WEBVTT 00:00.160 --> 00:03.560 And so we create another a new agent. 00:03.600 --> 00:05.200 The name is account manager. 00:05.240 --> 00:08.760 The instructions as are system prompt, the model as the model. 00:08.760 --> 00:10.120 And the tools we're passing in. 00:10.120 --> 00:12.520 We're no longer passing in an MCP server. 00:12.560 --> 00:17.120 We're passing in the tools that we've just reconstituted here. 00:17.440 --> 00:21.480 And the thing and this might be a little bit of a of a of a leap. 00:21.520 --> 00:28.880 The thing to keep in mind is that within those tools there is the function to be called, and that function 00:28.880 --> 00:31.440 to be called if the tool needs to be triggered. 00:31.560 --> 00:36.760 If I go back to the client, the function to be called that we've set up here, you'll need to look 00:36.760 --> 00:37.800 through this code. 00:38.000 --> 00:45.040 But it is to to to call back to this call accounts tool with the tool name in question and its arguments. 00:45.440 --> 00:53.080 So the bottom line is that if the LM chooses to run one of these tools, that's going to end up calling 00:53.240 --> 00:59.360 our MCP client, which is going to call our MCP server, which is going to call our business logic for 00:59.400 --> 01:00.920 that particular tool. 01:01.130 --> 01:05.970 All of that is going to happen, but this time it's going through our MCP client. 01:06.370 --> 01:08.330 And so we're still asking the same question. 01:08.330 --> 01:10.890 We're asking the my name is Ed. 01:10.930 --> 01:11.930 My account is under the name Ed. 01:11.930 --> 01:14.090 What's my balance and my holdings. 01:14.090 --> 01:20.130 But rather than using the MCP client that OpenAI provides us, we've built our own. 01:20.370 --> 01:22.370 And let's see if it works. 01:22.410 --> 01:29.170 So now going through our MCP client, spawning the MCP server and we get back the right balance. 01:29.170 --> 01:31.130 And that I have three shares of Amazon. 01:31.130 --> 01:35.770 So yes it works going through our MCP client and server together. 01:36.210 --> 01:39.730 And I can also show you how what it looks like to read a resource. 01:39.730 --> 01:46.650 So it's using the read accounts resource method function that we have in our client right here. 01:46.810 --> 01:54.370 Read accounts resource which launches a client uh and then creates a session with a server. 01:54.370 --> 02:01.370 So a Python process will launch to run this code right here, which is going to run our MCP server, 02:01.410 --> 02:03.970 which within it imports our business logic. 02:04.450 --> 02:11.490 And then we're going to call Read resource with this with Ed instead of name with this. 02:11.490 --> 02:14.370 And that is going to go to the server on the server. 02:14.370 --> 02:18.650 It's going to call this this read resource function. 02:18.650 --> 02:19.850 You see right here. 02:19.850 --> 02:25.890 And because name is is in the decorator at that point, that is how name is going to be provided. 02:25.890 --> 02:29.650 That is going to call account dot get for the name okay. 02:29.690 --> 02:30.690 What does that do. 02:30.810 --> 02:36.210 Let's go in here that calls uh this method get right here. 02:36.610 --> 02:38.570 And then it will call report on that. 02:38.570 --> 02:43.970 And you can see that the report down here is going to return a string representing the account. 02:43.970 --> 02:48.170 So that is the chain of events that's going to happen if it all works. 02:48.170 --> 02:51.010 When I simply call read accounts resource. 02:51.290 --> 02:54.290 And let's see and there is the result. 02:55.410 --> 03:00.690 And now if you're following all this, if you're not following all this it doesn't massively matter. 03:00.690 --> 03:07.420 The purpose of this is to show you the what goes on behind the scenes, the plumbing behind MCP. 03:08.180 --> 03:13.140 If you are following all this, you might be thinking to yourself, hang on a second. 03:13.420 --> 03:24.020 Okay, I see this, but surely I can also just say from accounts import account like the way it does 03:24.020 --> 03:30.380 that account get from accounts, accounts get report. 03:30.420 --> 03:33.540 I can just run that and I'll get the same thing. 03:34.500 --> 03:35.900 And you're absolutely right. 03:35.900 --> 03:36.500 You can. 03:36.540 --> 03:41.620 This is a rather faster way of doing it than going through an MCB client to an MCP server, to launching 03:41.620 --> 03:47.180 a Python process, to attaching to its input and output, to making that same call and to bringing back 03:47.180 --> 03:48.340 all of the results. 03:48.340 --> 03:55.540 And that's why I say that the magic of MCP is more about how this allows you to share tools and resources 03:55.540 --> 04:01.420 across different places, because if you're just writing it for your own purpose, then this just writing 04:01.420 --> 04:05.620 the business logic and just getting the resource is an easier way to do it, frankly. 04:06.260 --> 04:13.260 The reason you would want to build all of this scaffolding around MCP is because then you've got this 04:13.260 --> 04:22.420 consistent, universal USB-C way that anybody can read the ads account details, uh, just using this 04:22.460 --> 04:23.140 interface. 04:23.140 --> 04:24.580 And that's the big point. 04:25.420 --> 04:28.660 And so for the exercises to wrap it up, a couple of them here. 04:28.700 --> 04:30.700 Make your own MCP server. 04:30.740 --> 04:32.620 There's something you could do that's really simple. 04:32.620 --> 04:38.740 You can make a Python function to return the current date, expose it as an MCP tool, bring it into 04:38.740 --> 04:42.620 an agent and ask the agent the current date and make sure it works all the way through. 04:42.900 --> 04:46.980 It's easy to do, and it's really nice to show you to connect the dots. 04:47.100 --> 04:52.820 And then optionally, if you wanted, you could put an MCP client on top of that as well. 04:52.820 --> 04:54.140 So you can go through an MCP client. 04:54.140 --> 05:01.940 And then what you could do is use a native approach for calling OpenAI's code directly, just as we 05:02.070 --> 05:03.110 did in week one. 05:03.110 --> 05:10.270 Go through and make the JSON stuff and use the handle tool calls and see it working that way through 05:10.310 --> 05:16.910 your client and server so you're not taking advantage of the stuff that comes with the OpenAI agents 05:16.950 --> 05:17.510 SDK. 05:17.630 --> 05:22.750 And then you will really have seen you'll have connected the dots all the way through to what's going 05:22.750 --> 05:25.550 on behind the scenes, but that's an optional exercise. 05:25.550 --> 05:27.870 If you if you get it, you don't need to do that unless you really want to. 05:27.910 --> 05:29.030 The gory detail. 05:29.030 --> 05:35.910 But hopefully this lab has helped you to understand how things work behind the scenes and what it is 05:35.910 --> 05:39.830 that MCP is so good at, and where you perhaps don't need it as much. 05:39.990 --> 05:45.510 Well, with that, we've built an MCP server and MCP client, and we've used it with an agent, and 05:45.510 --> 05:46.630 it was great. 05:47.110 --> 05:51.910 And I hope you now have a pretty decent understanding of how to connect the dots. 05:52.190 --> 05:59.230 Next time we're just going to go crazy and use a bunch of local and remote MCP servers to get to get 05:59.230 --> 06:01.590 into this and to enjoy the ecosystem. 06:01.870 --> 06:03.110 I will see you then.