5 MUST Use MCP/Agent Skills

追加: 2026-05-23

[00:00:00]Things like MCP and AI skills are more important than ever before. This is because they plug directly in your AI harness, whether that's claw code, codecs, anything specific that you're using and they make a lot of things possible which you might think is not possible with your AI alone. In this video, I want to give you a runrough of top skills and MCP that you can find online today and some of them I actually use on a day-to-day basis as well. So, let's get started. All right. So first things first, in order to understand what an AI skill or an MCP is, we have to understand a few things before we get into my top list. First is an AI model, right? So this is something that we call as you know Opus 4.7, GPD 5.5, Deepc V4 Pro, all these models, right? A lot of them are there. Now on top of AI model, there is something known as an AI harness. And this is basically a software that is using this model to achieve something you want to be done.

[00:00:54]Now in this case for example cloud code or codeex over here these are like harnesses which are on top of built- on these models like these are supposed to use these models in some capacity. Now sometimes providers like Entropic would want you to limit cloud code for example with just OPUS 4.7 or you know entropic models generally but then there are harnesses like open code for example which support like a bunch of these models right including the paid ones like Opus and GPD but also free ones like you know some free ones whatever is free right now they keep on changing now a next step on top of this should idly be an AI skill or MCP right and a lot of confusion that I've seen a lot of people what end up saying is that MCPS are sort of or rather I'm saying it reverse AI as scale is sort of like an evolution on top of MCP it's not right and let me tell you why AI scale or agent scale rather I think that is what the cool guys call it very simply speaking this agent scale over here is nothing but a markdown file right as simple as that of course like you can keep it an HTML file as well but the typical idea is that agent skill would be a markdown file and that's it so what this means is that the skill over here is justformational it's purelyformational And it's telling your AI harness to do something or rather your AI model because your AI harness is going to feed this markdown file directly inside the model. Now depending on you know when it feeds and what context it gives that's like a different question altogether but all you have to understand is AI agent skill over here is nothing more than a simple markdown file that includes a bunch of instructions on what extr needs to know.

[00:02:28]For example, to give you an example, let's say internally you develop a skill that you call as coding practices.m MD, right? So this is that's fine. It's basically a skill at this point and in this skill you just write down you know how your code should look like. What should be the variable naming practices soft skills that you cannot enforce with a llinter or a type checker, right? So that could be like sort of like a skill.

[00:02:50]MCP on the other hand is interesting because MCP generally is powerful and it can also provide you context and the ability to execute stuff. So what I mean by this is imagine that you install a Postgress MCP inside your harness, right? So this Postgress MCP over here what it can do is it can not only give your AI harness or your AI model generally some information like saying that hey you can call these tools right?

[00:03:19]So you can give them MCP is basically like exposing a bunch of tools, right?

[00:03:23]Let's say you give it an execute query tool and you say that this would run SQL query. So MCP generally not only would say this instruction which is sort of again like a markdown thing only to your AI model but it will also provide the implementation and when your AI actually calls this this large language model over here when it decides to call this execute query in future your MCP is supposed to execute that right so this execute query would be a real programmatic function that is defined by hand not on runtime inside your MCP itself in order to execute that scale right So these are not independent things. In fact, your agent skill can include a bunch of MCP calls as well within them. It's completely possible.

[00:04:07]So your agent skill defines a bunch of instructions. You know, this has to be done this way, this has to be done that way. And then it also points down to a few markdown a few MCP calls rather that you should use these MCP calls in that specific order or you know when something like this is needed use this MCP tool call and so on and so forth, right? So it's it's all about packaging in a way. All right. Now we are done with the basics. Let's get to the brass stacks which is like what is the list.

[00:04:33]Now before we get to actually the number of skills or you know what skills are actually good. What I want to also talk about is the harness that I'll be using in this video and that would be claude code and you can basically use any harness of your choice. Right? It doesn't really matter. A lot of modern harnesses actually support skills MCP server out of the box. Right? So you can use agent skills as well as MCP servers directly with any harness. you'll have to look for some options here and there.

[00:04:57]But with cloud code because it's cleaner because the file system is exactly there. I can show you where the skills are. All right. So I created this empty directory in which I would try to install and use some of these skills now. So with number five I will start with and this is in no specific order.

[00:05:13]So this could be like you know out of order. These are like I'm not saying that this is better than the others that I'm showing you and vice versa. But the first and the number five skill that I have is stripe projects.dev skill. And this is something that I've actually done a dedicated video on as well. You can check out that video somewhere here if you want. Projects.dev is basically an initiative from Stripe of which you can install the skill over here. So if you click on this, if you visit projects.dev and click on this to install the skill, you'll be able to copy the skill, go back to your directory and hit this command over here. If you're using clot code, look at this. We are using npx skills add and then this specific thing right docs.stripe.com. So basically in this case what we want is the stripe project skill. So I'm going to hit space and I'm going to hit enter over here right and I'm going to select my agent to be cla and I want this only in this project and I'm going to sim link for now. So the idea is once this installation completes you're going to see that over here inside this folder we have a new structure. We have this agents and claude as well. So the problem is that every single AI harness tries to build their own sort of syntax for doing things right for example claude has this claude folder over here in which there's a skills folder and over here you can see that stripe projects is there uh it's a folder again and over here you can see that it's skill.md file is what we want and as you can see the skill.md is basically a markdown file right so this is all information you can even read through this if you want which just tells it you know what's going on that use when the user needs to provision a third party service available on projects.dev, create or retrieve a provider service API key and so on and so forth. So this is basically just knowledge. This is just instruction fed to your AI agent so that it's able to do things properly, right? It has some important things like non-interactive mode like always use JSON and yes to flags to supp surpress the interactive promps and even like confirm paid service sort of things, right? And a bunch of more knowledgeable things. So whenever our hardness uses the skill, it knows how to use it. This is actually good because it allows you to provision like infrastructure just using your AI.

[00:07:20]And this is not I'm not just saying that it'll just pull up free services and stuff. This can actually literally go to every of these websites whether that's Cloudflare, Clerk, Firecrawl, any of these websites and it will go ahead create an account for you create the API keys even rotate them for you and you can manage everything just from CLI just because they are partners of Stripe. So Stripe does the KYC for them. So it establishes that you are legit and in turn these companies have given sensitive API access like creating and registering users and you know every doing everything programmatically over to stripe. Number four that I have is remotion skill. Now this is a skill plus an MCP server that allows you to create videos using cloud code. So as you see if you go to reotion and you go to the skills columns over here we still have this interface like the same interface basically. And if you go back, let's say we go back to our home directory. Add the same command. You're going to you're going to sort of get the same UI. So I'm just going to enter cloud code and move forward with project sim link. Yes. And once this is done, you can go ahead and see inside your directory over here inside claude skills remotion and then skill.md. Right? So again, it teaches your agent on how to use reotion in the first place. It does not tell it how to execute code or what commands um you know it's available as an MCP server. It just teaches how to use it in the best possible way. Now let me show you how this works. So I'm going to go back a little. I'm going to launch claude in this directory and I'm going to say can you create a YouTube like and subscribe animation using reotion video scale. I'm not giving it a very good prompt. I just wanted to use the skill that it already has. As you can see, it all automatically loaded the scale because it figured out that I'm working with videos now and it's just going to create a project on that. All right. So, if you look at this reotion MCP that we have, let's see what it has done. So, this is the like subscribe animation folder. We have a full project going on over here. And I'm just going to do npm rundev over here. And it basically opened a full timeline project for me, right? And if I play this, you can see that I have this like, subscribe, and press the bell icon available to me. And this is actually pretty cool, right? I now have an asset file which I can just go to file or something to render and you know I can just render it in whatever quality and format that I want and I'll just get an M get a like a MP4 file. So this is pretty cool way of working with remotion where you just added scale. The scale includes enough commands and stuff that it does not need a separate MCP server but it can just work directly. All right coming to number three. I would put consensus, MCP and the agent scale over here which is a system that offers MCP as one of their tools which allow you to do academic research. You can search from peer-reviewed papers, you can build search strategies, generate structured outputs. If you're doing anything that requires research at some capacity and ideally if you're doing it for your academic work, research papers, any of that, you can use consensus MCB which also happens to be sponsor of this segment. and you're going to see that it works perfectly with chat GPD claude and MCP clients other MCP clients as well if you're not using any of them. Now let me show you how this works. Once you log in you can use either their interface that is over here and you can select the amount of research papers that you want in the interface that you are working on or you can also use their MCP server which is what I'm going to do. So in order to add their MCP server I'm going to quickly follow their documentation over here and I'm going to use this command cloud MCP add transport and then the whole thing. Now remember I told you that skills and MCP servers are completely different things. So far we have skills added right that's why we just have agents folder but MCP server requires it to be inside your harness.

[00:11:10]For example if I write MCP over here and hit enter you're going to see that I only have this claude.ai Google Drive and this computer use nothing else. So what I'm going to do is I'm going to go outside of claude paste the same command that I just copied over here with claude. Hit enter and as you can see they say that ooth will open in your browser automatically on first use. So MCP obviously is you know just a collection of tools exposed over here for example you know this is something that I mentioned like execute queries the tool but you also need to pass in more information right execute on what database is the user even allowed in the first place is the database user allowed in the first place and so on and so forth. So there are a bunch of ways to fix that like how do you bring authentication to MCP which by the way is a completely different video that I did a few weeks back so you can check it out I'll link it somewhere here again but the idea is that in case of consensus they are using oath as a strategy so we can also look into that so let me start cla I'm going to add MCP here and you can see like consensus is over here so if I hit enter I can click on this reauthenticate right and you see a browser window will open and as you can see it asks me to connect consensus to cloud code. I'm going to give it a connect and authentication is successful. I can go back to my clot code. Awesome. Now, one thing to note over here is that if I exit it from here, what you're going to notice is that we don't have a skill for it, right? If I go to clot skills, you're going to see we don't have any skill over here. Well, like I mentioned, agent skills are nothing but just markdown files, right? So, we can literally create a skill on our own. So, I'm going to create a consensus MCP as a directory. I mean, it's not really MCP.

[00:12:46]We can just keep consensus as a name. We can open consensus over here and I can just create a skill.m MD file. Right? So I'm not going to give it a huge skill obviously like that is how the thing goes that you generally give like all the instructions but I'm just going to keep it very simple. I'm going to say that if there is any research paper or anything that needs deeper research, you must run it through consensus MCP. You already have access to it, right? That's it. Our skill is very simple, very small. Nothing going on over here. But yeah, I hope this should work and Claude should be able to pick up the skill automatically. All right.

[00:13:32]So we have deepse 4 paper over here which I'm going to use as an example for consensus MCP and let's see how it plays out. So I just downloaded it and I want to tell it can you please tell me the gist of this research paper. So I want to do it in two parts. First is look at this where I'm asking it to review the Deepseek V4 paper but don't use any skills or MCP and I can check like whether it would use the skill or not.

[00:13:58]As you can see, it just started using or doing it without any MCP, right? So, we do not have the consensus MCP going on over here. So, it has given me a bunch of information over here which I'm not fully sure if it has actually gone through the PDF or it has cross-cheed the references and actually, you know, come up with the right numbers or not because a lot of times these AI do hallucinate, right? So what I what I'm going to do next is I'm going to place this one here and this time instead of this I'm going to say use the consensus MCP skill that is there and as you can see it has loaded the consensus skill which is not really anything but sure.

[00:14:39]And as you can see this time it's also calling consensus with the following phrase deepseek sparse attention. And I don't know if you noticed this or not but on the left which is the consensus MCP answer that is much more crisp in terms of like what it is listing out the optimizations and so on and so forth.

[00:14:55]And the good thing is that you can also go deeper into this conversation right it's not just a one prompt thing.

[00:14:59]Obviously the MCP can be used over and over again. And with a lot of these things it's also giving you references.

[00:15:05]For example this deepseek retained is a reference number 11. So I can jump over here and if I can literally open this and when I open this it basically throws me back into consensus portal which gives me the actual paper which is the reference and then you know you can just start asking the paper itself as well which is pretty cool. So use this consensus MCP for doing research work which is peer reviewed and it's actually reliable unlike the hallucinations that we see with AI a lot of times. All the links are in the description for you to check out these MCP servers including consensus. All right coming to number two for me. I'm going to give this Chrome Dev Tools MCP. And let me show you how this works. So, if you go through this Google blog post, you're going to find that there is a way for you to give access of your browser dev tools to your AI. What this basically means is that clawed code or any other MCP system that you are using, for example, any other AI system that you are using can access your Chrome tabs if you give them this MCP control. Right?

[00:16:05]But first things first, you have to go to this URL in order to enable remote debugging. If you don't do this, then this would not work. Once you do this, you're going to see that you get a line here that says server running at 127001 9222. And then you can use the Chrome Dev Tools MCP. Awesome. Now we have something as an MCP to add over here.

[00:16:23]You can see over here it does not give us that simple command that consensus MCP gave us. So we're going to just take this, copy this over here, go back, open our cloud again, and write MCP over here. You're going to see when you write this MCP command, you will get a path of this file over here which is your local MCP right where they are stored. So let's open this. And if you scroll down to your MCP server list over here inside the folder that you were using in this case I was at agent skills MCP in this file. You will see that there's a list of consensus is already there. I'm going to take this chrome dev tools part only and I'm going to add it over here.

[00:16:59]Right? So unlike consensus that's using an HTTP type, this one is using npx. So it runs locally on my computer. Let's save this and let's try again. So now you see that I have this chrome dev tools as a connected MCP over here. I can open this and I can see the tools and you're going to see that it has all these interesting tools which are there.

[00:17:18]Right? Awesome. Now you can go ahead and create a skill for this as well if you want just like I showed you for consensus MCB. But for now I can just mention it in the prompt itself and it's going to pick up that MCP. So I'm going to say can you go to example.com and check the performance of the website right because remember this has access to devto tools this is not something that you can actually say to any other random AI MCP server that's just able to browse the websites right this actually requires you to have dev tools because that is where the performance tab is so hit enter and let's see you can clearly see that it can call the chrome dev tools okay one thing to see here is that they asked us to add this channel beta which is only till Chrome 144. So you can just update your version which I already have and let's just remove that.

[00:18:03]So I'm going to remove So I'm going to remove this whole part and I'm going to save the file again and I'm going to copy this and start cloud again. Paste this again and let's see. And as you can see this time it just opened it automatically on its own. I'm going to allow debugging and as you can see it's opening tabs.

[00:18:20]It's browsing the website on its own.

[00:18:22]And if I go back you can see like it's also able to use the dev tools. If I expand this, right? Right. So it's starting this performance start trace and it can you know do a bunch of things. You would not see dev tools directly over here but it can this MCP system allows Chrome to expose important information to your agent. And as you can see you have a lab metrics field data all of that. And you know what can be a pretty cool way to work with this.

[00:18:47]You can create a goal and you can say that hey I want my website to hit 100 page score. Keep fixing bugs and deploying online and check and do not stop until you get the score. Pretty cool way of working right basically forever. This like this is like a great way of burning your tokens if you are on a heavy website. But anyway what I'm showing you is that it's great that you can do something like this. And last but not the least, context 7 MCP server. So this is basically a server that allows you to retrieve up-to-date documentation. If you're working with something that is probably new like for example a new NexJS framework or something which has not been indexed that well by LLMs, you can use this context 7 MCP server to get proper documentations. Right? Right. So I'm going to use cloud code. And as you can see just like consensus they have an interface over here. But this asks you to specify your API key directly over here instead of like doing an oath which is slightly inconvenient but sure. So I'm going to run this command with my API key which I'll reset after this. And as you can see that it has added this MCP server over here. And now we are good to go. Now to give you an idea of this look at this zero is a programming language that Warsel very recently released. Right? I created a video on this literally like last couple in the last two days. However, this is available on context 7. So now if you have the MCP with you for context 7, what I can do is go ahead and say can you create a small folder for zero programming language recently released by Versel and create an example for it. use context 7 MCP but the idea here is that you don't have to worry about what you are doing right if you are doing it with a new framework new version chances are that context 7 would provide fresh context inside that version inside that language whatever you're doing for you free of cost as you can see claude is working with context 7 to understand a little bit about the language because it's completely new so it wants to get it right and it's now doing the work so it has done everything I can ask it can you check the file fails by running the command and it can probably also give me the output for that. I mean you can technically argue that hey this could also work without you know having context 7 because claude can browse the web. True but then you will end up wasting a bunch of more tokens instead of you know just using something like this which can fetch the right things exactly at the right given time. So yeah that's pretty much it for the list of five top four five MCP servers and skills that I can share with you. And of course, this is not a technology which is like only restricted to people who are creating these skills.

[00:21:35]You can also do that. In fact, I have a video on how do you create your custom MCP skill as well. I will link it somewhere here. You can check it out.

[00:21:42]But yeah, do let me know in the comments what do you think or any of the favorite MCP servers that you use in your daily stack that I have missed out on. So that's all for this video. Hopefully you liked it. If you did, make sure to leave a like and subscribe to the channel. Do check out Consensus MCP. All the links are in the description. I'm going to see you in the next video very soon. If you're still watching, make sure you leave a comment I watched till the end below to tell me that you were still here. And let me know what do you think about the video.