WEBVTT 00:00.160 --> 00:04.680 Okay, so we've now covered the three ways to use MCP with Nan. 00:04.720 --> 00:10.320 The one that we're going to use most of the time that will be most useful for you, which is an MCP 00:10.320 --> 00:16.960 client node like a backup, a backup in cases where there isn't a tool readily available in Nan. 00:17.080 --> 00:23.280 And you can use the wonderful standard MCP to connect to anyone that's made an MCP server that suits 00:23.280 --> 00:23.840 your needs. 00:23.920 --> 00:25.200 That's the blue one. 00:25.680 --> 00:30.560 The yellow one is if you're trying to make your own MCP server that other people might want to reuse. 00:30.920 --> 00:36.640 And then the purple one is if you want all of your workflows to be MCP servers, so that that Claude 00:36.640 --> 00:38.640 say, could run any one of them. 00:38.680 --> 00:41.560 These are the three ways to use MCP. 00:42.200 --> 00:45.560 And now after all of this, I have I have some bad news. 00:45.760 --> 00:53.600 So the bad news is that MCP actually isn't as useful in Nan as it is in some other contexts. 00:53.920 --> 00:55.480 And I say that for a couple of reasons. 00:55.480 --> 01:02.000 The first of them is that Ann doesn't support one of those three transport mechanisms that I told you 01:02.000 --> 01:04.240 about the three different architectures. 01:04.480 --> 01:11.240 The first of them is called stdio standard input output, and that approach is not supported, at least 01:11.240 --> 01:12.000 as of now. 01:12.000 --> 01:16.880 With connecting to N810, you have to use one of the other two, and that rules out a bunch of different 01:16.880 --> 01:20.320 MCP servers that are designed for that particular way of working. 01:20.320 --> 01:21.520 So that's just one of those things. 01:21.520 --> 01:26.920 It just means that you've got a slightly smaller pool of MCP servers than than the rest of the world. 01:27.080 --> 01:27.880 That's one thing. 01:28.040 --> 01:29.160 And what's the other thing? 01:29.160 --> 01:36.560 Well, the other thing is kind of funny is to say that that actually, Nan is so good that you've already 01:36.560 --> 01:39.680 got a ton of tools to choose from. 01:39.680 --> 01:45.440 And remember, I just said to you that if there isn't already an existing tool that that covers your 01:45.440 --> 01:47.680 needs, then you should use that. 01:47.680 --> 01:52.800 That means that that the Nan team or the community have worked to package something into tools. 01:52.800 --> 01:54.920 For Nan, that's the first choice. 01:55.040 --> 01:58.720 Well, that first choice is available a lot of the time. 01:58.760 --> 02:04.160 It's actually quite hard to find something that doesn't have an Nan node. 02:04.560 --> 02:07.080 And so when there is an n node, you should use that. 02:07.080 --> 02:14.000 So the times when you need to fall back to like an MCP is actually quite rare because there's so many 02:14.000 --> 02:15.840 tools already in Nan. 02:16.000 --> 02:21.880 It's almost like saying Nan is so good that there's less need for this kind of generic MCP client MCP 02:21.920 --> 02:25.240 server setup, so that's worth worth recognizing. 02:25.240 --> 02:30.480 So when it came to, to me building some interesting projects with MCP, I struggled to start with because 02:30.480 --> 02:33.880 every time I thought of a good idea, there was already an end node for that. 02:34.080 --> 02:35.280 And that kept happening to me. 02:35.320 --> 02:37.240 It's like, ah, and it ends too good. 02:37.280 --> 02:39.120 So you may you may find you have that problem. 02:39.160 --> 02:42.200 You're determined to use MCP because it's all the rage. 02:42.200 --> 02:44.000 And and everyone wants to use MCP. 02:44.080 --> 02:46.680 But it turns out Nan already got you covered. 02:46.720 --> 02:48.520 You've already got so many tools. 02:48.520 --> 02:50.680 It's not like there's such a need for it. 02:50.680 --> 02:52.920 But still, I found a couple of needs. 02:52.920 --> 02:56.600 And that's what we're gonna do in today's practical. 02:56.600 --> 03:02.480 But first, there's one more thing I want to mention that's got nothing to do with MCP, but it's something 03:02.520 --> 03:03.600 important to mention. 03:03.600 --> 03:08.520 It's something that we touched on during the lab yesterday, which I sort of introduced in a bit of 03:08.520 --> 03:12.440 a rush, and I it deserves it warrants a little bit more time. 03:12.960 --> 03:19.240 I told you that tools are the major innovation that stands behind a genetic AI. 03:19.600 --> 03:26.560 Well, in addition to tools, there is a kind of secret weapon the the the the dark secret. 03:26.600 --> 03:34.400 That's not so much of a of a secret, uh, of agents and of using llms generally is this fabulous thing 03:34.400 --> 03:38.320 called structured outputs that we did in the lab yesterday. 03:38.560 --> 03:40.520 And I just want to reinforce it. 03:40.520 --> 03:47.600 Structured outputs is when you have an AI agent or really any NLM call and you say, you know what, 03:47.600 --> 03:50.320 I don't want you just to generate any tokens. 03:50.560 --> 03:57.600 I want you to output JSON, and I want you to output JSON that conforms to a particular spec. 03:57.960 --> 04:02.400 And as you may remember, the way you do it in Nan is by setting a flag. 04:02.520 --> 04:04.410 And that flag is right there. 04:04.410 --> 04:10.490 When you're editing the agent, there is like a switch, require specific output format, and you turn 04:10.490 --> 04:18.730 that switch on, and then this new little, uh, sub node appears in your node cluster AI agent that 04:18.730 --> 04:20.770 says output parser. 04:21.050 --> 04:26.970 And you can attach that to like a structured output parser as, as I've done in this diagram. 04:27.170 --> 04:34.330 And then within that you get to specify the kind of JSON that you want it to produce, and that that 04:34.330 --> 04:35.690 is the magic. 04:35.730 --> 04:44.090 That is where you say, I want to constrain the output of this AI agent to conform to this JSON, and 04:44.130 --> 04:51.690 it's super powerful, and it allows you to build a more bulletproof approach around your your AI. 04:51.810 --> 04:53.970 And I use it all the time and so should you. 04:53.970 --> 04:59.210 And as a pro point, it's worth mentioning that for many llms, the way this works is that it's actually 04:59.250 --> 05:01.290 implemented using tools itself. 05:01.290 --> 05:06.330 It is a tool to produce structured outputs, and you can see that behind the scenes when you when you 05:06.330 --> 05:10.010 look at the, uh, the actual information exchanged with the AI. 05:10.810 --> 05:11.450 Okay. 05:11.730 --> 05:18.650 So with all of that, we have the bits and pieces now to go and do today's lab to go and actually put 05:18.650 --> 05:21.290 this MCP thing into practice. 05:21.330 --> 05:24.450 We're going to use someone else's tools. 05:24.450 --> 05:29.330 We're going to use an MCP client to connect to an MCP server and use their tools. 05:29.530 --> 05:36.010 And we're also going to take a complete workflow and make it an MCP server so that someone else can 05:36.010 --> 05:36.890 call into it. 05:36.890 --> 05:41.970 We're going to do both of those things, recognizing that the first one is the more relevant one for 05:42.010 --> 05:43.570 for for you and me day to day. 05:43.730 --> 05:44.250 Okay. 05:44.450 --> 05:45.370 Let's go and do this. 05:45.370 --> 05:45.850 Okay. 05:45.890 --> 05:47.530 Here we are back in. 05:47.690 --> 05:50.330 And and we're going to start with some housekeeping. 05:50.490 --> 05:51.890 Let's go to projects. 05:51.890 --> 05:53.570 My project right here. 05:53.730 --> 05:57.290 And from here let's go over to Project Settings on the right. 05:57.290 --> 05:58.930 And let's change the name of this. 05:58.930 --> 06:07.330 Let's make the name of this amplify, which is going to be the name of our project Save Project amplify. 06:07.370 --> 06:08.330 Save successfully. 06:08.330 --> 06:09.050 That's satisfying. 06:09.050 --> 06:11.090 And over on the left here, it becomes amplify. 06:11.330 --> 06:13.650 And I'm going to change this this folder name. 06:13.690 --> 06:17.370 Let's let's rename this folder from my first folder. 06:17.370 --> 06:20.890 Let's let's take this make this uh testing. 06:21.570 --> 06:23.250 Oh let's make it prototyping. 06:23.290 --> 06:24.010 That feels better. 06:24.010 --> 06:27.170 Prototyping rename. 06:27.330 --> 06:30.850 And then we're going to make a new folder in here. 06:30.970 --> 06:34.770 And we're going to have the new folder be uh to to. 06:34.930 --> 06:37.490 So right click here create folder. 06:37.490 --> 06:40.170 It's going to be um MCP. 06:40.210 --> 06:41.450 That's what we're going to call it. 06:41.490 --> 06:45.210 It's where we're going to build out our MCP workflows right now okay. 06:45.250 --> 06:47.810 So here we go I'm going into MCP our new folder. 06:47.810 --> 06:49.250 We're going to create a workflow. 06:49.250 --> 06:50.170 Nothing there yet. 06:50.170 --> 06:51.130 Let's change it. 06:51.410 --> 06:52.210 Uh okay. 06:52.250 --> 06:56.130 First step we are going to do the usual on chat message. 06:56.250 --> 06:58.490 Click escape to come back here again. 06:58.530 --> 06:59.730 Let's hide the chat. 06:59.850 --> 07:02.450 Nice and clean screen okay. 07:02.650 --> 07:04.730 Next up I wonder what noble add here. 07:04.730 --> 07:06.890 Let's make it an AI agent. 07:07.210 --> 07:08.610 That's what it's all about. 07:08.850 --> 07:12.210 And now here is the usual AI agent screen, and we're normally pressing escape right away. 07:12.210 --> 07:17.490 But no, I want to add one option that becomes increasingly important as we get a little bit more in 07:17.530 --> 07:18.090 depth. 07:18.090 --> 07:21.690 I'm going to click here and Max iterations. 07:21.690 --> 07:26.850 This is how many total parts of the loop, how many times it will run the loop before it says, okay, 07:26.890 --> 07:27.450 that's enough. 07:27.450 --> 07:31.250 I've done too many iterations as a sort of way to avoid it going on forever. 07:31.290 --> 07:37.170 We're going to bump that up from 10 to 30, because ten just isn't enough when it gets to hardcore stuff 07:37.170 --> 07:40.210 like we're going to be doing today and, uh, there we go. 07:40.250 --> 07:47.690 I'm also going to turn on require specific output formats because we're all about structured outputs. 07:47.690 --> 07:51.010 So there we have our AI agent. 07:51.010 --> 07:55.810 And just to keep it simple, we will have a chat model that I'm going to use open AI. 07:56.210 --> 07:59.050 You can of course use whichever one you're now comfortable with. 07:59.330 --> 08:03.730 Um but I'm going to make it a strong one this time because I want to mess around with doing some hefty 08:03.770 --> 08:08.850 stuff, but you for sure can try with a smaller one if you'd prefer, if you'd prefer the, uh, the 08:08.850 --> 08:12.650 price point or whatever, I'm going to go with GPT 5.2, the very latest one. 08:12.930 --> 08:19.570 Uh, but I am I'm going to I'm going to say the reasoning effort, which is how long, how much time 08:19.610 --> 08:23.490 does it spend to think things through before coming to the the outcome? 08:23.490 --> 08:28.250 I'm going to say low so that we're we're a little bit faster and more economical. 08:28.250 --> 08:32.450 We don't need it to be on on slow, uh, long thinking times. 08:32.450 --> 08:38.970 So this is a nice combo, most powerful model on the planet for me and with a low reasoning effort. 08:38.970 --> 08:40.570 I like this setup. 08:40.770 --> 08:43.010 Okay, so here is our agent. 08:43.010 --> 08:49.770 Let's just have some simple memory for now and we are almost ready to get started. 08:49.770 --> 08:56.010 Okay, just before we start with our MCP, I want to set this structured outputs this output parser, 08:56.050 --> 09:00.010 and I should explain that what we're going to be building here is kind of taking things we've done in 09:00.010 --> 09:01.330 the past a step further. 09:01.370 --> 09:08.860 I want to build an engine that's able to find sales prospects from the internet, so it's able to find 09:08.900 --> 09:14.220 people that we could sell our product to by being like an intelligent salesperson. 09:14.540 --> 09:14.820 Okay. 09:14.860 --> 09:21.700 Which means that what I want it to output is ultimately going to be leads leads for me or sales prospects. 09:21.740 --> 09:25.780 So let's see how can we how can we how can we do that with structured outputs. 09:25.820 --> 09:30.420 Well, we start by clicking the plus here and selecting Structured Output parser. 09:30.460 --> 09:35.500 And you'll remember that what goes here is the JSON that we want our model to produce. 09:35.580 --> 09:36.860 So what should that look like. 09:36.900 --> 09:39.140 Okay let's delete whatever's there already. 09:39.140 --> 09:41.780 And let's have our best shot at this. 09:41.940 --> 09:46.740 So what we want is something that's going to respond with sales prospects. 09:46.820 --> 09:48.500 So we want a JSON object. 09:48.500 --> 09:50.940 So open and close curly braces. 09:50.940 --> 09:56.100 It's going to have a field called prospects prospects like so. 09:56.220 --> 09:59.340 And this is going to be a list of prospects. 09:59.340 --> 10:04.260 So a list you remember in JSON is surrounded in square brackets like that. 10:04.260 --> 10:07.660 And each each element of this list is going to be another object. 10:07.660 --> 10:09.980 So we need more curly braces like this. 10:10.140 --> 10:11.900 And that's going to have some fields. 10:11.940 --> 10:12.460 Okay. 10:12.500 --> 10:15.300 This is each of the prospects in the list. 10:15.300 --> 10:17.140 So let's let's come up with each of them. 10:17.380 --> 10:19.700 Each field it's going to have a first name perhaps. 10:22.500 --> 10:25.140 The uh prospect. 10:27.340 --> 10:29.620 First name here. 10:32.340 --> 10:33.740 Last name perhaps. 10:34.020 --> 10:35.020 Sounds sensible. 10:36.340 --> 10:36.740 Uh. 10:36.980 --> 10:38.420 Last name? 10:40.020 --> 10:42.140 Uh, then we could have their company. 10:44.340 --> 10:50.900 Where they work, we could have their role, what they do. 10:51.060 --> 10:57.860 And I'm just putting in here in the values, just English language that describes to the LM what we 10:57.860 --> 10:59.300 really want in there. 10:59.340 --> 11:01.460 There's no hard and fast rules to this. 11:01.500 --> 11:06.420 It's just something that will be read in by the LM as it figures out what do we want it to generate. 11:06.580 --> 11:07.940 So that's the role. 11:07.940 --> 11:08.860 What else do we want? 11:08.980 --> 11:12.100 How about, uh, email? 11:13.460 --> 11:22.460 Predicted email address for this prospect. 11:23.460 --> 11:25.860 And then finally, rationale. 11:28.780 --> 11:41.300 Why this prospect is, uh, qualified, uh, uh, as a potential client. 11:42.220 --> 11:43.260 Oh, okay. 11:43.940 --> 11:50.580 Uh, so those are sensible fields that we might want our LLM to generate. 11:50.580 --> 11:56.380 And this is why Structured Outputs is such a powerful technique, because this gives us an opportunity 11:56.380 --> 12:01.980 to constrain what the LLM generates and to make sure that it generates something consistent with our 12:01.980 --> 12:07.300 business objective generating prospects that conform to this schema. 12:07.660 --> 12:09.220 That's that's why it's so powerful. 12:09.220 --> 12:10.900 And that's why I love to use it.