1 00:00:00,080 --> 00:00:02,600 So here we are back in the lab. 2 00:00:02,600 --> 00:00:08,680 And what we're looking at here are parameters being set for our very own MCP server that we just wrote. 3 00:00:09,240 --> 00:00:16,640 So the command is you've run because that's exactly what we would type at a command line to run this 4 00:00:16,640 --> 00:00:17,320 module. 5 00:00:17,480 --> 00:00:20,680 You've run account server dot Pi. 6 00:00:20,920 --> 00:00:24,160 Just run the account server.py that I just showed you. 7 00:00:24,160 --> 00:00:30,040 And we know that when that's called it's going to create an MCP server, a fast MCP. 8 00:00:30,320 --> 00:00:31,840 It's going to call run. 9 00:00:31,840 --> 00:00:35,080 And that we've got some functions decorated as MCP tools. 10 00:00:35,560 --> 00:00:36,160 All right. 11 00:00:36,440 --> 00:00:42,320 So then we use the open AI agents SDK context manager with MCP Server Studio. 12 00:00:42,640 --> 00:00:44,480 And we pass in those parameters. 13 00:00:44,520 --> 00:00:46,000 Remember this timeout. 14 00:00:46,240 --> 00:00:50,280 And we then are going to call server list tools. 15 00:00:50,400 --> 00:00:52,960 So just by running this what's going to happen. 16 00:00:52,960 --> 00:00:55,040 It's going to create an MCP client. 17 00:00:55,160 --> 00:01:00,640 It's then going to spawn our MCP server by carrying out this instruction right here. 18 00:01:01,080 --> 00:01:04,600 And it's then going to ask it, what tools do you offer us? 19 00:01:04,840 --> 00:01:05,760 And we'll print them out. 20 00:01:05,760 --> 00:01:08,240 And we're hoping to see the things that we decorated. 21 00:01:08,560 --> 00:01:10,080 Uh, let's see if this works. 22 00:01:10,320 --> 00:01:11,400 Can it be that simple? 23 00:01:11,720 --> 00:01:12,520 Off it goes. 24 00:01:12,600 --> 00:01:13,800 Took 1.4 seconds. 25 00:01:13,800 --> 00:01:14,880 Let's just print that. 26 00:01:15,480 --> 00:01:16,360 And there we go. 27 00:01:16,400 --> 00:01:19,200 These are the functions that we just decorated. 28 00:01:19,200 --> 00:01:20,280 Get balance, get holdings. 29 00:01:20,280 --> 00:01:20,760 Buy shares. 30 00:01:20,760 --> 00:01:21,400 Sell shares. 31 00:01:21,440 --> 00:01:22,520 Change strategy. 32 00:01:22,680 --> 00:01:23,600 It's available. 33 00:01:23,960 --> 00:01:24,400 Wow. 34 00:01:24,920 --> 00:01:29,560 Um, so as you can see, it's not super easy, but it's also it's pretty easy. 35 00:01:30,040 --> 00:01:31,840 Okay, let's try and put this to action. 36 00:01:31,840 --> 00:01:33,520 Let's have some instructions. 37 00:01:33,520 --> 00:01:37,600 You're able to manage an account for a client and answer questions about the account. 38 00:01:37,640 --> 00:01:38,680 My name is Ed. 39 00:01:38,800 --> 00:01:40,360 My account is under the name Ed. 40 00:01:40,400 --> 00:01:42,720 What is my balance and my holdings? 41 00:01:42,880 --> 00:01:46,360 So it's going to need to make use of these tools. 42 00:01:46,640 --> 00:01:48,520 And we'll give it the latest model. 43 00:01:48,520 --> 00:01:49,200 Why not. 44 00:01:49,640 --> 00:01:51,720 Uh, let's take that. 45 00:01:51,960 --> 00:01:55,440 And now this again is the same code as before. 46 00:01:55,600 --> 00:01:58,720 We use the context manager with MCP Server Studio. 47 00:01:58,760 --> 00:01:59,960 We pass in our parameters. 48 00:01:59,960 --> 00:02:03,960 Let's put in that timeout client session timeout is 30. 49 00:02:04,600 --> 00:02:12,080 Um, and uh, we will pass in our instructions, our model, and we will then run a run and display 50 00:02:12,080 --> 00:02:12,640 the output. 51 00:02:12,640 --> 00:02:18,440 We're hoping that it's going to spawn the server, call the tools and be able to tell me about my six 52 00:02:18,480 --> 00:02:23,440 Amazon shares, uh, and the balance and everything like that. 53 00:02:23,880 --> 00:02:27,400 So enter your current cash balance is 8000. 54 00:02:27,440 --> 00:02:29,960 Your holdings includes six shares of Amazon. 55 00:02:30,080 --> 00:02:32,760 If you need any further details then let me know. 56 00:02:33,080 --> 00:02:40,800 So that is us successfully calling our very own MCP server that then calls our business logic. 57 00:02:40,800 --> 00:02:44,600 And now it's time to show you what it's like to write an MCP client. 58 00:02:44,600 --> 00:02:49,440 But I should first explain that it's not a common task to write an MCP client. 59 00:02:49,440 --> 00:02:51,360 You don't really need to do it anymore. 60 00:02:51,640 --> 00:02:57,120 When I when I first started working on this lab, the way that you worked with OpenAI agents SDK, they 61 00:02:57,120 --> 00:02:59,680 didn't sort of natively support MCP. 62 00:02:59,840 --> 00:03:04,810 You had to write your own client and then provide the tools into OpenAI agents SDK. 63 00:03:05,210 --> 00:03:11,730 And it was the day that I finished this project and checked it into GitHub and did the push that same 64 00:03:11,730 --> 00:03:18,290 day, about about 2 or 3 hours later, they released an update to OpenAI SDK that simplified everything 65 00:03:18,290 --> 00:03:20,050 and meant that all my code was useless. 66 00:03:20,730 --> 00:03:24,250 And you can, you can, you can look at the git timestamps if you don't believe me. 67 00:03:24,530 --> 00:03:25,170 It was crazy. 68 00:03:25,170 --> 00:03:27,770 I was I was infuriated, but it is a great thing. 69 00:03:27,770 --> 00:03:33,370 And basically it means that just with this, this context manager construct, you automatically get, 70 00:03:33,450 --> 00:03:36,490 uh, OpenAI SDK creating the client for you. 71 00:03:36,730 --> 00:03:40,570 But anyway, it's a good exercise to show you what it's like to make a client. 72 00:03:40,690 --> 00:03:44,450 And also this this construct works for tools. 73 00:03:44,450 --> 00:03:47,810 But if you want to use resources, I think you still need to write a client. 74 00:03:47,850 --> 00:03:48,890 And it's not so common. 75 00:03:48,890 --> 00:03:52,490 But but but we'll do it anyway and you'll see how it works. 76 00:03:52,770 --> 00:03:53,130 Okay. 77 00:03:53,170 --> 00:03:56,970 So this is in a Python module called Accounts clients. 78 00:03:57,930 --> 00:03:59,770 And this this is where the magic happens. 79 00:03:59,770 --> 00:04:01,250 That is no longer needed really. 80 00:04:01,610 --> 00:04:01,970 Uh. 81 00:04:02,010 --> 00:04:09,690 So this is this is a this is a an MCB client for use with my accounts, MCP server. 82 00:04:09,690 --> 00:04:15,810 So at the top here I specify these are the parameters that will be used to launch the MCP server. 83 00:04:15,850 --> 00:04:17,410 Now this could be something configurable. 84 00:04:17,410 --> 00:04:23,250 You could make a sort of generic MCP client that takes this as configuration and then and then spawns 85 00:04:23,250 --> 00:04:23,330 it. 86 00:04:23,330 --> 00:04:28,770 But for now this is just a fixed client for our accounts MCP server. 87 00:04:28,770 --> 00:04:31,210 And so this is an accounts MCP client. 88 00:04:31,530 --> 00:04:33,570 And so there are a few things we need to be able to do. 89 00:04:33,570 --> 00:04:35,450 We need to be able to list the tools. 90 00:04:35,490 --> 00:04:40,090 And so this first function here is an example of the function that lists tools. 91 00:04:40,090 --> 00:04:44,690 And we're basically using a bunch of anthropic code here for our client. 92 00:04:45,010 --> 00:04:48,530 And you can see that there's some context managers that we manage a session. 93 00:04:48,650 --> 00:04:50,210 We initialize the session. 94 00:04:50,210 --> 00:04:55,330 And then we call list tools on the session and return the tools. 95 00:04:55,370 --> 00:04:59,210 And so it's just sort of plumbing stuff that that you need to know about. 96 00:04:59,250 --> 00:05:03,810 If you want to write something that will contact your server and list the tools. 97 00:05:04,250 --> 00:05:06,530 This second one is actually calling a tool. 98 00:05:07,250 --> 00:05:12,330 This is how you go about reading a one of the resources, and this is reading the other resources. 99 00:05:12,330 --> 00:05:15,850 You can see I specify the resource name right there. 100 00:05:16,290 --> 00:05:19,290 Um, and then finally the types. 101 00:05:19,290 --> 00:05:27,570 The way that MCP returns tools is very similar to the JSON, the standard JSON that you use when you 102 00:05:27,610 --> 00:05:30,290 call a tool, as we did in week one. 103 00:05:30,450 --> 00:05:31,770 But it's not identical. 104 00:05:31,770 --> 00:05:34,450 There is some couple of slight differences. 105 00:05:34,450 --> 00:05:40,610 So if you're writing your own MCP client, you also have to know how to map between the MCP description 106 00:05:40,610 --> 00:05:46,370 of a tool and the kind of JSON description of a tool that's used generally when you're calling LMS. 107 00:05:46,530 --> 00:05:51,290 And that is what this this function here does that I painstakingly wrote. 108 00:05:51,290 --> 00:05:56,530 But all of this comes for free packaged in OpenAI agents SDK. 109 00:05:56,690 --> 00:06:00,770 So you don't need to do all of this, but you could look through it should you be interested and should 110 00:06:00,770 --> 00:06:05,250 you want to understand what does it mean to have your own MCP client? 111 00:06:05,290 --> 00:06:09,570 Okay, so now it remains for us to actually run the MCP client that we just created. 112 00:06:09,570 --> 00:06:11,450 So I import from that module. 113 00:06:11,450 --> 00:06:14,370 We just looked at some of the functions we just saw. 114 00:06:14,490 --> 00:06:19,650 And the first function I call is the list account tools function, which is the one that's going to 115 00:06:19,690 --> 00:06:25,010 spawn an MCP server and ask it for its tools that will come back as MCP tools. 116 00:06:25,330 --> 00:06:30,650 I'm then going to call the other function I showed you, where I take the MCP tools, and I reconstitute 117 00:06:30,650 --> 00:06:32,850 them as open AI tools. 118 00:06:32,890 --> 00:06:38,170 I turn them into these function tools, just as if I'd had a function and decorated it with app function 119 00:06:38,170 --> 00:06:38,450 tools. 120 00:06:38,450 --> 00:06:39,170 Same thing. 121 00:06:39,290 --> 00:06:40,930 And then I'm going to print that. 122 00:06:40,930 --> 00:06:45,250 So I'll print these two separately so you can see the results side by side. 123 00:06:45,650 --> 00:06:52,410 So this first line here, this is the actual MCP tools themselves with their description, the arguments 124 00:06:52,450 --> 00:06:53,130 of all the tools. 125 00:06:53,130 --> 00:06:55,690 And you can take a look at them should you wish. 126 00:06:56,130 --> 00:07:02,810 And underneath it is the same thing but reconstituted as a function tool, which is the kind of object 127 00:07:02,810 --> 00:07:05,180 that needs to be passed in to OpenAI. 128 00:07:05,740 --> 00:07:08,820 And indeed, that's what I'm going to do right here. 129 00:07:09,100 --> 00:07:16,060 So this is the way you had to do it before OpenAI built this, so that you can now you can just pass 130 00:07:16,060 --> 00:07:17,020 in the MCP servers. 131 00:07:17,020 --> 00:07:19,220 You don't have to build a client and do all of this stuff. 132 00:07:19,260 --> 00:07:23,820 But before that, in fact, very recently this was the way to do it. 133 00:07:23,860 --> 00:07:27,060 You'd actually have to pass in the tools themselves. 134 00:07:27,140 --> 00:07:33,020 So we've used our MCP client to find the tools and then we're passing them in here. 135 00:07:33,060 --> 00:07:37,660 So as I say, you just follow along for interest to see what this is doing. 136 00:07:37,660 --> 00:07:39,660 But it's not like you'd need to do this yourself. 137 00:07:39,980 --> 00:07:47,140 So as always, I've got a trace, and then I'll pass in my instructions and I will pass in the tools 138 00:07:47,140 --> 00:07:50,940 that when these tools are called, it's actually going to call the MCP client. 139 00:07:50,940 --> 00:07:55,580 That's going to call the MCP server that's going to actually run our business logic. 140 00:07:55,700 --> 00:07:55,980 Okay. 141 00:07:56,020 --> 00:07:58,980 So now we're going to run our MCP client like this. 142 00:07:59,180 --> 00:08:06,340 So we're passing in the tools we are calling the model and it has now responded. 143 00:08:06,620 --> 00:08:08,100 I asked what's my balance? 144 00:08:08,100 --> 00:08:09,860 And it's responded with the balance. 145 00:08:09,860 --> 00:08:14,940 And so just as a recap, what's going on there, we provided it with some tools right here. 146 00:08:14,940 --> 00:08:21,820 Those tools are in fact wrappers around MCP tools, which actually created an MCP client which launched 147 00:08:21,820 --> 00:08:28,180 our MCP server and which ran our function through the business logic on the MCP server and got back 148 00:08:28,180 --> 00:08:29,900 the results to the LM. 149 00:08:30,020 --> 00:08:32,860 So quite a lot was going on there behind the scenes with this. 150 00:08:32,900 --> 00:08:37,820 And as I say, you really shouldn't need to do this yourself because open AI agents SDK does it all 151 00:08:37,820 --> 00:08:38,340 for you. 152 00:08:38,340 --> 00:08:41,940 But it's useful to see how the plumbing works behind the scenes. 153 00:08:42,340 --> 00:08:44,860 Now here's another example of using the MCP client. 154 00:08:44,860 --> 00:08:50,620 This time we're using this read accounts resource client function that I wrote. 155 00:08:50,620 --> 00:08:52,460 And I think that you do need to do it this way. 156 00:08:52,460 --> 00:08:57,940 If you want to use resources, then that doesn't come for free with the OpenAI agents SDK package. 157 00:08:58,060 --> 00:09:01,820 So this is an example of reading the resource called editor. 158 00:09:02,100 --> 00:09:04,140 So let's see what happens if we run that. 159 00:09:04,420 --> 00:09:06,140 Let's run it through our client, through our server. 160 00:09:06,140 --> 00:09:07,980 And that's what comes back. 161 00:09:08,180 --> 00:09:11,180 This is a description of my account. 162 00:09:11,420 --> 00:09:17,020 And just to show you, I could also, of course, just import the business logic directly. 163 00:09:17,020 --> 00:09:19,980 And this is ending up just calling this function report. 164 00:09:19,980 --> 00:09:22,500 So this should give exactly the same answer. 165 00:09:22,500 --> 00:09:23,740 You could see the same stuff. 166 00:09:23,940 --> 00:09:27,100 So essentially what we've done is we've taken this piece of business logic. 167 00:09:27,100 --> 00:09:31,820 We've wrapped it in an MCP server to be available at a certain resource. 168 00:09:31,820 --> 00:09:36,900 We've written an MCP client that allows you to expose that resource. 169 00:09:37,060 --> 00:09:39,380 And that is what this function is. 170 00:09:39,540 --> 00:09:42,940 And so you can call this MCP client to get this. 171 00:09:42,940 --> 00:09:45,420 You could also just call the business logic directly. 172 00:09:45,700 --> 00:09:47,420 Um, and so why would you want to do this. 173 00:09:47,460 --> 00:09:53,700 Again it's if you wanted to share this resource with other people, it gives everyone a simple, streamlined 174 00:09:53,740 --> 00:09:59,180 way to access your resources rather than having to understand how your business logic works. 175 00:10:00,100 --> 00:10:07,990 Okay, that wraps up creating MCP servers and MCP clients with most of the emphasis on the servers. 176 00:10:08,030 --> 00:10:12,070 And I would now like to give you an exercise to go and create your own. 177 00:10:12,270 --> 00:10:19,150 So one super simple one to do would be to write an MCP server that can tell you the current date, give 178 00:10:19,150 --> 00:10:26,110 the current date, and you can expose it as a tool that you can equip OpenAI agent with so that it can 179 00:10:26,110 --> 00:10:27,110 find out the current date. 180 00:10:27,110 --> 00:10:31,590 And it can make sure that what it's the the content and the questions it's answering, it's doing so 181 00:10:31,630 --> 00:10:33,430 mindful of what the current date is. 182 00:10:33,670 --> 00:10:35,430 Um, and uh, yeah. 183 00:10:35,470 --> 00:10:43,190 Then then as a harder exercise, you could not only build that MCP server, but also build an MCP client 184 00:10:43,190 --> 00:10:44,270 to accompany it. 185 00:10:44,470 --> 00:10:50,550 Taking a look at exactly how I just did that for the accounts client compared with the accounts server, 186 00:10:50,550 --> 00:10:51,710 same approach. 187 00:10:51,750 --> 00:10:57,670 And then you could actually write something that looks like a native call to OpenAI, as we did in week 188 00:10:57,670 --> 00:11:01,990 one, like a simple native call with a tool as a JSON. 189 00:11:01,990 --> 00:11:07,110 And then you're getting really bare metal and you're seeing how you can call an MLM and have an MCP 190 00:11:07,110 --> 00:11:10,870 client and an MCP server where you're writing all the nuts and bolts. 191 00:11:10,990 --> 00:11:14,470 The only real use of that is to give you some direct exposure to it. 192 00:11:14,470 --> 00:11:16,110 So only do that if you're if you're interested. 193 00:11:16,150 --> 00:11:17,550 You won't have to do that day to day. 194 00:11:17,670 --> 00:11:19,790 But but it's a it's an interesting exercise. 195 00:11:20,110 --> 00:11:26,150 It's also worth pointing out that actually answering the current date isn't a great tool isn't super 196 00:11:26,150 --> 00:11:31,350 useful in the real world, because if you need to tell your agent about the current date, it's better 197 00:11:31,350 --> 00:11:36,550 just to state it in the prompt so that it always has it available so that the model doesn't have to 198 00:11:36,550 --> 00:11:41,030 go through the extra work of knowing to call your tool to collect the date. 199 00:11:41,190 --> 00:11:44,710 So if you feel like it, try make some more interesting tools. 200 00:11:44,750 --> 00:11:50,350 Have a shot at making a calculator that can do some calculation operation on two inputs or something 201 00:11:50,350 --> 00:11:50,950 like that. 202 00:11:51,030 --> 00:11:51,390 So. 203 00:11:51,390 --> 00:11:57,670 So try and make some interesting tool that appeals to you, and then write an MCP server and perhaps 204 00:11:57,670 --> 00:12:01,670 also explore writing an MCP client should you wish and enjoy that. 205 00:12:01,710 --> 00:12:05,270 And we will then go over to wrap up for today.