mCP has really exploded past few weeks if you want to learn what mCP actually is there are probably tall videos and tweets that you can go through already so I wouldn't dive too deep into that but today what I really want to talk about is what kind of startup opportunity I saw with this mCP ecosystem blooming and also a stepbystep process of how can you build a mCP server from scratch po play and distribute to others and at extremely high level mCP basically provides a unified way for your AI agents or AI application to access external assistance as example a few weeks ago I post this video about how I use all sorts different mCP in my AI coding workflow that made AI coding agents a lot better and if you have been building AI you probably wondering how is this mCP different from T call which is something introduced by open AI almost two years ago that also allow AI agents and large L models to take actions external assistance where the key thing here is a standard even though large L model can already take actions in real world through function calling but each one of those large L model provider has different format like open a will support a two core format looks like this with name arguments and become part of conversation history while claw have a totally different Json format same for Gemini and llama so each one of those large L model provider have different default ways of interacting with external system but mCP provide a unified format of how two code should be communicated if you want to draw some analogy this somewhat similar to the early days of Internet there's a portol called tcpip where before TCP IP each company has its own standard about how different device should communicate each other like IBM have their own protocol called sna and even Apple has their own thing called Apple talk they allow macing touchbased Network to talk to each other but if you want to build something that integrate across different device that means you need to build custom things for every single standard but with introduction of standard like TCP IP everything became unified which means if you're raer you just need to build once with this protocol standard and it allow you to build for the whole internet and this changed the value dramatically new protocol like worldwide web https all came out and enable business like ncape yaho Google and internet eventually became where it is now and this also part of the reason why I found mCP is so exciting and we should pay a lot of attention to it because there will be lots of opportunities and companies to be built around this ecosystem for example with unified protocol like mCP the entry point for building new AI agent clients like cursor Wing serve Manu becomes lower and lower because you don't need to build all the Integrations yourself you can just connect to existing mcps and this kind of draw the analogy in my mind would be similar to iPhone windows or even browser like nascap that directly interact with end user and those AI agent clients will be the main driver of adoption for all sorts different mcps for example this concept mCP isn't new is introduced almost 6 months ago but only exploded recently because with finally have killer AI agent use case like cursor and Wing surve which is just one type of use case like coding agents but with more vertical AI agents show up like sales agents customer support agents or even general purpose agents there will be needs and requirements for different mCP servers and on the other hand there also a lot of opportunity for creating a Marketplace for mcps there already existing website like GL and Smith where they help people to discover what type of mcps are out there and a lot of hard work in terms of testing and curating those mcps it's kind of similar to how those existing ecosystem has app store or early days of Yahoo has index page for people to find the interesting websites but the most important one is mCP itself they kind of like apps in app stores the any iPhone and Android users can go and access to enhance their experience and even though at moment mCP is like a wi web where everything seems open source and still quite difficult to use but there are already companies like 20 First Step introduce new products that is the mCP the existing in clients like cursor Wing surf that enhance their coding experience specialized in helping AI agent to get more context of how to build beautiful UI and they add G monetization for those mcps so I actually expect a lot of new business to come up that just building mCP and distribute to All Sorts different users and to monetize from that that's why today I want to show you how can you build a mCP from scratch so that you can either build mCP for yourself to enhance your own experience or you can distribute that to others and even try out building a mCP based business and the good thing is that it is not hard it's actually pretty simple and straightforward I'll take you through example of how did I build a figma mCP from scratch and list down those Marketplace like glama and symetry but before I dive into this if you are launching your own mCP or any type of product building the product itself just half the work after the product is built the next challenge is how to get your product to the market Market successfully nine out of 10 business venture fail and 10% don't even make the first year and the key difference here is often how solid your go-to Market strategy is especially when launching new prodot type like mCP that's why I want to introduce you to this free goto Market playbook for launching new products then made by hosital they interview top experts from successful startups about their best practice process on PCT launch it covers step-by-step process of how can you design a lunch blueprint between Market product and distribution Channel then where and how should you go find signals of the market that your product is targeting at how to craft the perfect Market story and step-by-step process to build your viral tweet it even including ready to use metric dashboard to track your new product launch with real world insights from goto Market experts giving Hands-On guidance I learned a lot from this package and I highly recommend you go and try out if you're planning to launch a product I put a link in the description below for you to download for free now let's get back to how to build your own mCP so entropic actually provide specific SDK for python typescript Java codling so you can use the language that you are most familiar with and for me I will be using python SDK so if you take a look at GitHub here I'll do pip install mCP so our open Terminal and just do pip install mCP we can create a new file called test mCP dopy so they provide a class called Fast mCP we can just call this fast mCP class to initiate a new mCP server with this one we can just do a decorator mCP do tool and then Define function here and when writing the function you can give a description about what does the function do as well as what kind of inputs does it expect and the output to give large L mod or more context and you can save this and this mCP is already working we can quickly go to your mCP settings in cursor we will right click this file and copy past and I will just add a new mCP here I give name my mCP and command will be mCP run and pasting the pass here and click add and after it is running you can see this my mCP has been added already with the function we defined here so with this turn on I can already open the chat and then prompt the cursor agent help me calculate my bmi where I'm 70 kg with 173 CM height you can see here it call the calculate BMI function pass on this two variables and then get result back so this is how easy it is to actually set up your mCP server and to practice this further let's try to build a figma mCP where can read your figma file and also convert Your figma Design into hity react page so first let's take a look at the figma API so figma actually only have four end points for you to get data out one is get file which will return everything in your specific project file and get file node allow you to specify a very specific node ID so in figma every single UI components you see here on the left side is a note and you can actually copy the link to selection and if you paste the link here you can see that it also has node ID that point to that specific frame so this endpoint will allow to get a very specific component or page out of figma file as well as get image and get image will return the image screenshot of the mup but unfortunately AI coding agents like cursor or Wing serve can't really take image response from mCP yet but with the first two endpoint we can already do a lot of cool stuff so I will be using this endpoint of get file notes to return the details of specific frame so I'll create a new EMV and go to my figma click settings security and click generate new token given name test and generate it so you can copy this token over and paste in here so I go back to our test mCP file and add a few new Imports and our load the DMV and load the figma API token first they create a helper function called Fetch figma nodes where it will pass on the file key and node ID and calling this endpoint which is what do we see here and then we will do mCP do2 create function called G Noe where where we pass on file key and node ID and our add some descriptions here that it is to get specific node from figma file and I will also pass on a bit of information about what's the file key and what's node ID and its format and then our Define function where our Cod API and point return the response so we can quickly test out I'm going to copy the link paste in here and extract file key and node ID so I'll do python test mcpi cool and you can see it return uh a Blog of Json of how does that mup look like great so this function is working and all we need to do now just go back to cursor mCP and if we refresh a little bit you can see a new function called G node is showing up here and just to test this out I can add a new agent and I will just copy the link and here I have extremely basic mop so I just right click copy the link to selection and go back here and give Implement UI like figma mop here in Pixel Perfect manner in index.html you can see it called again note mcp2 and it return whole bunch of information about this UI that we Define in figma then start creating this HTML page cool so now if we open this webs page you can see this is Page generated it kind of get most part right but some part is looks a bit off and you might be wondering why and the key reason here is because if you look at response in return including whole bunch of information that is not relevant or very complicated for large L model to understand like the color is rgba format which you can probably communicate in a lot less token so what we want to do is we want to clean up the information a little bit more and this is exactly what I did so I have a new file called clean node. py and in here I have a whole bunch of helper function that can help you return more clean information and I actually got a lot of Inspirations from another mCP server created by Graham but basically do a whole bunch of clean up of the information and I can load those functions from clean node in transform fig adjacent and come back here at a clean response equal to this and return clean response and go to mCP and refresh again and now let me try this again and now you can see the UI looks exactly like the UI we present in figma file so this how you can build a new mCP from scratch that's running on locally but you can also distribute to others the next thing you want to do is probably distribute your awesome mCP server to others and there platforms like GL or Myster they're providing existing repository of mcps for you to distribute to others so all you need to do just write down a doc with all dependencies and read me so other people know how to run this and cursor is actually pretty good at it so you can just give a prompt I have this mCP server implemented help me write the required. dxt and read me and once it's done you can upload this project to GitHub and then go back to like GL click on ADD server given name description as well as your GitHub Ripple once you add that it will go through a process of reviewing and if everything passed then other people will be able to search and install the server directly and I have already put my figma mCP on glama and Smith Tre where you can just go and download and for the figma mCP I will continue to improve that I already added new functionalities like get components so that it can extract all the reusable components in your figma file as well as workflow so it can get the figma protype workflow to better understand the use flow if you have any feedback or question about mCP feel free to DM me on Twitter or you can join my AI Builder Club Community where I will continue sharing interesting mCP and AI coding workflows that I'm experimenting with detailed breakdown of some my favorite mCP that I use day-to-day as well as more in depth course and tutorial about AI coding and building production agents but most importantly we have this community of top AI Builders who might already experience a problem that you are facing right now as a solopreneur so you can come and ask any question you have to get support I have put the link to join the community in the description below I hope you enjoyed this video please comment about what type of mcps that you are building or you really want others to build thank you and I'll see you next time