Hey everyone, I am currently working inside cursor and I want to talk about a problem I have been running into. I use the Gemini 2.5 model and like most language models, it has a cutoff date. This means it does not have knowledge of newer tools or libraries unless you manually explain them or paste in the code yourself. Cursor tries to improve this limitation by allowing you to link documentation. For example, when I am working with the MCPU's library, I can simply add the GitHub read me and cursor is able to read the content. The issue is that it loads the entire documentation and attempts to process everything at once, which creates problems with context and can lead to confusion. To address this, I tried using context 7 MCP. It is a solid platform that hosts updated documentation and uses retrieval augmented generation to serve only the relevant information. This helped in some cases, but the results were inconsistent. Even when I explicitly told cursor to use context 7mcp, it often ignored that instruction and search the web instead. At times, it pulled in unrelated content that made things more difficult, especially when working with overlapping tools such as MCPUs. These issues became more noticeable when I was dealing with tools or frameworks hosted on GitHub because many repositories and resources are interconnected. Cursor ends up retrieving context that does not align with what I am actually working on. That is where git MCP becomes useful. It transforms any GitHub repository into its own dedicated MCP server with focused documentation. The setup process takes just a few seconds and provides Cursor with exactly the context it needs. In this video, I'm going to show how Git MCP implemented the new A2A protocol correctly without introducing any errors. It offers a straightforward way to improve cursors accuracy and makes the experience of working in coding editors much smoother. This is the GitHub repository for the GitHub MCP tool and it contains a lot of useful information. There is a great example that demonstrates how they built a 3JS project both with and without GitHub MCP. So, let's take a look at that. You can clearly see the level of detail that resulted from using GitHub MCP in comparison to not using it. The key difference lies in the context the AI model receives and how specific that context is. When the context is highly specific, the model is able to generate much better graphics as demonstrated here. That is not even the most impressive part. Other tools can feel overwhelming to set up. As I mentioned with context 7 MCP, there are times when it completely ignores the MCP server and begins searching the web instead. In this case, the setup is so minimal and lightweight that there is no friction involved. You simply configure it and begin coding. It is also completely free. You can self-host it with MCP servers if you prefer, although I do not believe it is necessary. It connects with any integrated development environment including cloud desktop, windsurf, VS code client or cursor. They have also demonstrated how to use it with a specific repository or even a GitHub pages site. This means that if the documentation is hosted on a GitHub pages site such as Langraph, you can feed it to the agent. However, if the documentation is not hosted there, you will not be able to use it with this MCP at this time. Now I will show you how quick the setup process is and how fast you can add the MCP server and begin working with it. This is the repository for the Google A2A protocol which is a relatively new standard designed for communication between agents built on any framework. We also have a video that covers this topic in detail and you can find the link to that in the description below. If you want to provide this GitHub repository to cursor so that it can begin building with it, especially since this is a new protocol, you will need to supply the right context. The process is very straightforward. You simply replace github.com with gitmcp.io in the URL and press enter. That action will give you an active MCP server based on the repository. There is no need to create any rules manually because it is already configured for all major coding environments. You just need to copy the MCP rule and paste it into cursor. Once that is done, cursor will have access to the documentation and the MCP server will deliver accurate and relevant instructions for using this protocol to build agents. You also have the option to interact with the documentation directly. When the page opens, you will see a clean chat interface that is powered by language models. These models are available for free and while they are not the most advanced, they perform well enough to answer questions and assist with using the MCP tool. You can ask any question related to the Google A2A documentation and the system will help you. I have a pretty cool example to show you. I implemented the A2A protocol between three separate agents and I did not read a single line of the A2A documentation beforehand. After completing the implementation, I went through the documentation to confirm that everything had been done correctly. I began by prompting it to explain the A2A protocol and how it works and it immediately fetched the documentation. One of the things I really appreciate is that when I mention the term A2A protocol, it automatically retrieves information from the A2A MCP because that is the name under which it is saved. When I used context 7 MCP, it would occasionally hallucinate based on how the prompt was written, unless I specifically instructed it to use the MCP. That added a layer of friction to the overall process. Once it retrieved the documentation, it provided a complete explanation of how the protocol works and what it is intended to do. After that, I gave it the actual task. I asked it to create three agents, including one main agent that I would interact with directly, one that discusses only animals and another that discusses only plants. Although this is a fairly simple setup, it demonstrates an important concept. You can connect smaller language models to rag databases that are focused on specific domains. There is no need to rely on one large model with extensive training data. Instead, you can break the problem down into focused areas of knowledge or tools. I then instructed it that the main agent should communicate with the second and third agents using the A2A protocol and that was the key requirement. It began building the agents and continued to call the MCP tool throughout the process which was helpful because it consistently referenced the correct documentation. It generated all of the necessary files and separated the logic for each agent. I also asked it to include a readme file and it created that along with a requirements file. I followed the steps outlined in the readme and the agents are now live and successfully connected. If you're enjoying the video, I'd really appreciate it if you could subscribe to the channel. We're aiming to reach 25,000 subscribers by the end of this month, and your support genuinely helps. We share videos like this three times a week, so there is always something new and useful for you to explore. I have started the agents in the terminal, and this is the main agent. You can see that it is running and is connected to both the plant agent and the animal agent. I asked it a question specifically what a lion would eat and it correctly identified that the question was related to the animal agent. It sent the request to the appropriate agent and received the correct answer in response. Here is the plant agent which is running on its own server and has not received any requests yet. And here is the animal agent running on another server where you can see that it received the lion related question and responded with the appropriate answer. This demonstrates that the request was routed correctly. Now I will ask another question this time about the most important requirements for plants to grow. After sending the question, the system routed the request and we received the answer. If we go back, you will see that the animal agent did not receive anything for this query, but the plant agent did. This clearly confirms that the system is functioning as expected. Once I confirmed that everything was working properly, I reviewed the agent structure to ensure that the implementation was correct. That was the only manual verification I needed to do in order to be certain for the purposes of this video. What stood out to me the most was the retrieval process. While it was building the agents, it was continuously retrieving relevant information. That is the aspect I found most impressive and what I believe makes this MCP server so effective. Before ending the video, I wanted to share something cool that I also found pretty funny. I actually used the Git MCP chat to learn more about the Git MCP tool itself. I placed its GitHub link into the prompt box and started chatting with it to see how it would respond. The main question I had was whether the tool uses vector databases and implements rag. It turns out that it does not follow that approach specifically. Instead, it uses large language model text from each GitHub page and the readme files and it navigates the codebase based on that content. It does support textual search, but it does not apply retrieval augmented generation in the traditional way. What impressed me the most was how it created the A2A agent without introducing any errors and implemented it correctly by even generating the agent cards. If you are not sure what I am referring to, you can check out my A2A video for a more detailed explanation. It is definitely an excellent tool and I would recommend giving it a try. I am not suggesting that you stop using the Context 7 MCP, but this was something I came across and found genuinely interesting. After testing it, I really like the results. You are welcome to try it yourself and see how well it works for your use case. That brings us to the end of this video. If you'd like to support the channel and help us keep making tutorials like this, you can do so by using the super thanks button below. As always, thank you for watching and I'll see you in the next one.