01:36.490 It's an application. 01:37.490 --> 01:42.970 An MCP client is a is something which runs within the host. 01:43.250 --> 01:46.290 And there are there can be many MCP clients. 01:46.290 --> 01:54.330 In fact, you need an MC client for every MCP server, every set of tools that you're going to use in 01:54.330 --> 01:56.610 your host, like your desktop. 01:56.810 --> 02:03.820 So supposing you wanted to give Claude desktop access to Google Maps, Then you would need an MCP client 02:03.820 --> 02:09.180 for Google Maps running inside the the host inside your desktop. 02:09.180 --> 02:13.540 If you wanted to access the local file system, that would be another MCP client. 02:13.700 --> 02:18.140 If you wanted to look up stock prices, that would be another MCP client. 02:18.340 --> 02:22.100 You'd have each of these sets of tools offered by somebody. 02:22.460 --> 02:28.060 They would be, uh, managed by an MCP client living inside that host. 02:28.060 --> 02:32.420 And if for some reason you had a couple of different hosts, you had Claude Desktop and you also had 02:32.420 --> 02:37.580 cursor running some tools, and you also had an agent app, then all three of these hosts would all 02:37.580 --> 02:41.780 have their own MCP clients for each MCP server. 02:41.980 --> 02:44.980 So now that I've said MCP server, I better introduce that. 02:44.980 --> 02:54.020 So an MCP server is the actual process which provides tools and context or resources as the proper MCP 02:54.060 --> 03:00.380 word for IT tools, resources represented, context and prompts, and most importantly, tools. 03:00.380 --> 03:02.380 That's what we'll mostly use. 03:02.800 --> 03:09.560 uh, and an MCP server is one particular offering that supplies a group of tools. 03:09.560 --> 03:15.560 It might be like a file system MCP server that gives tools that give you access to the file system. 03:15.560 --> 03:20.960 It might be a weather MCP server that is able to look up the weather for any given city using a weather 03:20.960 --> 03:21.560 API. 03:21.840 --> 03:27.000 So these different MCP servers are these processes which can carry out some capabilities. 03:27.000 --> 03:30.160 And they advertise, hey, I can do these tools. 03:30.360 --> 03:33.640 I can look up these resources, I can give you these prompts. 03:33.920 --> 03:38.560 And an MCP client running in a host takes advantage of it. 03:38.600 --> 03:41.800 And don't worry if you're a visual type, there's going to be a visual in just a second. 03:42.440 --> 03:50.320 Uh, and so yeah, to give a quick example, fetcher is the name of an MCP server that can fetch web 03:50.360 --> 03:50.840 pages. 03:50.840 --> 03:56.560 And the way it does it behind the scenes is it actually uses that playwright Microsoft product that 03:56.560 --> 04:02.560 brings up a browser heedlessly without a screen and is able to to collect a web page. 04:02.680 --> 04:05.050 And so that is the name of an MCP server. 04:05.610 --> 04:08.650 And you can configure Claude Desktop. 04:08.650 --> 04:14.850 If you want to use that you have to set up an MCP client inside Claude. 04:15.090 --> 04:21.490 And that will launch the fetcher MCP server on your computer as a separate process. 04:21.970 --> 04:25.210 And we actually did that last week with Autogen. 04:25.530 --> 04:29.050 That was the thing we did, which I sort of talked quickly through. 04:29.050 --> 04:36.610 But we we launched we can we used some autogen code to have an MCP client running within our Autogen 04:36.650 --> 04:44.570 agent, and that spawned a process on our box, which was an MCP server that was the fetcher server 04:44.570 --> 04:47.650 and that was able to bring up a headless browser. 04:48.370 --> 04:52.290 So hopefully you're halfway following me, mostly following me, hearing these new terms. 04:52.290 --> 04:55.610 It's going to be much clearer when we see it visually. 04:55.650 --> 04:56.450 Okay, let's do it. 04:56.450 --> 04:57.490 Let's look at it visually. 04:57.490 --> 05:02.050 The architecture diagram you've been waiting for, and you may have seen this a few times because anyone 05:02.050 --> 05:06.100 that's done any googling and people have seen this, but maybe now that we've walked through the terms 05:06.100 --> 05:07.300 and stuff, it's going to become clearer. 05:07.300 --> 05:10.540 And then I'm going to point out a few things that are like a common misconception. 05:10.540 --> 05:12.580 So you have your computer. 05:12.620 --> 05:14.300 That's what that big box there is. 05:14.300 --> 05:18.900 It represents your machine, your box, which is going to be an important player here. 05:19.220 --> 05:24.020 Uh, that blue box, there is a host running on your computer. 05:24.020 --> 05:29.020 Remember, whilst sometimes people use the word host to refer to a whole computer, in this case we're 05:29.020 --> 05:34.460 talking about like a host application, an MCP host is like an application that could be your desktop, 05:34.500 --> 05:40.500 it could be your agent application, it could be a cursor, the application, or many other applications 05:40.500 --> 05:43.980 that are able to to take advantage of MCP. 05:44.900 --> 05:47.620 And then we collect a couple of MCP servers. 05:47.620 --> 05:50.180 Maybe the one on the bottom left is an MCP server. 05:50.180 --> 05:53.380 That's to do with using the local file system. 05:53.420 --> 05:58.500 And the one on the bottom right is one that is able to to look for the weather using Google's weather 05:58.540 --> 05:59.180 API. 05:59.460 --> 06:01.420 Let's say, uh, okay. 06:01.860 --> 06:08.510 Well then within the host we will need to have two MCP clients, because there needs to be an MCP client 06:08.510 --> 06:11.270 in the host for every MCP server. 06:11.270 --> 06:12.870 And so that's what we have. 06:13.230 --> 06:18.230 And really those white arrows you see there, that is what MCP actually is. 06:18.270 --> 06:19.750 That is the protocol. 06:19.750 --> 06:26.910 It's a protocol for talking from a thing running in your host called the MCP client, to a process running 06:26.910 --> 06:29.710 outside your host called the MCP server. 06:29.710 --> 06:32.950 And anthropic have given libraries to help build these. 06:33.110 --> 06:37.670 But really, the MCP protocol itself is just the way that they communicate. 06:38.710 --> 06:44.550 And let's suppose we also have a remote server, a box somewhere else. 06:45.030 --> 06:52.910 Then that could also be running an MCP server, and you could have another MCP client on your host that's 06:52.910 --> 06:58.070 talking remotely to that MCP server that's running on a different server. 06:58.070 --> 07:00.270 And that would be another configuration. 07:00.670 --> 07:03.350 And so that is also something that that happens. 07:03.710 --> 07:07.360 Um, so those are the different ways that it can fit together. 07:07.480 --> 07:13.080 And now, just before I explain that the misconception that some people have about MCP, let me just 07:13.080 --> 07:14.760 quickly clarify that. 07:14.800 --> 07:20.960 You remember I said that that that box on the bottom right could be, say, an MCP server that's looking 07:20.960 --> 07:24.720 up the weather in Google using Google's weather API. 07:25.160 --> 07:30.040 Well, of course, what that means is that there is, in fact another remote computer, Google's remote 07:30.040 --> 07:36.200 computer, in this case, and we are calling a remote process on that computer to look up the weather 07:36.240 --> 07:37.880 passing in Google's API. 07:38.880 --> 07:41.160 So it is going off to another server. 07:41.600 --> 07:46.560 But in that architecture, the MCP server is running on our box. 07:47.120 --> 07:49.120 So there's two ways that this could happen. 07:49.160 --> 07:51.680 When you can see there's three different ways. 07:51.680 --> 07:53.520 Let's talk about all three different ways. 07:53.680 --> 07:57.440 On the bottom left you've got example of where the client is on. 07:57.440 --> 08:00.960 Our host is on our box, the server is on our box. 08:00.960 --> 08:04.160 And the work it's doing is happening right there on our box. 08:04.320 --> 08:05.920 That is one model. 08:06.420 --> 08:12.060 The second model is if you come a long one and you see the client is on our box, they always are. 08:12.420 --> 08:19.260 The server is also on our box, but to do its work, it's calling out to an external API like like the 08:19.260 --> 08:19.740 weather. 08:20.300 --> 08:24.780 Or it could be looking at your Google Drive, or it could be looking up stock prices or whatever. 08:24.780 --> 08:30.380 There's many, many, many MCP servers that run on your box, but call out to different APIs. 08:30.380 --> 08:34.860 But then the MCP server turns it into a tool which can be used on your host. 08:35.220 --> 08:37.700 And then the third model is that model. 08:37.740 --> 08:39.580 Now we're moving anti-clockwise. 08:39.580 --> 08:44.900 The third model is that model on the on the top right which is saying, okay, you have an MCP client 08:44.900 --> 08:50.020 on your box as always, and it's calling out to an MCP server running somewhere else. 08:50.060 --> 08:51.980 And that is the third model. 08:52.180 --> 08:54.180 And so what is the misconception? 08:54.180 --> 09:02.100 Well, the misconception is that I think when people hear about MCP servers and they hear that MCP servers 09:02.100 --> 09:08.190 are like open source and and hosted on various hubs that will have a look at they sort of assume that 09:08.190 --> 09:15.110 the architecture is that top right architecture that most of the time MCP clients are calling out to 09:15.150 --> 09:20.950 MCP servers running somewhere else, running remotely, because the word server sounds like you're talking 09:20.950 --> 09:22.790 about it running on a remote server. 09:23.190 --> 09:25.550 But that's just a bit of a of a confusion. 09:25.670 --> 09:32.550 Most of the time, it's actually quite rare that you deal with remote MCP servers. 09:33.310 --> 09:36.190 By far the most common, at least in my experience. 09:36.230 --> 09:43.030 I'd shout if this isn't your experience, but almost all the ones that I work with are either the one 09:43.070 --> 09:47.350 on the bottom left, either that model or the one next to it, and most commonly the one next to it. 09:47.390 --> 09:54.870 Most commonly, you're running an MCP server that is calling out to get stock prices to look on Google 09:54.870 --> 09:57.750 Drive to post on slack to do whatever it's doing. 09:57.750 --> 10:00.310 It's doing that by connecting remotely. 10:00.310 --> 10:05.070 But that connection is happening in an MCP server that's running on your box. 10:05.070 --> 10:07.720 That's the most common Configuration. 10:07.720 --> 10:12.400 And it's confusing because, uh, because MCP servers are provided online. 10:12.440 --> 10:13.640 Like, they're open source. 10:13.640 --> 10:15.320 You go online, there's lots of them. 10:15.320 --> 10:16.280 And people talk about that. 10:16.480 --> 10:16.960 They. 10:17.000 --> 10:18.800 Host MCP servers. 10:18.800 --> 10:22.880 But what they're talking about there is that they make MCP servers available. 10:22.960 --> 10:30.480 For you to download and run them locally on your box in order to call out to an API just like that. 10:30.520 --> 10:34.840 That diagram on the bottom, just like that diagram, that's the most common. 10:35.000 --> 10:36.520 And there are indeed cases. 10:36.560 --> 10:40.840 When you call MCP servers running remotely, and we will do one of them this week for sure. 10:40.840 --> 10:42.800 So you get to see that and experience it. 10:42.840 --> 10:44.440 And understand what that's like. 10:44.680 --> 10:48.400 But in my experience anyway, that is not as common as the. 10:48.440 --> 10:49.920 Other configurations. 10:50.080 --> 10:53.760 And I don't think people explain this this distinction often. 10:53.800 --> 10:55.120 And as I say, because. 10:55.160 --> 11:00.920 An MCP server is called a server, you might naturally think that it runs remotely when most of the 11:00.920 --> 11:01.480 time. 11:01.520 --> 11:06.240 The the client obviously runs in a box and so does the server most of the time. 11:06.280 --> 11:07.320 And so there's one more. 11:07.560 --> 11:10.890 Technical point to make about this, which is worth knowing about. 11:10.890 --> 11:15.050 You don't need to know this in any detail, but you'll see these terms all the time. 11:15.050 --> 11:16.850 So it's it's good to recognize them. 11:16.850 --> 11:22.930 At least there are these two different transport mechanisms, which are two different ways that clients 11:22.930 --> 11:25.090 and servers can communicate to each other. 11:25.090 --> 11:27.410 And in theory, there might be many more in the future. 11:27.450 --> 11:33.170 The protocol would allow for many different types of ways, but as of now, anthropic has has produced 11:33.170 --> 11:37.450 two different, uh, technical approaches for them talking to each other. 11:37.770 --> 11:43.090 And the most common and the one that needs to be used, uh, that can only be used locally, uh, for, 11:43.130 --> 11:47.850 for these kinds of connections that we've got on the, on the, on the screen right now, the first 11:47.850 --> 11:55.490 two yellow boxes, uh, is called studio or Stdio, and it stands for standard input and output. 11:55.490 --> 11:58.370 And this transport mechanism is very simple. 11:58.530 --> 12:04.570 The client spawns a new process, spawns a process to run the server. 12:04.570 --> 12:12.030 It launches something and it connects to that using that processes standard input and output just uses 12:12.070 --> 12:13.910 the input output. 12:14.070 --> 12:19.350 And that is how it sends its messages and how it gets its responses from that process. 12:19.590 --> 12:24.670 Um, and you know, that's you don't need to know that technical detail, but you do need to recognize 12:24.710 --> 12:30.070 the stdio, because when we work with this, you'll see that there'll be the terms like server stdio 12:30.150 --> 12:35.230 that will be creating to connect from our MCP clients to our MCP server. 12:35.390 --> 12:38.630 And uh, so that is one of the two mechanisms. 12:38.630 --> 12:41.270 And the other mechanism is called SSE. 12:41.790 --> 12:46.870 And SSE is saying that the connection will be over Https. 12:46.910 --> 12:49.630 So it's going to be a secure HTTP connection. 12:49.630 --> 12:52.750 And the way that data will be sent is by posting. 12:52.750 --> 12:59.470 And the way it will be received back is using a technology called Server Side Events SSE, which is 12:59.510 --> 13:06.990 actually the same technology that's used by AI labs to stream back results when you're using the API 13:06.990 --> 13:08.550 or when you're looking in the screen. 13:08.550 --> 13:14.240 It's a it's a it's a technique for long running HTTP queries that allows you to sort of drip feed back 13:14.240 --> 13:16.880 data, which is how the streaming works. 13:16.880 --> 13:20.680 So SE is the name of this connection type and it uses SSH. 13:21.440 --> 13:23.600 And uh, it's, it's a type. 13:23.600 --> 13:29.880 I do believe that you can use this transport either for remote or for, for the, for the local ones 13:29.880 --> 13:30.440 as well. 13:30.800 --> 13:34.600 Um, but you can't use stdio for the remote ones. 13:34.800 --> 13:36.320 Uh, you're not spawning a process. 13:36.320 --> 13:38.040 Obviously, it's a remote box. 13:38.240 --> 13:45.640 So, um, but generally speaking, what you see is that stdio is used for the local, uh, use cases. 13:45.680 --> 13:49.800 The first two here and SSH is used when you go remote. 13:49.800 --> 13:52.240 So we will do that, uh, this week at one point. 13:52.240 --> 13:57.600 But mostly we'll stick with stdio, which is the most common one that there is. 13:57.640 --> 14:02.280 Okay, that's enough concepts, enough terms, enough diagrams, enough architecture. 14:02.320 --> 14:03.160 I know what you want. 14:03.160 --> 14:03.960 You want to see. 14:04.000 --> 14:05.080 You want to see it happening. 14:05.080 --> 14:06.480 You want to see some code? 14:06.680 --> 14:07.680 Let's go to cursor. 14:07.680 --> 14:08.640 Let's see some code. 14:08.640 --> 14:10.080 Let's work with mic.