WEBVTT 00:00.280 --> 00:05.640 And you should be starting to feel really quite knowledgeable about MCP and MCP servers as we head into 00:05.680 --> 00:10.200 day three of our journey with MCP in week six. 00:10.440 --> 00:13.440 And one more time, I'm going to throw up the architecture slide. 00:13.440 --> 00:15.760 At this point, it's probably familiar to you. 00:16.120 --> 00:21.760 It's showing the computer with its host, potentially multiple hosts, and the host, which is something 00:21.760 --> 00:28.440 like our agents application ourself contains MCP clients, each containing each connecting to an MCP 00:28.480 --> 00:29.120 server. 00:29.360 --> 00:32.240 And we're going to be looking at different MCP servers today. 00:32.240 --> 00:35.280 We're going to be bringing in a whole zoo of them. 00:35.520 --> 00:38.680 And there's going to be all three kinds. 00:38.680 --> 00:40.080 We'll look at the first kind. 00:40.080 --> 00:40.800 You'll remember. 00:40.840 --> 00:46.200 Perhaps the simplest is when you have an MCP client running on your host, connecting to an MCP server, 00:46.200 --> 00:53.280 running on your computer in a different process outside, typically connecting over stdio. 00:53.360 --> 00:59.000 It can also connect over SSH, but but it's more normal to use the standard input output into that process 00:59.000 --> 00:59.920 that is spawned. 01:00.280 --> 01:03.120 And then the second configuration is basically the same thing. 01:03.120 --> 01:07.840 But we're just saying that this would be an example of when you have a server which isn't doing all 01:07.880 --> 01:11.560 of its processing locally, it's calling out to the internet to do it. 01:11.560 --> 01:13.680 It's using a web service or something like that. 01:14.040 --> 01:21.240 And then the third configuration, not as common, is when we call out to a server running remotely, 01:21.280 --> 01:27.840 an MCP client running on your computer in its host calling out to a remote MCP server. 01:28.280 --> 01:31.600 Okay, so let's just go and do all three of them right away. 01:31.640 --> 01:33.040 And we're in cursor. 01:33.040 --> 01:36.160 We're going to the sixth directory MCP. 01:36.160 --> 01:38.320 And we're going to the third lab. 01:38.520 --> 01:39.080 Welcome. 01:39.080 --> 01:40.400 Week six day three. 01:40.800 --> 01:45.320 And we do some usual imports including loading dot env. 01:45.600 --> 01:46.280 All right. 01:46.320 --> 01:47.040 Type one. 01:47.240 --> 01:51.600 Type one is going to be the MCP server that runs locally. 01:51.600 --> 01:53.080 And everything is local. 01:53.080 --> 01:54.600 And we're going to use a cool one. 01:54.800 --> 01:58.400 It's a knowledge graph based memory. 01:58.400 --> 02:01.640 So it's something which remembers information. 02:01.800 --> 02:09.830 And it does so by storing entities observations about those entities and relationships between the entities. 02:10.030 --> 02:17.550 And it's an MCP server that provides a suite of tools that can be used by our server to achieve all 02:17.550 --> 02:18.630 of those things. 02:19.150 --> 02:27.390 So it's one of these it's it's the parameters is we're making the mpc's command to run a node and we're 02:27.390 --> 02:28.230 passing in. 02:28.430 --> 02:30.750 This is actually an official anthropic implementation. 02:30.750 --> 02:31.670 So it's from them. 02:31.710 --> 02:35.110 It's from a model context protocol themselves. 02:35.110 --> 02:37.350 And it's called server memory. 02:37.510 --> 02:39.750 And let's collect the tools. 02:39.750 --> 02:41.630 And as always it's just as simple as this. 02:41.630 --> 02:46.630 This is the beauty of MCP that just a couple of lines of code like that and we're connecting to a bunch 02:46.630 --> 02:47.350 of tools. 02:48.110 --> 02:50.510 And so there's one called create entities. 02:50.510 --> 02:56.030 Create new entities in the Knowledge Graph, create relations, add observations, delete entities, 02:56.030 --> 03:01.910 delete observations, delete relations, read graph search nodes, open nodes. 03:01.910 --> 03:04.350 So you get a sense of what this is about. 03:04.670 --> 03:05.030 Okay. 03:05.070 --> 03:06.990 So now we have some instructions. 03:07.670 --> 03:14.150 I'm saying use your entity tools as a persistent memory to store and recall information about your conversations. 03:14.150 --> 03:18.710 So you do have to give it that, that that prepping because we're just arming it with all these tools. 03:18.710 --> 03:21.670 We're not necessarily telling it what to use them for. 03:21.830 --> 03:27.910 And as they explain in the documentation, in your instructions, you need to make it clear how frequently 03:27.910 --> 03:33.470 you expect it to use its memory or give it some, some kind of context to work with in terms of how 03:33.470 --> 03:35.990 to relate what it has to do with these tools. 03:36.750 --> 03:39.990 So my request my name is Ed, I'm an LM engineer. 03:39.990 --> 03:46.150 I'm teaching a course about AI agents, including the incredible MCP protocol, and SAP is a protocol 03:46.150 --> 03:51.310 for connecting agents with tools, resources, and prompt templates and makes it easy to integrate AI 03:51.350 --> 03:53.710 agents with capabilities. 03:53.710 --> 04:01.310 That is some a request that we are going to make a prompt to this LM so let's do that. 04:01.350 --> 04:10.870 And this is the familiar code again so simple with we frame it with the with MCP service studio passing 04:10.870 --> 04:13.550 in the parameters to to associate with this. 04:13.950 --> 04:20.830 Um, and then we simply call our agent with the system prompt, with the model with the server that 04:20.830 --> 04:22.430 we've just created. 04:22.710 --> 04:27.230 And then we run it or run passing in this request, this, this prompt right here. 04:27.230 --> 04:28.830 So that is what I'm going to do right now. 04:28.830 --> 04:33.310 And that is running, doing its thing, thinking stuff is happening. 04:35.310 --> 04:36.550 Hopefully good stuff. 04:36.590 --> 04:37.550 It's certainly take its time. 04:37.550 --> 04:38.070 But there we go. 04:38.190 --> 04:41.470 Okay, I've noted your details and information about the MCP protocol. 04:41.470 --> 04:42.590 Here's what I've recorded. 04:42.870 --> 04:45.350 Uh, me the MCP protocol agents. 04:45.350 --> 04:47.510 And you can see it's got like observations. 04:47.830 --> 04:50.750 And if there's anything else you'd like to add or modify, let me know. 04:51.070 --> 04:52.030 So that's cool. 04:52.310 --> 04:55.470 Well, let's find out if it really has remembered any of this. 04:55.470 --> 05:00.910 So we are going to, uh, make make a we're going to pass in the same instructions, the same system 05:00.910 --> 05:02.830 prompt, this system prompt. 05:02.870 --> 05:04.950 There's nothing about me in that system prompt. 05:05.310 --> 05:08.550 And then as the user prompts, we're going to say, my name is Ed. 05:08.590 --> 05:09.670 What do you know about me? 05:10.190 --> 05:13.630 So we're calling the stateless GPT four mini. 05:13.670 --> 05:15.630 It doesn't have any other context about me. 05:15.870 --> 05:18.510 Let's see how it responds to that question. 05:20.350 --> 05:21.350 Here's what I know about you. 05:21.390 --> 05:21.590 Editor. 05:21.590 --> 05:22.710 You're an LM engineer. 05:22.750 --> 05:24.270 You teach a course about AI agents. 05:24.270 --> 05:26.350 You're involved with the MCP protocol. 05:26.950 --> 05:31.790 Additionally, you teach about AI agents, which are intelligent systems capable of performing tasks 05:31.790 --> 05:33.870 autonomously and utilized for enhanced functionality. 05:33.870 --> 05:34.350 There you go. 05:34.390 --> 05:35.790 It seems pretty decent to me. 05:36.110 --> 05:41.270 Uh, I guess you're involved with the MCP protocol as perhaps taking it a bit far, but never mind. 05:41.310 --> 05:42.550 We'll we'll give it. 05:42.590 --> 05:43.590 We'll give it that one. 05:43.790 --> 05:45.150 Uh, and, uh. 05:45.350 --> 05:45.630 Yeah. 05:45.670 --> 05:47.950 Have a, have a, uh, hopefully you're playing with this yourself. 05:47.990 --> 05:49.390 Have an experiment with it. 05:49.550 --> 05:51.310 Uh, and we'll look at the trace in a sec. 05:51.550 --> 05:56.070 And just before we look at the trace, let's just look at the documentation for a second for this, 05:56.110 --> 05:57.870 which is right here. 05:58.110 --> 06:04.470 Uh, so this, uh, describes using the graph, uh, memory server tells you about the concepts, what 06:04.470 --> 06:05.910 to do, how to use it. 06:06.030 --> 06:12.390 Uh, the fact that that it stores relations and observations in addition to entities and how that works. 06:12.390 --> 06:14.470 So it's interesting to read through that. 06:14.470 --> 06:15.470 So take a look. 06:15.790 --> 06:19.940 Uh, and then we will now look at our trace. 06:20.460 --> 06:22.180 Uh, where is that now? 06:22.180 --> 06:22.780 Here we go. 06:22.820 --> 06:23.980 Look at the trace. 06:23.980 --> 06:28.260 So this is again going into platform Openai.com traces. 06:28.540 --> 06:30.660 And this is the conversation right here. 06:31.020 --> 06:35.100 So you can see that it, uh, it called search nodes. 06:35.100 --> 06:40.700 This is for the um, the, the second, the second call to, to read in its memory. 06:40.820 --> 06:46.540 And you can see that it read in some observations that I'm teaching with a course about, I'm involved 06:46.540 --> 06:51.740 with MCP protocol and it got its information from its memory. 06:51.940 --> 06:54.380 Uh, so there you have it. 06:54.380 --> 07:02.980 There you can see the, uh, the use of a type one of a, of an MCP, uh, server that was using local 07:02.980 --> 07:09.180 storage and using that adding functionality to store a memory in that format. 07:09.180 --> 07:16.140 And that means that just just through using this MCP server, you can now add this kind of structured, 07:16.420 --> 07:20.140 relationship driven memory to any of your agents.