Cursor and other agentic IDs have a big limitation. Every model they use has a training cutoff. I'm using Claude 3.7 right now and it stops its learning beyond a certain point. You might have noticed this already if you've used it to build full stack apps. If I tell it to install Shad CN UI in a project, it still runs the old install command. That command has been replaced, but the model doesn't know the latest version because it lacks updated context. I tried using the add docs feature to fix this. I added documentation for different frameworks and it helped a bit. But the agent pulls from both the project and the documentation at the same time. That makes the context heavy and the output code is often poor. Things go more smoothly when the project is small. For example, with a lightweight library like MCP use where everything fits into one Redmi file. But the moment you move to something bigger like a Nex.js app with microservices, the system starts to struggle. Now there's a much better fix available. What I'm talking about is the context 7 MCP server. Context 7 is a site with a curated list of documentation. It lets you pull specific pieces of information with code examples which when given to cursor agent help it code accurately. This way it doesn't mess up when handling unseen code. You can see that the tools are active here. Don't worry, I'll show you how to set everything up. But first, let's test it out. I'm going to open the agent and ask it to build an MCP Python agent for Airbnb using the MCPUs framework. This framework is used to build AI agents with MCPS. It's a solid option. I've already made a video on it, so check that out if you haven't. Let's give the agent the prompt. It has started working. And in this case, we're using the docs feature from cursor agent. The documentation here is small. There's no real chance of errors. The full documentation for MCPUs is that one file from the GitHub repo. It's not a lot, but let's see how much of a difference it makes. The agent has opened and I can see it made the correct imports. Everything matches the source. Nothing unexpected. I've accepted the changes. This is the final file. I ran it and it works. Now let's try the same prompt using the context 7 MCP. I'm using the same prompt. So let's see how it builds the agent. This time you can see it's searching the MCPU's documentation to understand how to build the agent. When we open it, it tells us what it's doing. Shows a basic MCPU's agent implementation and gives a code example too. In larger projects, like I mentioned before, this helps avoid breaking the code base. Let's look at what it builds. It's done. Let me just accept the changes and then I'll look at the code. This is exactly what I meant. It included the Playright MCP, created an advanced Airbnb agent, and then added more features to it. It extended the agents capabilities, and that would never have happened if we used the simple docs. The simple docs just show how to implement an MCPU's agent. But here, because it pulled specific pieces of information, the agent had more room to work with. Now I'll test it. Everything runs fine. The agent works and everything is set up correctly. I tested it out and it works. It's actually pretty amazing. Not overloading the context really does help the agent perform better. Another improvement in cursor is security. You can see that certain AI features are disabled in myv file where I've added my API keys. The file is declared as cursor ignore so your information won't get leaked. So for anyone who says you shouldn't build your backend in cursor or handle things involving API keys, here's another layer of security added. This is the GitHub repo for the context 7 MCP server. And the installation instructions down here are now pretty clear. Since all MCP clients now use a standardized way of connecting to these MCP servers, they all use the same methods. There's no authentication or API key needed. So it's really straightforward. We have installation support for cursor windsurf and now also for VS Code thanks to its GitHub copilot agent which can now use MCPS. You just have to copy the command and paste it. For this video we're going to be using cursor. Inside cursor go into your settings and from there navigate to the MCP section. As you can see I don't have any MCPS configured which is why it's giving me this error. I'm just going to add a new MCP. Paste the configuration I got and save it. If we go back, you'll see that the context 7 MCP has now been added. In case the tools don't appear like this, you can try refreshing first and it should show the tools. If that still doesn't work, just quit cursor and reopen it. That should fix it. What's actually happening behind the MCP server? Let's go to their site context7.com. You'll see updated documentation for different frameworks. Many of them are already supported. For example, we have Nex.js here. And I also noticed MCPUs listed. The documentation is really up to date and if something feels missing you can request more. Right now they have documentation for around 800 frameworks which is impressive. If we open one like the NexJS documentation you'll see everything laid out clearly. It shows the total number of tokens. You can scroll through and find exactly what you want to learn. For example, search for pages router and all the related information shows up. There's also an option to increase the token limit to get broader results. You can copy any of this content and use it directly with an LLM or inside any code editor you prefer. There are also a lot of features being requested. It's a really active project. And one feature I'm especially excited about is documentation versioning. In larger projects, once you're working with a specific dependency, you need to refer to the right version of the documentation. This is going to help a lot with that. So, I hope you enjoyed the video and you're going to start using the MCP server. It's totally free. Doesn't really take much to set up and it provides a whole lot of value. If you see that anything is missing, please request a documentation to make it as much better as possible. And if you like the video, do subscribe. And if you want us to keep doing this, click the link in the description to donate to us. Thanks for watching. We'll see you in the next video.