WEBVTT 00:00.200 --> 00:00.640 Okay. 00:00.640 --> 00:02.080 Welcome back to the lab. 00:02.080 --> 00:07.360 Then we're at week two, day two, and we're going to run some imports and the load dot env. 00:07.680 --> 00:09.880 And now we're going to import account. 00:09.880 --> 00:11.640 This is the business logic class. 00:11.640 --> 00:14.760 Nothing to do with MCP servers yet I just want to show you this. 00:14.960 --> 00:19.520 When I import it it automatically creates a new SQL database with SQLite. 00:19.520 --> 00:21.680 Because that was in the code at the top. 00:21.680 --> 00:24.400 When you import, if it doesn't already exist, it creates one. 00:24.680 --> 00:30.640 And also if I get an account that doesn't exist, it creates a default account with $10,000 in it, 00:30.680 --> 00:31.480 which is kind. 00:31.560 --> 00:32.320 So there we go. 00:32.400 --> 00:33.440 This is all a simulation. 00:33.480 --> 00:33.760 Of course. 00:33.800 --> 00:37.760 It's just just to have something stored so we can watch trading happening. 00:37.920 --> 00:42.160 So here we go a new account called Ed with a fake balance of $10,000. 00:42.200 --> 00:45.640 No holdings, no trades, no time series. 00:45.880 --> 00:50.920 And we're now going to buy three shares of Amazon in this fake simulation. 00:51.120 --> 00:54.480 And we're going to pretend that I'm thinking of doing this years ago. 00:54.680 --> 01:01.080 And my my rationale for doing it is because this bookstore website looks promising. 01:01.600 --> 01:03.840 Oh, if only I had had that foresight. 01:04.120 --> 01:05.520 Uh, so there we go. 01:05.560 --> 01:09.000 We buy shares completed, and it gives some details of that. 01:09.000 --> 01:10.360 The balance, the strategy. 01:10.560 --> 01:15.720 And I should say it actually, there's some code in here that looks up the current share price of Amazon. 01:15.720 --> 01:18.360 And later we'll come back and I'll talk about how we do that. 01:19.040 --> 01:22.920 Then we are going to, uh, ask for a report of the account. 01:22.960 --> 01:28.040 And here it is with the strategy of the transactions, the rationale. 01:28.040 --> 01:33.000 And over at the end here, there is a total portfolio value and a profit and loss. 01:33.040 --> 01:38.280 And you'll notice that the total portfolio value is slightly less than the $10,000 it was originally. 01:38.280 --> 01:41.120 And there's a tiny negative, uh, profit and loss. 01:41.440 --> 01:47.680 And for those that care about this, that want to dig into this, uh, I have made a small assumption, 01:47.720 --> 01:54.200 a fixed assumption of, um, a fee that you pay every time that you trade in the form of what they call 01:54.240 --> 01:56.200 a a bid offer spread. 01:56.200 --> 01:57.400 And you'll see that in the code. 01:57.400 --> 02:03.050 It's just a very simplistic assumption to mean that every time you trade, you lose a little bit. 02:03.530 --> 02:03.970 Okay. 02:04.010 --> 02:04.810 So let's list. 02:04.850 --> 02:06.730 Transactions in our portfolio. 02:06.770 --> 02:08.530 We have three shares of Amazon at. 02:08.570 --> 02:09.370 That much each. 02:09.570 --> 02:12.410 And that is a quick demo of the business logic. 02:12.450 --> 02:19.050 The next step is to write is to is to use our MCP server and see if we can access this directly through 02:19.050 --> 02:20.170 the MCP server. 02:20.570 --> 02:21.170 All right. 02:21.170 --> 02:22.730 So how hard do you think it will be. 02:22.770 --> 02:24.490 It's going to be very easy. 02:24.850 --> 02:27.090 Uh so remember this params. 02:27.130 --> 02:30.130 Params is how you say what it is that you want to do. 02:30.170 --> 02:32.250 What what server do you want to run. 02:32.490 --> 02:35.450 And literally the command is you've. 02:35.730 --> 02:37.450 And the arguments is run. 02:37.450 --> 02:40.330 And then account server dot pi. 02:40.610 --> 02:43.650 It's just going to spawn a process. 02:43.970 --> 02:45.170 It's going to do it itself. 02:45.170 --> 02:46.570 It's going to run this process. 02:46.570 --> 02:49.450 You've run account server.py. 02:50.210 --> 02:50.770 Okay. 02:51.130 --> 02:56.530 Now I have this line async with MCP server studio as server. 02:56.530 --> 02:58.170 So what is MCP Server Studio. 02:58.170 --> 03:04.410 So this this is a class from OpenAI agents SDK. 03:04.770 --> 03:14.090 And it's basically some code which essentially creates an MCP client and attaches to an STD IO server 03:14.170 --> 03:15.570 that you want to run. 03:15.570 --> 03:21.490 It will it will spawn that server and then run it and allow us to do something with it. 03:21.490 --> 03:25.170 So all of that is simply in this one class. 03:25.490 --> 03:31.050 And as I think I mentioned before, this is actually relatively new in OpenAI agents SDK. 03:31.090 --> 03:36.930 And as of just recently, you would have to do that last part yourself of writing the code that spawns 03:36.930 --> 03:37.570 the server. 03:37.850 --> 03:41.210 And fear not, we will do that anyway, just so you see that side of it. 03:41.210 --> 03:44.890 But you don't need to because OpenAI agents SDK has written that for you. 03:45.050 --> 03:49.810 So this is the client and spawns that service. 03:50.090 --> 03:56.610 And then we are just going to call list tools, just as we did for fetch, for all of the other tools 03:56.610 --> 03:58.090 that we used last time. 03:58.090 --> 03:59.130 This is the same code. 03:59.130 --> 04:00.610 It's just three lines. 04:00.620 --> 04:04.220 we run it, we get the MCP tools and here they are. 04:04.340 --> 04:08.780 Get balance, get holdings, buy shares, sell shares with a description. 04:08.780 --> 04:14.140 And then you'll see if we scroll along that there are going to be all of the different bits and pieces 04:14.140 --> 04:14.740 to it. 04:14.900 --> 04:24.180 So if we now just remind ourselves if we go to the accounts server, you'll see here is our MCP server. 04:24.180 --> 04:30.740 And so get balance right here was one of the one of the decorated functions from our server. 04:30.740 --> 04:36.460 And if we go back here you'll see that get balance with that description is exactly what has been surfaced 04:36.460 --> 04:36.980 here. 04:37.180 --> 04:40.860 And so it really is that straightforward okay. 04:40.860 --> 04:43.860 So we've seen how easy it is to list the tools. 04:43.900 --> 04:45.140 Now let's use them. 04:45.340 --> 04:49.100 So it really is going to be again very straightforward. 04:49.380 --> 04:54.780 Uh writing MCP servers is a little bit of a few extra decorators here and there. 04:54.820 --> 04:56.540 Using them is a cinch. 04:57.060 --> 05:00.460 So here here's uh, some instructions or a system prompt. 05:00.460 --> 05:04.140 You are able to manage an account for a client and answer questions about the account. 05:04.180 --> 05:04.860 My request. 05:04.860 --> 05:05.900 My name is Ed. 05:05.940 --> 05:07.500 My account is under the name Ed. 05:07.540 --> 05:09.540 What's my balance and my holdings? 05:09.660 --> 05:11.540 Model GPT four mini. 05:11.780 --> 05:12.740 That sets us up. 05:12.980 --> 05:14.020 And so here it is. 05:14.020 --> 05:17.780 This is the code with MCP server studio. 05:17.820 --> 05:24.420 That is the OpenAI agents SDK thing passing in the parameters of our server, caching the tools. 05:24.420 --> 05:27.940 Why not uh, as uh, as MCP server. 05:28.300 --> 05:32.860 We don't need to do this, particularly for one call like this, especially to to show you how simple 05:32.860 --> 05:33.780 the code looks. 05:34.140 --> 05:38.580 Uh, so with an we create an agent called account manager. 05:38.580 --> 05:44.340 We give it the system, prompt the model, and we pass in the server, the MCP server that we want to 05:44.380 --> 05:48.140 use, which is our MCP server, our very own homemade one. 05:48.140 --> 05:54.860 And then we put in a trace and we say runner dot run with the agent and we pass in our request and we'll 05:54.860 --> 05:56.060 display the final output. 05:56.060 --> 05:57.940 The question is is this going to work? 05:57.940 --> 05:58.700 What it's going to do? 05:58.700 --> 06:04.830 If it works, it's going to create an MCP client, which will then spawn an MCP server. 06:04.830 --> 06:10.910 It will basically call you've run account Server.py launching that process. 06:10.910 --> 06:16.230 It will talk to that process using standard input and output, and it will collect the tools. 06:16.230 --> 06:22.390 It will pass those tools into our agent by constructing the JSON gumph that will go in. 06:22.630 --> 06:28.790 And when the LM says it wants to use a tool, then this code will then run that using standard input 06:28.790 --> 06:32.190 output to ask our spawned process to run that tool. 06:32.350 --> 06:35.790 And then the results will come back and be given to the LM which will complete. 06:35.790 --> 06:42.030 So there is a lot of stuff happening backwards and forwards that's all been elegantly wrapped into this. 06:42.070 --> 06:42.790 Will it work? 06:42.790 --> 06:50.190 Will we find out our holdings in natural language from asking GPT four mini based on our very own server? 06:50.470 --> 06:51.470 Yes we will. 06:51.790 --> 06:55.390 Your current balance is 9470. 06:55.430 --> 06:59.030 In terms of holdings, you have three shares of Amazon. 06:59.470 --> 07:01.350 That is mission accomplished.