Transcription
We all saw when MCP was released by Anthropic, the hype was huge. So much content and so many products were being built around it as Anthropic has just released the USBC for AI models. But to this day, many people misunderstand MCP. It wasn't some new technology. It was a standard for AI models to communicate with other tools. That's all it was.
But the main thing was that Anthropic had solved a major problem in this AI space. And now OpenAI has gone ahead and done the same thing by solving another huge problem for us. Now the main problem that MCP solved was that before its introduction developers had to build custom integration code for every combination of AI model and tool. For example, the Figma MCP server developers were writing different code in different languages to integrate its APIs as tools for the agent with the model so that it could call it. But when MCP was introduced, it gave them a standard, a way that any MCP client with any model could connect to any other tool because standard MCPs were made for them and we just had to paste the config in.
But now with the advancement in AI coding, we have come across another problem. Now, a quick break to tell you about today's sponsor, Hostinger Horizons. What's the idea you've been sitting on and what's stopping you from launching it? I turned mine into a fully working website in minutes with zero coding. This is Hostinger Horizons, an all-in-one AI partner that acts as your designer, developer, and copywriter. All I did was tell the AI chat box, "Make a coffee shop site with a menu and online orders," and within seconds, it spun up a mobile friendly SEO ready draft. I changed the entire layout to a fresh new design, quickly edited the website text in seconds, and even added a full online store with just a few clicks. It supports 80 plus languages, automatically fixes code errors, and integrates with tools like Stripe, PayPal, and Google AdSense. When I was ready, I launched it with one click. With a yearly plan, you get free hosting and a free domain, plus 24/7 expert support. And on top of that, if you use my code AIABS, you get 10% off. And yes, it's backed by a 30-day money back guarantee. Stop waiting. Turn your idea into a real business today with Hostinger Horizons and launch your dream project in just one click.
Now, the main problem is that ever since we've had these AI agents with proper tools, the only way to use them effectively is by properly following the principles of context engineering. This means you need to have the rules and context ready for whichever agent you're using. That way whenever you initialize a new session, it isn't completely oblivious to what has already happened, what needs to happen, or what's happening in the current repository. For example, if you use Claude code, you already know it has this Claude.md file. Whenever you initialize a new instance of Claude code in a folder containing that Claude.md file, everything written in that file automatically loads into its memory. In my case, this is just a tool I built for making an automation with Claude code and it contains a workflow. So Claude code already has this workflow inside it. If I ask it "What is your primary workflow?" after initializing it in a new repo, it doesn't know anything. It has no memory except for the Claude.md because that file is automatically loaded into its memory without me saying anything. It reads the Claude.md on its own because that is its primary set of instructions. Claude does this because it has been programmed to by Anthropic.
Now the thing is these new AI tools don't really stop coming. Every day there's a new AI agent out there. As a channel creator, I have to test them out and I know most of you do the same to see if they offer a significant performance boost or not. A clear example is Cursor and Claude Code. For a long time, we thought Cursor was one of the best tools out there, but slowly Claude Code completely overtook the market. Now personally I only use Claude Code for coding but occasionally when new tools come out I do have to try them all like when CodeX was released by OpenAI or when the Gemini CLI launched with its usage limits and then Quen Code which is based on the Gemini CLI. The thing is whenever I test them I always try them on an existing project and to give them context what I need to do is copy the content. Let's say I'm in Claude Code right now and I want to switch over to the Gemini CLI. To use the Gemini CLI in this repo, I would copy everything, create a new file, name it gemini.md, and paste it there again. But as you continue testing projects with different tools, this becomes a really big hassle.
Here on X, Shad CN pointed this out by clearly stating that it has gone into complete madness with the number of configurations we have now. With the sheer amount of tools being released, each one has its own MCP configuration and its own folders that need to be set up. But an even more frustrating problem is their rules files, their context files. You can clearly see that right here. There's one for Claude, then so many other MD files for different rules. One for Windinsurf, one for Kira, one for Cursor, and another for GitHub Copilot. There are separate files for all of them. Even if you think it's fine to keep creating files again and again, that it's not a big problem for you. The real issue comes when you try another tool. For example, if I test Gemini in this repo and it updates something, it's only going to update its own file. Every single tool only updates its own context. And if you're really working in the flow and actively testing them out, you might forget to prompt them. Prompting them to refresh memory and put the progress or context back into the original file becomes another hassle on its own. So, this has turned into quite a huge problem, and it's only going to get worse with the constant stream of new tools being released.
But OpenAI has taken a step forward and has come up with a solution to this problem. Meet agents.md which finally solves the problem of having multiple context files by providing a single standard for agent context files. If I scroll down right now, you can see how many agents have already adopted this as the standard context file. Among them, obviously since this is from OpenAI, we have CodeX on here. The Gemini CLI can also use the agents.md file and tools built on top of it will also be using it if they're regularly updated such as Quen Code. Even Cursor and other web apps like Factory and Google Jewels are adopting it. So a lot of tools we use daily are already on board. The main goal of OpenAI here is to push this forward so that other tool builders also adopt it. Making sure that agents.md becomes the default context file their agents follow.
But one clear tool missing here and one that most of you are using right now just like me is Claude Code. Its absence entirely depends on Anthropic. They need to step up and start adopting this even though it's coming from OpenAI and honestly this would make things much better. Right now most of us are still using Claude Code. It is undoubtedly the best agent out there. But if we need to occasionally test or work with other tools having this integration would be amazing. Other than that, there's nothing particularly special or different about it. They've given some examples, but these are just standard examples of how to structure your context files. To add it, you simply need to create the file, and the tools agent should be programmed to look for agents.md at the start of its session. And it's not even necessary for Claude Code to completely remove its own file. They could simply add support for this as well. That alone would make life so much easier, especially for me as someone running a YouTube channel who constantly needs to try out different tools on real-world projects. If Claude Code adopted this, it would honestly be a game-changer.
Now, if you scroll down on the site, you'll see some instructions for tools such as Adria and Gemini CLI. These tools can be configured, which is why the instructions are there. But since Claude Code is not open source, it's completely up to Anthropic to make sure it recognizes this file when you initialize sessions. If they do this, we can finally get rid of at least one side of this mess. Unfortunately for MCP, we're still dealing with the same situation right now.
That brings us to the end of this video. If you'd like to support the channel and help us keep making videos 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.