WEBVTT 00:00.320 --> 00:04.080 So what we're going to do now is create our own MCP client. 00:04.280 --> 00:05.560 And this is a bit unusual. 00:05.560 --> 00:08.200 It's not actually that common to need to write an MCP client. 00:08.200 --> 00:09.160 When would you do it? 00:09.160 --> 00:09.600 You would do it. 00:09.600 --> 00:14.600 If you're building your own application, that's going to be a host, that's going to be able to call 00:14.760 --> 00:15.640 MCP servers. 00:15.640 --> 00:21.480 And in your application, you want to be able to call a number of different, uh, MCP servers. 00:21.480 --> 00:23.800 And so you're going to want to write an MCP client. 00:24.040 --> 00:29.160 You'd also have to do this if you rewound the clock a few weeks and wanted to, to use OpenAI agents 00:29.160 --> 00:31.440 SDK, because they didn't used to have one built into them. 00:31.440 --> 00:33.760 But now they do, which makes this not as useful. 00:34.400 --> 00:37.920 But I did this the first time because I needed to, but no longer. 00:37.920 --> 00:43.400 And if you're using another framework or you're going natively to an LLM and you didn't have, uh, 00:43.480 --> 00:46.080 the client that you could use, then you would have to do this yourself. 00:46.080 --> 00:47.320 And it's pretty straightforward. 00:47.320 --> 00:48.520 So we might as well do it. 00:48.520 --> 00:49.440 Let's go do that now. 00:49.440 --> 00:57.720 So I'm going to go into this, uh, module called Accounts Client, which is our MCP client implementation. 00:57.960 --> 01:03.560 And we're using a bunch of MCP anthropic code from clients, standard Input-output. 01:03.560 --> 01:07.360 We're importing the standard Input-output client from MCP. 01:07.560 --> 01:12.440 We're importing this standard input output server parameters, which is going to use the parameters 01:12.440 --> 01:14.200 in the format that we're used to now. 01:14.400 --> 01:20.400 And this is something I'm importing from OpenAI, because I want to translate between Anthropic's description 01:20.400 --> 01:24.520 of tools for MCP into OpenAI's version for this purpose. 01:25.000 --> 01:29.040 So the parameters that we set up is one of these studio server parameters. 01:29.040 --> 01:33.040 Objects is basically the same parameters we were looking at before. 01:33.200 --> 01:39.880 You've run account server dot Pi uh to to to load to run our server. 01:40.320 --> 01:40.840 Okay. 01:41.120 --> 01:50.280 And so there is here a list tools function which is going to create our client, spawn a server and 01:50.280 --> 01:51.840 get it to list its tools. 01:51.840 --> 01:53.200 And this is how it works. 01:53.520 --> 02:00.360 Uh, you uh, use the context manager with to create a, a standard input output client. 02:00.560 --> 02:03.760 You start a session and then you initialize the session. 02:03.760 --> 02:07.880 And then we call this list tools on our session. 02:07.880 --> 02:09.440 And we get back the tools. 02:10.080 --> 02:14.560 We can also call a tool using this call tool function. 02:15.230 --> 02:19.390 We can read a resource, and this is one that we will actually continue to use. 02:19.390 --> 02:21.550 We'll use this approach for reading resources. 02:21.670 --> 02:22.510 Read a resource. 02:22.510 --> 02:28.030 We start the client the session and then we call read resource on our server. 02:28.790 --> 02:35.110 And so each of these three that you see here to list the tools to call a tool and to read the account's 02:35.150 --> 02:36.870 resource to read a resource. 02:37.110 --> 02:41.590 What they're all doing is they are spawning a Python process. 02:41.590 --> 02:46.070 They're calling this which which triggers which spawns another process. 02:46.070 --> 02:49.270 They're connecting to that process using its standard input output. 02:49.270 --> 02:50.830 That's how they're communicating. 02:50.830 --> 02:57.190 And they're using the agreed MCP protocol as the way to communicate and to communicate things like list 02:57.230 --> 03:00.270 tools and call tool and read resource. 03:00.270 --> 03:02.030 That's part of the protocol. 03:02.190 --> 03:03.150 That is what's happening. 03:03.150 --> 03:05.750 That is how the client talks to the server. 03:05.750 --> 03:07.550 And that's really at the root of all this. 03:07.550 --> 03:08.750 And that's why I show you this. 03:08.750 --> 03:11.870 So you see that's that's how it all fits together. 03:12.590 --> 03:19.510 And then I've got a utility method here a function get accounts tools OpenAI which is going to call 03:19.550 --> 03:20.910 list accounts tools. 03:21.340 --> 03:22.740 and what comes back. 03:22.780 --> 03:29.900 It's in a format that is the MCP format, and OpenAI for its tools, has a slightly different format, 03:29.940 --> 03:31.060 just very slightly different. 03:31.060 --> 03:37.340 And so I map between them and I return that in the OpenAI version of things. 03:37.700 --> 03:44.460 And this code here, along with this client code, is basically what OpenAI has now written and provides 03:44.460 --> 03:45.460 to us out of the box. 03:45.460 --> 03:46.780 So we don't need this anymore. 03:46.940 --> 03:51.940 But if we wanted to, we could just do it this way and then and then we've written our own client. 03:52.180 --> 03:57.500 So as I say, not not necessary, but it shows you that writing an MCP client is no big deal. 03:57.500 --> 04:04.540 You just have to to use some of Anthropic's code to spawn a process, connect through its input, output 04:04.540 --> 04:09.220 its standard and standard out, and then ask questions like what are your tools? 04:09.260 --> 04:12.940 Let me call a tool and can I read your resources? 04:13.500 --> 04:18.460 And so that's all it takes to talk to our server. 04:18.460 --> 04:21.140 And of course these parameters could be anything that you want. 04:21.180 --> 04:23.820 So so it doesn't have to be specific to accounts. 04:23.820 --> 04:27.460 This could just be a sort of generic server that takes care of this. 04:27.900 --> 04:28.700 All right. 04:28.740 --> 04:30.340 So with that, let's go and give this a try. 04:30.380 --> 04:32.220 Let's try using our client. 04:32.220 --> 04:34.980 So we're back in in our lab. 04:35.220 --> 04:40.900 And from that client module we were just looking at I'm going to import two things get accounts tools 04:40.900 --> 04:45.220 open I the thing we were just looking at and read accounts resource. 04:45.220 --> 04:49.100 Another thing that we'd seen that just calls the server for the resource. 04:49.500 --> 04:51.740 Um, and uh, let's run this. 04:51.740 --> 04:54.260 And if I print open AI tools, we'll see there. 04:54.420 --> 04:56.340 These are the open AI tools. 04:56.820 --> 05:01.260 Uh, in fact, what I can also do, I can just call the list accounts tools. 05:01.260 --> 05:03.500 We can just directly call that as well. 05:03.500 --> 05:13.780 Let's just do this to show you, um, I will say list accounts, tools, um, uh, MCP tools equals 05:13.780 --> 05:14.500 await list. 05:14.660 --> 05:15.140 There we go. 05:15.180 --> 05:15.780 That's what I want. 05:15.820 --> 05:16.140 Thank you. 05:16.180 --> 05:16.700 Cursor. 05:16.820 --> 05:19.060 And then we will print those. 05:19.300 --> 05:26.060 And so just to show you what this means, if we run this, this first response here is exactly what 05:26.060 --> 05:31.220 comes back from the MCP server through our MCP client printing it. 05:31.220 --> 05:37.740 And this second line, the the OpenAI version is just what happens when I switch it up into being in 05:37.740 --> 05:42.140 the format that OpenAI expects, but that's the only distinction between them. 05:42.140 --> 05:48.220 But either way, making these calls it it started our client, it had that session in it. 05:48.260 --> 05:50.380 It spawned a process. 05:50.420 --> 05:52.260 It made a call to that process. 05:52.260 --> 05:56.300 The standard output was collected and returned right here. 05:56.300 --> 06:00.220 And that is an MCB client and server at work. 06:00.420 --> 06:05.180 It's also perhaps worth pointing out that the way I've written this is not efficient, that every single 06:05.180 --> 06:08.460 call in here is spawning another process. 06:08.460 --> 06:11.500 It's running this again, creating another session. 06:11.540 --> 06:16.740 In practice, you could do this more efficiently with one context manager that handles all of these 06:16.740 --> 06:17.660 different steps. 06:17.660 --> 06:22.060 But right now this doesn't matter to us particularly starting this Python process is super quick. 06:22.260 --> 06:25.460 So that that that that's not of consequence. 06:25.460 --> 06:29.020 But you may have noticed that that's probably not the way you would you would want to write it. 06:29.020 --> 06:33.380 If you needed to have this scale a lot, you would just have the one context manager. 06:34.260 --> 06:34.820 Okay. 06:35.020 --> 06:42.700 So, uh, going back to the lab, then let's actually use this client with an agent.