WEBVTT 00:00.210 --> 00:01.920 Instructor: So let's go back to the code 00:01.920 --> 00:04.050 and let's go to main.py file, 00:04.050 --> 00:08.580 and we'll start by creating the Stdio client 00:08.580 --> 00:11.550 'cause that's going to be the proxy for our MCP host, 00:11.550 --> 00:13.590 which is going to be our LangGraph agent. 00:13.590 --> 00:15.420 And the client needs to know two things. 00:15.420 --> 00:17.910 It needs to know how to run the MCP Server, 00:17.910 --> 00:19.680 so which commands to execute 00:19.680 --> 00:21.330 in order for the server to work. 00:21.330 --> 00:24.690 And second, it needs to know how it'll communicate with it. 00:24.690 --> 00:27.270 And those two pieces of information, we're going 00:27.270 --> 00:31.860 to supply the Stdio client with the Stdio server parameters. 00:31.860 --> 00:34.680 So when we initialize this variable here, 00:34.680 --> 00:38.250 we give it the information of how to run the MCP Server 00:38.250 --> 00:42.750 and the object itself, the Stdio server parameters, 00:42.750 --> 00:44.610 it has the information that it's going 00:44.610 --> 00:48.090 to communicate via the transport of Stdio. 00:48.090 --> 00:50.580 So it has all of that information. 00:50.580 --> 00:52.440 Anthropic handles this for us, 00:52.440 --> 00:55.500 and this is what's convenient about MCP SDK. 00:55.500 --> 01:00.030 Wonderful so let's go and use the Stdio client 01:00.030 --> 01:03.540 context manager and let's define it 01:03.540 --> 01:06.930 and supply it with the Stdio server params. 01:06.930 --> 01:11.190 So to do that I'm going to write async with 01:11.190 --> 01:13.950 as the context manager, stdio_client, 01:13.950 --> 01:16.060 and I'm going to give it the argument 01:16.978 --> 01:18.330 of the Stdio server params, 01:18.330 --> 01:21.030 and what we're going to get back from this, we're going 01:21.030 --> 01:23.010 to get back two objects here. 01:23.010 --> 01:25.110 One object is going to be the read 01:25.110 --> 01:27.810 and the other is going to be the write transport. 01:27.810 --> 01:29.460 All right, so after we did that, 01:29.460 --> 01:32.460 we want to create the client session 01:32.460 --> 01:35.910 because every client connects to a server via session, 01:35.910 --> 01:38.490 and we're going to feed that session with the write 01:38.490 --> 01:40.110 and read we just got here. 01:40.110 --> 01:42.630 So let's go and use async context manager 01:42.630 --> 01:44.490 with client session. 01:44.490 --> 01:47.340 And here we're going to give it the read stream 01:47.340 --> 01:49.590 of read we got from before, 01:49.590 --> 01:52.650 and the write stream of write we got from before as well. 01:52.650 --> 01:55.650 So now we have a client session ready. 01:55.650 --> 01:59.160 And this MCP client session, this object is what's going 01:59.160 --> 02:01.380 to do the heavy lifting of communicating 02:01.380 --> 02:03.237 between the client and the server. 02:03.237 --> 02:05.310 And to be honest, we don't really need 02:05.310 --> 02:06.780 to understand this in depth 02:06.780 --> 02:10.890 how this communication is working as the users of this SDK. 02:10.890 --> 02:12.120 If you want, please let me know. 02:12.120 --> 02:14.400 I'd be happy to make a deep dive into it, 02:14.400 --> 02:16.620 but we should use it as a black box. 02:16.620 --> 02:19.170 And this is the convenient thing about using frameworks 02:19.170 --> 02:21.660 with abstract information for us, which is not 02:21.660 --> 02:25.320 that relevant because as the developers of the agent, 02:25.320 --> 02:27.420 we don't really care about the ins and outs 02:27.420 --> 02:30.270 about the communication between the client and the server. 02:30.270 --> 02:31.980 We just assume that it's working 02:31.980 --> 02:35.550 and that the Anthropic software engineers did a decent job 02:35.550 --> 02:38.700 implementing this and abstract away this for us. 02:38.700 --> 02:41.880 Cool, so now the client knows exactly how to run 02:41.880 --> 02:44.550 and how to communicate with the MCP Server. 02:44.550 --> 02:48.060 So let's now see the flow in action, the MCP flow. 02:48.060 --> 02:50.910 We first want to initialize this connection here. 02:50.910 --> 02:53.520 So right now we have all the details in order 02:53.520 --> 02:55.410 to make this initialization. 02:55.410 --> 02:59.730 To do that, let's go and await session.initialize. 02:59.730 --> 03:03.403 And let's also print that the session was initialized. 03:05.220 --> 03:09.240 So after the client initializes the connection, 03:09.240 --> 03:13.320 the MCP Server is going to respond with its available tools. 03:13.320 --> 03:16.320 So after we done this initialization, 03:16.320 --> 03:19.590 then the client is going to have the tools 03:19.590 --> 03:21.810 that the MCP Server exposes. 03:21.810 --> 03:23.700 And I'll be more accurate here. 03:23.700 --> 03:26.370 It's not only the tools, it's also the resources 03:26.370 --> 03:28.950 and the prompts which the server exposes. 03:28.950 --> 03:33.630 So if we were to ask the MCP client here, what are the tools 03:33.630 --> 03:36.930 that the MCP Server exposes, it'll have this information 03:36.930 --> 03:39.450 and this is exactly what we're going to do now. 03:39.450 --> 03:42.870 So let's define a variable which is called tools, 03:42.870 --> 03:46.440 and it's going to be equal to the awaiting 03:46.440 --> 03:50.220 of the code routine of sessions.list_tools. 03:50.220 --> 03:52.680 So this is going to list all the available tools 03:52.680 --> 03:56.130 that the client has and let's go and print it. 03:56.130 --> 03:58.770 So now let's go and run this program. 03:58.770 --> 04:01.110 I'll run uvrun main.py 04:01.110 --> 04:04.050 and we can see that session was initialized 04:04.050 --> 04:08.100 And we can see processing request of type ListToolsRequest. 04:08.100 --> 04:09.953 So this log that we see right over here, 04:09.953 --> 04:13.050 this is from the MCP client itself. 04:13.050 --> 04:16.410 And this is from the MCP SDK implementation. 04:16.410 --> 04:18.480 So we didn't write this log. 04:18.480 --> 04:20.340 After we printed the tools, 04:20.340 --> 04:22.740 we can see here we have a list of tools 04:22.740 --> 04:26.040 and we have here with the add tool, the name is going 04:26.040 --> 04:30.000 to be add, we can see all the metadata, all the information, 04:30.000 --> 04:32.820 the description, the input schema, which arguments, 04:32.820 --> 04:36.270 which does it receive, which output does it output, 04:36.270 --> 04:37.950 and all the information relevant 04:37.950 --> 04:41.280 to help the LLM decide whether to call this tool or not. 04:41.280 --> 04:44.130 So we have this information about the adds tool 04:44.130 --> 04:46.950 and we have it about the multiply tool. 04:46.950 --> 04:51.300 So those are the tools that are math MCP Server exposes. 04:51.300 --> 04:54.930 So at this point, we have an MCP Server which is running 04:54.930 --> 04:57.480 because the MCP client ran it. 04:57.480 --> 05:00.960 It then created the connection in order to communicate 05:00.960 --> 05:02.970 with it via Stdio. 05:02.970 --> 05:06.180 The client now knows the tools which can be invoked. 05:06.180 --> 05:08.880 So those are the tools that the MCP Server exposes. 05:08.880 --> 05:13.140 And now we want to create the host application, 05:13.140 --> 05:16.710 which is going to contain this client information. 05:16.710 --> 05:18.870 And in this case, in this example, 05:18.870 --> 05:20.190 the host application is going 05:20.190 --> 05:22.710 to be a LangGraph react agent. 05:22.710 --> 05:25.443 So the user is going to communicate with that. 05:26.580 --> 05:29.910 All righty, so let's go and create our agent. 05:29.910 --> 05:32.910 And we're going to do this by using the LangGraph 05:32.910 --> 05:36.180 pre-built create react agent function. 05:36.180 --> 05:40.110 And this function returns us a LangGraph react agent. 05:40.110 --> 05:44.280 Now I am assuming that you know what's a react agent? 05:44.280 --> 05:46.530 What's LangGraph? Cool. 05:46.530 --> 05:50.850 So the react agent is going to receive the LLM, 05:50.850 --> 05:54.300 which is going to be the default LLM from Chat OpenAI 05:54.300 --> 05:57.060 I think it's GPT-4o mini. 05:57.060 --> 06:00.180 And it's going to receive the tools that we want 06:00.180 --> 06:02.010 to supply our agent with. 06:02.010 --> 06:06.690 And notice in line 27, I gave it the argument of tools, 06:06.690 --> 06:09.720 which is an MCP tool object. 06:09.720 --> 06:12.270 It's not a LangChain tool object. 06:12.270 --> 06:16.230 So by running this right now, we are going to get an error, 06:16.230 --> 06:17.340 and I want to show you this. 06:17.340 --> 06:21.000 So to show you the value of the LangChain MCP adapter 06:21.000 --> 06:22.620 and what it should provide us. 06:22.620 --> 06:25.950 So let me go and run this code over here, 06:25.950 --> 06:28.080 and let me show you that we get an error 06:28.080 --> 06:29.280 and you can see that we get here 06:29.280 --> 06:31.170 that the first argument must be a string 06:31.170 --> 06:33.750 or a callable for tool decorator. 06:33.750 --> 06:35.157 Instead it got a tuple. 06:35.157 --> 06:38.280 And the TLDR for this is the mismatch. 06:38.280 --> 06:41.250 The create react agent of LangChain expect 06:41.250 --> 06:45.300 to get in the arguments, a list of LangChain tools 06:45.300 --> 06:49.890 and instead we gave it a list of or tuple of MCP tools. 06:49.890 --> 06:51.390 So they both are going 06:51.390 --> 06:54.630 to contain the same information about which tools to invoke 06:54.630 --> 06:56.580 and how to invoke them, which arguments, 06:56.580 --> 06:58.230 what's the return values? 06:58.230 --> 07:00.870 But the format is going to be a bit different. 07:00.870 --> 07:02.490 And to mitigate over this, 07:02.490 --> 07:05.730 to translate the MCP tool object into a LangChain tool 07:05.730 --> 07:08.700 object, we're going to use the load mcptools 07:08.700 --> 07:10.410 function of LangChain. 07:10.410 --> 07:14.880 And all this function needs to receive is the MCP session, 07:14.880 --> 07:17.940 which holds the information on the tools 07:17.940 --> 07:20.520 that the MCP Server exposes. 07:20.520 --> 07:24.300 So if we'll now run it, we can see 07:24.300 --> 07:27.420 that we get here coreturn is not returnable. 07:27.420 --> 07:28.680 I forgot to await it. 07:28.680 --> 07:30.213 So let's run it again, 07:32.730 --> 07:34.770 and everything is running right now. 07:34.770 --> 07:39.120 So our agent is ready to work. So now we can even test it. 07:39.120 --> 07:43.110 So let me go now even and print the tools here. 07:43.110 --> 07:45.150 Let me go and print the tools, 07:45.150 --> 07:47.280 and we're going to see that now you're going 07:47.280 --> 07:48.900 to get the LangChain tools 07:48.900 --> 07:51.960 because the structured tool class is a class 07:51.960 --> 07:53.703 from the LangChain library. 07:54.570 --> 07:59.570 Wonderful. So let's go and test our LangGraph react agent. 07:59.910 --> 08:04.910 So here we are going to run the agent.ainvoked 08:05.130 --> 08:08.190 because we are in asynchronous context 08:08.190 --> 08:10.620 and the argument that we're going to give it is going 08:10.620 --> 08:12.480 to be a dictionary, which is going 08:12.480 --> 08:13.950 to hold the key of messages. 08:13.950 --> 08:17.190 So, and here we're going to give a list of messages, 08:17.190 --> 08:20.490 and we're going to put here the human message object. 08:20.490 --> 08:22.980 Again, this is a LangChain construct, 08:22.980 --> 08:26.130 and here the content is going to be what is two plus two. 08:26.130 --> 08:30.360 And don't forget to import it because we did not import it. 08:30.360 --> 08:34.773 So let's go and simply add this import, let's say over here.