Learn MCP Servers with Python (EASY) transcript

good morning everyone how's it going today welcome back to the channel in today's video we're doing an mCP servers crash course okay so the idea is that by the end of this video you will understand everything that you need to know about how to use mCP servers and how to create your own okay so this is what we're going to be doing first in less than 10 minutes we're going to be covering what is mCP so that you understand actually the theory behind it and the uh agentic workflow that takes place in it after that we're going to be creating our own server using Python and the server that we're going to be creating is one that will allow our AI assistance to check the latest documentation of the library that we are using uh before actually suggesting code to us okay so that the code that our AI assistant suggests to us is always up to date um and then we're going to be seeing how to use that server that we just created in both Cloud desktop and in Cloud code okay we're going to walk you through everything uh on how to set this up because sometimes it's a little bit tricky and uh then we're also going to show you how to debug these mCP servers cuz yeah that's also a little bit tricky if you want to debug them it's not just like running the function um so there you go that's what we're going to be doing today uh as usual feel free to uh drop a like And subscribe if you like these kind of videos and uh let's get right into it [Music] all right so let's start this off by going over what an agent is if you are already familiar with how an agent Works uh feel free to skip to the part of the video where I talk about how this affects the mCP servers but if you're not it really just takes a couple of minutes don't worry uh we're going to go over it really quick and uh it's actually important to understand how an agent works because pretty much all of the llm applications that you currently use are actually agents so chpt CLA clein cursor all of those use agents uh they're not just calling the llm ones and um to understand what an agent is itself and it's very straightforward an agent is essentially a loop of calls to a language model that simulate a thought process okay a Chain of Thought So in this way um when you send a query to your agentic framework it enters this Loop that you see right here it starts with a thought then it already knows because of its system prompt the available actions that it can take this can be functions usually uh just API calls or anything and then the observation is the result of that of executing that action okay so this kind of tries to emulate when you articulate the thoughts in your head before acting so let's say that for example you're making in a cake and at some point you need to mail but you don't have it with you so you're like okay so I don't have it here oh it has to be in the fridge I'm going to go to the fridge and then you go to the fridge so that articulation of the thought uh LED you to actually perform the action and I know that's that's not always how it works but um in an agent it's very useful to use a language model to simulate that thought and then uh it uh then choose the action from the tools that it has available that we have specified in the system and then once the application executes the action the observation is when it uh when you send the output of that execution to the llm and then it can start thinking again so let's say that for example uh let's say that you give your agent a tool that allows it to uh read your email so if you ask your um agent hey give me a summary of my email um it will be like okay so I have to check the summary the email so I have to check the emails um the latest emails in the Inbox and there's a tool that reads the latest emails in the inbox then the observation is going to be the return value of that which is going to be the actual uh content of the emails and now we can go back to the thinking part and the agent's going to be like okay so now I have the emails now I'm going to summarize them and now it can output the summary uh to the user okay that's essentially all that's going on and this is just another way of um of showing this so you have the user query that enters this thought uh loop then if it has to use a tool it uses the tool then it Returns the observation if it has to use the tool again it uses the tool and Returns the observation and then if by the end it already has the response it gives you the response okay that's essentially how an agent works and uh that is why it is useful uh mCP servers because the idea behind an mCP server is that here you have a toolkit but uh imagine that this toolkit you could have multiple toolkits and they all have the same input output protocol so now everyone can can create their own toolkits and you can just switch them over and be like okay so now I I want to use um now I want to use this other toolkit or now I want to use the two of them and your agent is going to be able to choose which one to to use okay um so that's essentially how an agent works and how it relates to the mCP servers um now let's actually take a look at how the mCP servers work okay so now that we have understood how an agent works now it's actually way easier to understand how an mCP server works because essentially every single server is going to be treated as a toolbox in this case so here we have a documentation toolbox that is going to allow us to search I mean in this example we have a documentation toolbox that is going to allow us to search for the document ation of a given library for example here we have another toolbox that is going to allow us to read email and send emails and this one itself is going to be another mCP server and this one right here is going to be another server that allows you to search the web okay and the idea right here is that this AI assistant can be pretty much any assistant that you want so this one can be chbt clot uh or any assistant that you wrote If you make it compatible with the mCP protocol and that's actually the kind of revolutionary part right here is that all of this is possible because we have this mCP protocol right here so happy protocol and um the idea of this is that you can use this mCP protocol to connect to all of these servers and um as long as you make your application compatible with the mCP protocol it has access to the servers that people are going to be developing themselves uh even if they are not at all related to your application okay so that's kind of the um new thing now let me actually show you the actual diagram that is the one that is shown by theanthropic and uh this is what it what we actually finally get to so we have the mCP client which is the one that you have to create for your application so if you have your chat application your chat put your assistant whatever you can use uh the mCP um SDK to make it compatible with the mCP protocol and now it is going to be able to connect to all the servers mCP servers that are essentially these toolboxes for specialized uh tools and specialized features that you can add to your server so um I mean and to be clear your servers can expose I mean can give your assistant access to of course tools like we saw before but they can also give it access to uh custom prompts and uh to resources themselves like for example uh markdown files Etc that that can be useful for documentation for example okay um so now that we have understood how an mCP server Works let's actually create our own all right so what we're going to be building right here is essentially uh this uh server that you see right here uh we're not going to be adding this resource like this we're just going to be creating this tool and we're going to be adding it to an mCP server that we're going to start and that we're going to uh use with our Cloud desktop application and also with code uh with clo code okay and the idea right here is that usually when you use an llm to help you code um you have the common problem that it does not know about the latest releases of the libraries that you're using so for example if you're using Lang chain or L index maybe there were some updates to the library a couple weeks ago or a couple months ago and that naturally is not within the pre-trained material but that uh trained the language model so they will not be aware of that so an alternative to this is that you create this mCP server that allows your coding assistant to go look for the latest documentation before answering question or before generating a code suggestion for the given um for the given library that you're using okay and that's what we're going to be building right now so let's actually start doing that okay so what we're going to be building first is let's just uh start by setting up our environment uh all the commands that I'm going to be using right here are actually in the blog post and the description so feel free to check that out and um so the first thing is that actually the guys at anthropic they recommend using UV as a package manager and I've actually been using it um for a few days now it's it's amazing I really like it um way it's kind of like pip but faster so feel free to use one if you want and um so the first thing that we're going to do is we're going to initialize our project and in this case I'm going to call my project documentation okay so I go UV init documentation and after that I am essentially just going to go into my documentation and open it with cursor okay so let's do that okay so as you can see this has created um our project right here with a very quick get ignore path and version we have our main entry point file and uh basically just the starting point of the entire thing okay um now let's actually continue with the setup just going to open my terminal right here and uh let's do a couple of things first first thing we're going to have to create our virtual environment and uh there you go now it's created right here so what I'm going to do is just um activate it and there you go now I have my virtual environment working within here and uh next thing is you're going to have to install these dependencies so I'm going to be using I can just give it more space right here I'm going to be using the mCP CLI and the httpx uh library to make the calls to Google which is what we're going to be using to search for the documentation okay so I execute this there you go and uh now I have pretty much everything thing I need now I can actually start building my mCP server all right so let's start actually building our server okay so the first thing that we're going to do is we're going to import fast mCP from mCP server fast mCP and this is the one that we're actually going to be using to initialize our server okay and there we go so then you just do mCP equals fast mCP and we name it in this case I'm just going to name it docs like this and and uh there we go uh we're also going to be needing some API keys so I'm going to be just importing uh from from tnv let me just copy it from here from tnv I'm going to import load. tnv and then I'm just going to run it before anything else happens right here okay and there we go then we're just going to initialize a few constants the first one um actually the only one that we're going to be using is this one the US user agent and um let's actually also initialize the uh serer URL like this okay and this one we're going to paste it a little bit later when we see how serer works and just to give you an idea of what the whole thing that we're going to be building we're going to be building uh three methods right here three functions the first one is going to be search web so search web this one is essentially going to allow us to this one's going to to return uh the web results from a Google search um then we're going to let me just like this uh we're also going to be creating another one called Fetch URL and this one is going to get us the results from the from just uh to scrape uh URL and to get the actual contents off it and then we're actually going to get the get docs which is actually going to be the tool that we're going to be building okay uh and this one is going to be decorated with mCP do tool like this and this is the one that the agent is going to be able to call and this one is going to first search the web to find the relevant documentation Pages for our library and then we're going to fetch those pages to actually get the contents and actually in order to do that we're going to have to use a quick little uh hack of YouTube that allows us to check only search results from a given URL and we're going to give it this URL for python this URL for L index and this URL for open AI okay you're going to see how that works in just a moment but let's actually first take a look at how serper Works to show you how to uh create this uh search web um function right here all right and just a quick uh Showcase of what we're going to be using to build this we're going to be searching the documentation using Google and in order to do this we're going to be using serer it's essentially just Google within an API and it gives you 2500 free queries when you start your first account we can see that I've been testing it a little bit and uh essentially let me just show you how it works so you just send the query and it sends the query within the payload right here under the Q parameter and then you send that to the your this URL right here and um then you just send it all to this endpoint and it returns to you the results so if we search for Apple going to get actually just uh say uh three results per per query there you go oh I can only get 10 here in the playground all right so there you go I send this and then I get from this organic results this are the ones that I'm actually interested in I get the actual results from my Google search Okay and actually what I'm going to be doing is I'm going to be appending it or prepending it with site and then the URL of where I want to search for the documentation so here if I want to say for example I want to look for Chroma DB right here now all of my results are going to be from this URL ling.com dooc so now I am sure that I'm getting actually the latest documentation from um langing and the same goes for L index say for example if you wanted to search for this thing on L index only I can prepend this with this and now all the results actually come from Doc L index stable so I'm sure that this documentation is going to be up toate and uh the idea right here is that I'm going to have my agent first uh perform the search and then go through the link that's listed right here and actually get the results using beautiful soup okay uh so let's actually first get our API keys I'm going to copy it and then I am going to paste it right here into mymv file that I'm just going to create and I'm going to say serer API key I'm just going to paste it right here and uh now I'm going to actually be able to use it with my load. tnv file right here and uh now we're ready to actually start coding actually one thing I forgot is to copy this URL right here now we can actually start uh creating this search uh web function all right so now let's actually start creating these two helper functions that we're going to be using okay so as a reminder our agent is going to first search the web for the query that it requests and then use those search results to actually access the URL of the search results and get the contents okay so this one gets the contents of the URL and this one gets the results in the form of a URL okay uh so let's first do the search web function it's going to take a query that is going to be a string like this the string and uh it's going to return a dictionary or none okay uh let me just copy and paste this one right here because this is not um um exactly uh this I mean this is not a video about how to create a search um web function so there we go um as you can see we're using async in here so I'm going to do async to and I'm gonna have to also import J httpx Json and Os for this to work so just to go over this very quickly I initialized the payload and I send the query so this search web function is essentially just going to uh uh search the web for any query that I sent to it and it is going to return a number of two uh it's going to return two results only okay so I'm just going to be focusing on the top two results um the headers the first header that it takes is my API key and then we have the content type right here as well and uh secondly we're going to initialize this Asing client and we're going to call our server API using this and actually the URL is going to be my server URL right here and there we go my headers we send my data and then I just set a time out of 30 seconds as well and uh there we go then we just response race for status and then return the Json object as a dictionary okay that's all that we're doing with this search web function so essentially just returns our search results in a dictionary and I'm going to do the same thing with the fetch URL it's essentially just the same thing but uh instead of using serper it just sends a get request to my URL and gives me the results and then I just parse them using beautiful soup okay again I am using async here so I'm just going to async it like this and I'm going to have to import beautiful soup too so there we go I essentially just send the SK request and then I parse it with beautiful soup and now I just get the text and return only the text of the uh URL okay so what my agent is going to do when they when it calls this get docs function it is going to first search the web get two results in the form of the title and the and the URL of the search result and uh then we're going to open the UR L using the fetch URL and it's going to return by the end only the results of the two top results on Google that come from a given uh Library uh documentation okay so there we go that's our helper functions we can now start actually coding the get documentation um tool all right all right so now it's time to actually start creating the tool um now very importantly you have to create well to create your tool that your agent is going to be able to use you have to decorate it with this decorator right here which is mCP do tool okay and this essential is going to convert the function definition that you're doing right now into an actual tool that is compatible with the mCP protocol okay and uh importantly I'm just going to add here uh my uh parameters as we mentioned before we're going to be taking the URL and the query um so uh something extremely important that you have to do when you initialize a tool for any agent but I mean of course right now for mCP is that you have to add a doc string specifying very clearly what your tool does this is the description that is going to appear to the user when they see the mCP server that is connected and also this is going to be the description that is available for your llm your language model to see what this tool actually does and to select it when it is when it is um useful because if your um doc string is not specific enough your agent or your application your language model is not going to be able to use it so here as you can see I specified that this one is the sear allows you to search for the documentation of a given of um for a given query in library and for now it supports Lang chain open Ai and llama index as I mentioned up here okay now the arguments also specify them very importantly because these are going to be parsed as Arguments for the function goal of your language model and then the actual uh returned value and here it's not actually going to return a list of dictionary it's going to return um the text uh from the documentation okay and let's actually start by doing this so I am going to just right now uh the first thing that I'm going to do is I'm going to test if the library that was passed right here is actually within the supported libraries that I have right here okay if it is not in uh wait if it is not in that Library I'm just going to raise an error and uh then I'm going to build the query that we're going to be sending okay remember as I mentioned before we're going to be using site uh colum and then the actual URL where we're going to be um searching and then the actual query from the language model that wants to for whatever documentation it wants to find and then I am just going to get the results right here from the search web um helper tool helper function that I had right here now I'm using a weight right here so I'm just going to have to turn this into an ASN function and uh there you go let's see what this does now if there are no results I'm just going to return results not found and then for every single one of those results I'm going to tap um into the result URL link if you remember correctly that's what the results look like if I show you again right here if I go to the results I go to the organic part and then I go to the link this is what we are going to try to fetch actually here um yeah We're looping through the organic pot and then for each one of those we're tapping into the link one and we're fetching the URL which remember with beautiful soup it returns just the text of that page and then uh instead of printing this we are going to be adding them to a dictionary so the text is going to be like this and then we just return the text like that okay so there you go now we have the tool that's working correctly and we have that it is already returning um all the text from the documentation which is going to be added it's going to be returned by the tool and it's going to be added to our um language model function call okay uh so there we go now last thing to do is to come right here to if name equals Main and we're going to do mCP do run like this mCP do run and transport equals St standard input output and uh then we save this and now we have actually complete finished our mCP server it was that easy I mean you can of course have your coding assistant write it for you but I just wanted you to understand how this is all working so essentially the tool is the one that is going to be called and um you're Avail you're able to have as many helper functions as you want um and as many tools as you want as well all right so let's actually um run this and connect it to our Cloud desktop application okay all right so quick little thing that I wanted to mention before actually showing you how this works in cloud is that I actually do not have a Plus subscription to clot because I do everything with the API so uh this tool is probably not going to work with the free version because it Returns the entire contents of two different web pages and that's a lot for the that's a lot of context for the free tier of the clot desktop app so if you have a Plus subscription to Cloud this should work perfectly uh if not then uh you probably want to uh Plus subscription if you want to use this with with Cloud desktop um that being said for the demo I am actually going to be mocking the response with a shorter response uh just so just to make sure that it fits within the context window of uh Cloud desktop okay and uh after that I'm going to show this to you on cloud code and for that I'm actually going to be using the real tool uh just to show you that uh because that one works with the API okay so let's actually uh show this on CL desktop and then on cloud code all right so what I'm going to show you right now is how to use uh this mCP server that we just created with CLA okay and we're going to be using Cloud desktop for this in order for actually be able to handle this and uh the first thing you're going to want to do is you're going to go to settings developer edit configuration and it's going to show you this Json file now you're going to edit it with it any text editor you want going to be using Sublime here and under mCP servers you're going to paste the configuration for your server okay now in my case I created my server in Python so I'm going to be using UV to run it uh there is a common error uh sometimes when you do not add the entire binary so if you just do uh if you set S command UV uh this may raise an error depending on your path but um there you go I mean we can take a look at this in a moment uh so you're going to add the command to actually run your server which is UV then you're going to move your um terminal to the directory where you'll build your server in my case this is let me show you this is this one right here puid is print working directory by the way so you copy this come right here and you right here going to paste the full directory of where you're server is located then you're going to add run and main.py which is the name of our file where our server is okay this is essentially telling Claud to run the server at start okay so I'm going to save this I'm going to go back to Cloud I'm going to open it again restarted as you can see we have these errors and uh most of the time the reason for this is that your your cloud is not finding the binary for UV so what I want to do is I want to find which UV I am using here where I was actually using it and it was working so I'm just going to copy the actual full path to the binary going to back go back to my um Cloud config file and instead of just saying UV I'm going to add the entire binary path right here so now I can close this and I can restart cloud and uh there you go so as you can see now I have this little Hammer right here which uh if I click it I see the list of the tools that my uh Cloud desktop has available for me right now so in order to uh use these tools I can only I can just go right here and ask Claud for a question that will likely use this tool so for example how do I um Implement a chroma DB in blank chain and as you can see it's going to try to run this tool it going to ask me for my permission to actually run the tool it's calling library and the query then it got the answer strives to run it a few times and um there you go uh and uh there you go so now it is actually answering the question and it is actually giving me the implementation to do it using the latest uh documentation from from Lang chain and chroma so there you go now this works correctly let me show you how this looks so chroma DB implementation and now you have the result is this uh the answer right here on how on the actual documentation so there you go and that's how you do this with CLA desktop now let's actually take a quick look at how to do this with um with Cloud code and um and then let's check on debugging all right so now I'm back here at my ID and to actually run this uh mCP server with code CL uh Cloud code okay so what we're going to be doing first is I'm going to be removing this mock um uh tool result to actually use the real tool and uh what I'm going to do is add it to Cloud code first okay so I'm going to do CLA uh code to show you that I currently don't have any tools or any mCP servers running I'm going to close this and in order to add it what we're going to be doing first is we're going to do Cloud mCP at okay and this is this is going to open an interactive uh helper to actually add the mCP server or if you know the parameters you can just add them right here okay I'm going to open the interactive tool and it asks me to create a server name going to call it documentation docs because it's going to look for the latest documentation um the project is going to be the local one and enter I command to start my mCP server okay so in my case I wrote this uh mCP server in Python so I'm going to be using UV to run it however remember that sometimes if you just do UV it may u uh do some errors so I'm going to check where my binary for my UB UV is so I'm going to paste the entire binary and then for the parameters remember just like with uh uh CLA desktop we're going to do directory and then we're going to paste the working directory where my um where my uh server is so I'm going to come right here I'm going to paste this then we're going to run the actual uh mained py file that we have right here okay and the main the P file is the one naturally where my server is coded okay uh so I'm going to hit enter for the environment variables I am using some environment variables but I don't have to add them here because I have already loaded them in mymv file I'm going to leave this empty I just check that everything is correct directory uh Etc documentation I can confirm this and just hit enter to finish and if you want to check uh the mCP servers running on that you have on cloud you can do m mCP uh list clot mCP list and you will see that your new server is here just added and now if I run Cloud you can see that I have my docs uh right here okay so everything seems to be working correctly now I'm actually going to test it so let's uh ask it to create a new function that implements a chroma DB database um yeah V database with lank chain and um make sure to use the latest uh integration let's run this and hopefully it's going to use my tool so let's see if that's if it actually does it it actually used the tool um and this is the results from my tool I had actually already tested this before and it asked me to validate this and I validated that I wanted to do it so if this is the first time that you're running this tool um it is going to ask you for the validation if you actually want to run the tool so I'm going to say yes to everything to all the changes that it wants to do just to show you instead of actually TR having to validate every single thing and um as you can see it's now thinking and it used the tool that I added to it and it's using the latest now document ation from uh Lang chain to to perform this integration okay I'm going to cut the video uh right now and get back when it's finished to show you the final result okay that was actually quite fast so it says that I've created a comprehensive implementation blah blah blah and as you can see it used the tool that I asked it to use um doc uh get docs and the query was Chrome integration with Lang chain and this is what it got okay now if I close this I can come back to my main.py and see that actually there were um there's this new tool create from a vector store okay um so there you go that's how you uh use this with cl code now what we want to what I want to show you uh afterwards is how to deug this because I mean sometimes when you're building this there you can encounter errors so and that's normal so let me show you how to deug this all right so uh in order to show you how to debug this I'm going to go back to the the previous state of my mCP server which is the one that I developed right here and save this and actually let me just close uh clot code so you can see that costed me uh 14 cents for just one query so not the most affordable AI assistant but uh there you go so now what I want to do is I want to open npx uh um you're going to run this like this right here let me show you um like this so you're going to do npx at model context protocol SL inspector and then you're going to run your server within it okay so in other words in this particular case my server is on main.py and remember that in order in order to run it you do UV run and then the file of your server and in this case it is UV run main.py so I'm going to enter this and it's going to open my inspector on Local Host 5173 so here I now open my inspector so you can see and uh this is essentially just a a tool that allows you to debug your mCP server so now you can do UV is the command the arguments are this now I can do connect I'm connected to my server uh as you can as you remember from the explanation you can add uh resources prompt Etc to your mCP servers but in this case we're focusing on the tool side so in order to create to check the tools you go to tools and list the available tools that your server has right here and as you can see here's the tool that I created so you open it and here you can actually start testing it so the query what's going to be the query uh let's say that in my case it's going to be Chrome ADP and the library that I want to check for is uh say LMA index and this naturally comes from the fact that I have support for this right here in my documentation URLs and then I can run the tool and I can check the results right here and this is what is going to be shown to my language model when it calls this uh this tool so I can see that it's working correctly just a very straightforward way to test your your mCP server tools because not a straightforward to run in a notebook or in the script itself because it's using the mCP decorators so this is a way to test them and it's a very reliable and very very neat way to actually check that your mCP server is working correctly okay uh so there you go so there you go we have successfully finished the video we explained at the beginning what is mCP exactly and how it works we created our own mCP server using python um in the future maybe we can make a video about how to do this using node um in JavaScript uh we connected it to both Cloud desktop and Cloud code and uh we also learned how to buug it real quick so that's it for this video feel free to subscribe if you want to more videos like this uh feel free to drop a like and also feel free to check out the boot camp that the AI engineering boot camp that I set up the link is also in the description to take you from Zero to Hero in the AI engineering world okay so uh drop any questions in the comments and I will see you next time [Music]