Transcription
A new war officially broke out on X and Reddit. Many people said that Claude Code was scamming their users and that they were switching to Codeex. Others said they didn't experience any degraded performance and that it was just a simple trend.
Now that even Claude admitted the performance loss and are trying to fix it, there's still one thing that Codeex has that Claude doesn't, and it's the larger context window with 272K tokens. Now, if you want the larger 272K window, you're going to lose some features that only Claude Code has and Codeex doesn't yet. But even if you are going to switch, you need to know these three tools that give you the added functionality, such as getting sub agents inside the Codeex CLI. But getting another $200 agent to play around with isn't cheap.
So, here's a message from our sponsor, make.com. A powerful visual platform that lets you connect apps, AI agents, and automations in real time, all without writing complicated code. With its simple drag and drop interface, you can design workflows that fit your exact needs. With over 2500 pre-built apps and full AI assistance, make.com makes it easy to combine Gen AI tools, LLM, and your favorite productivity apps into workflows that save hours of manual work. Instead of juggling repetitive tasks, you can automate them once and let Make handle the rest. As your business grows, Make.com scales with you. Its real-time monitoring, analytics, and unique make grid give you complete visibility and control over every automation you run. Whether you're building solo or part of a large team, make.com transforms complexity into clarity. Click the link in the description and start building today.
So, while I was testing codecs, I thought that to show you comparisons, I'd quickly build a site or something like that. But the thing is, both of these models are extremely good at React Nex.js JS and making user interfaces even when they use shad CN components. If we're talking about real project use, it doesn't only depend on the capability of the model or the agent. It depends a lot on the context it's been given and the context engineering workflow it's been through. I wanted to actually oneshot something complex. So, I asked it to build an Angry Birds clone, but I didn't give it a normal game in 3JS or simple HTML. Instead, I asked it to implement it with phaser 3, which is a proper game engine, and with Kenny assets, which is an asset library with the blocks and textures like they have in Angry Birds. I gave the same prompt to both Claude and Codex.
Now, this first result is from Claude. I tried to oneshot both of them and Claude came up with this. In the first version from Claude, when I clicked on the play button, it just got stuck. So, I had to reiterate it. After that, it came up with this. I have no clue what this is. it completely couldn't get it right and didn't use the assets either.
As for Codeex, this is what it came up with. In the end, both of them failed to make a fully working game, but to me, the implementation from Codeex is much better. At least it got some functionalities to work, and this at least looks like a game. Of course, the 272K context window really helps here. Oneshotting wasn't possible in this implementation. I'll have to set up a whole context engineering workflow for codecs. I'll tell you more on that later.
As far as the capabilities of the models, the model I used here was GPT5 Codeex High. I also tried to get them to fix the game, but it was too hard and just got too complicated for the model.
When people switched over to Claude Code, one of the biggest issues for new users coming from cursor was that they didn't find it easy to use Claude Code inside the terminal because it was a new and unfamiliar interface. As a result, many people began creating UIs for Claude code that worked much like cursor. All you had to do was drag and drop and a lot of problems like doing heavy configuration in the terminal were solved. For claude code, one of the best ones was op code which was previously known as Claudia. We featured it here on the channel as well. They also had a coding agent and an IDE extension that allowed you to use claude code directly inside cursor or VS Code, but to me that wasn't really great. There were also very functional UIs such as conductor which was excellent at orchestrating different agents. Its main function was basically to monitor your git work trees.
So, you're probably wondering if you're switching to codeex, what would you use if you don't want to use codeex itself? I found open agents which is a desktop GUI app for codecs and it looks like this. It's a small tool without many features, just the standard ones. You have your chats right here. It fetches previous ones which is causing some issues but other than that you can start new chats and adjust the reasoning levels. And that's pretty much it.
I also found another one called Codeexia. At first, when I looked at the preview, I was really excited because even though it doesn't look great, it seemed to offer a lot of functionality with MCP and different agent modes. But when I downloaded it for Mac OS, it didn't work for me. The file was corrupted. Maybe it works on other operating systems like Windows or Linux. I haven't gotten it to work here on Mac OS yet, but it's still actively in development with the latest release only coming out about 18 hours ago. I hope that they fix it. I'll be providing the links to these open-source repos as well, so you can monitor them.
In my opinion, the best UI that you could use is the IDE extension for codecs. Unlike cloud code, this one is actually really good. For example, it's a lot like the cursor window. The experience is really smooth. You can select different modes, choose different models, and if you set up a cloud environment for codecs, you can also have external agents running there automatically doing your tasks. On top of that, MCP configuration with this is really easy. You can directly open the file and edit it right here. No need to open the terminal. There are also some small settings you can change. Overall, it's just a really nice user interface. Pasting images works great here, too. You can see the previews as well. So, yeah, this is definitely better than the other options I've tried, but if you want a standalone app like Claudia for Claude Code, then you might have to use those.
Now, one of the best features of Claude Code was its sub aents. Sub agents basically gave you separate agents with specialized expertise and totally separate contexts which allowed them to fully take advantage of the context window and do one thing really well. In cloud code, you'd have your main agent which acted as an orchestrator and then you'd make use cases out of your agents such as a UI agent. All we had to do was write a good description. Well, Codex doesn't have anything like this right now.
So, I actually found an open-source repo that solves this, and it's called Codex Sub aents MCP. Basically, it's an MCP server that gives Codeex the ability to run these clawed code style agents. Here's what happens. Whenever the MCP is called, a new codeex instance with a completely separate context is initiated and a context for that specific agent is injected via the agents.m MD file for that instance. For example, let's say you want to do a security review. The main codeex instance has a tool called sub aagents.delegate which picks the appropriate agent based on the description and gives it the task. The delegate tool then looks up the agent, makes another instance, adds the new context layer, and that instance completes the task using the specialized context it's been given, then returns the result to the main codeex agent. This way, you're replicating the functionality of claude code.
If you take a look, you'll see all the agents that are already present in there and you can create your own as well. Essentially, you give it a task. Then first it lists all the agents as a confirmation call and then after that it delegates the task to the appropriate agent. The main agent also uses the agents.m MD file and is instructed not to do the work itself but to route tasks through the other sub aents. However, this is an open-source repo and as you can probably tell from the number of stars, it's not very popular. the author hasn't done much ongoing work, so there are some issues I had to manually fix to get it working.
If you do want a detailed video on this and enough people are interested in switching over to Codeex and actually using it, then I'll make a whole separate video showing how to set it up and create proper agents so you can use it. If you want to explore it yourself, you can do what I did, fork the whole repo, connect it to Claude Code, and experiment with it.
Now, you clearly saw how easily Codeex failed when it came to actually implementing something new with Phaser 3. Oneshotting things, giving your AI agents a single prompt and expecting them to add features again and again just isn't how software works. For that, you need to pre-plan everything. And this is where your context engineering comes in.
One of the best methods out there for context engineering is the BMAD method. What it essentially does is put your AI agents on track with the agile method where they plan everything out first and then break the implementation down into chunks. Those chunks are then worked on one by one, tested one by one, and that's how you build the whole software. Each task contains everything your AI agent needs to know about the task. You can see that it has made its own tasks here, but this is just for tracking. With the BMAD method, each task that needs to be implemented would also include information on how to implement it and where to implement it, completely eliminating hallucinations and problems, just like we saw with this game we tried to implement here.
Installing it is really easy. I have a whole video on it and I won't be able to teach you everything about the BMAD method here. So, I'll link the video down in the description, but essentially this is the BMAD method. You can see we have a Phaser 3 dev pack right here. They've engineered a specific context workflow that allows you to build these games in a proper way.
The thing I wanted to show you is that you also have the option of codec cli in here and you can just go ahead and install it. I'm going to open this up in cursor to actually show what we just installed here. An app was already built but basically what we installed was this BMAD core and in this you have your agents. Now one advantage Claude has is that it has slash commands and you can delegate your agents via that but you won't have that in codeex. In codeex, you'll essentially have to add them manually like this, but it will get the job done and you'll be able to follow the BMAD method. If you're wondering, this SM is the scrum management agent which basically takes your BRD architecture document and other things, then makes the tasks and ensures we're on track and that tasks are being followed. It's basically an entire engineering team fit into one GitHub repo. And it's really, really amazing. If you're switching to codeex or using any agent for that matter, then I highly recommend it.
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.