Transcription
Okay, so OpenAI just dropped Codex app, which is their uh UI experience around Codex CLI. For some reason, OpenAI really likes this name Codex because this is the fourth product with Codex name in it.
Now, this is designed for multitasking without an IDE. So, you can work on multiple different projects simultaneously, which is pretty neat. Also, it has built-in voice dictation. Although I highly recommend to use my voice dictation app because that works across all apps, not just this.
Also, it has built-in work tree support, which means within the same project, you can create multiple work trees and work on different features and just make PR, which is pretty neat, and it's kind of becoming a standard feature for different coding agents out there.
Now with this, I think OpenAI is trying to look for non-developer audience as well. We have seen this shift that developers move from CLI to terminals with tools like cloud code or codec cla, but now we are seeing this trend that these companies are trying to create tools which are easier to use not only for developers but for non-developers as well. Claude co-work is a good example of this, and now with the Codex app, seems like OpenAI is trying to do a very similar thing here.
You can collaborate on multiple projects simultaneously. I really like this idea of command center for your agents. This, I think, first was shown to work in projects like anti-gravity from Google or the Gemini team, where they have this agent manager which lets you work on multiple different projects simultaneously.
So in this video, we're going to go through the setup process. I will share my firsthand experience of downloading it and setting it up, and then we're going to see what type of work you can do with Codex app. I am on Mac OS, and right now it's only available for Mac OS. The Windows version is going to be coming soon. So let me download this, and we're going to install it. Okay.
So here's how it looks like when you open it. On the left-hand side, we have new thread. This is where you can start building things. Then there are different automations available. You can create a new automation. These are basically cron jobs that you can run on regular intervals. And then we have skills. So there are some really interesting skills available. I'm going to install Playwright. It's actually a pretty good set of skills. You can also include your own skills if you want. And then we have different automations. Let's go to a new thread. Here are a few examples that we can run.
So now in here, you can see that you have a local project. If you are working on a repo and you want to implement different features, you can use work trees to kind of isolate different features and different branches. And there's a cloud version as well, which I think you can basically offload the work to the cloud version of Codex CLI.
Okay, so I think I'll have to create a new project for it. Let's work in the local version, and we can enable plan mode. Now, plan mode is extremely helpful if you're just working on an idea. Let's build one of these. Now, let's run this, and we're going to figure it out. Okay.
We can also create or work on another project. So, we're going to use, I guess, like the medium version because this does seem to be a very simple app. So, let's send this in. And you can see that it's currently working on two different apps simultaneously. Now, for the first one, we enabled plan mode. So it's asking me some questions. So let's do this. All right. This seems to be inspired from the ask user question tool in Claude Code. Okay. So more questions.
Okay. So in this project, I did not ask it to use plan mode. So it went ahead and it started thinking about how to implement this PDF skill. Okay. So it came up with the plan that we can read through. Now, usually when I'm working on a project, I spend a lot of time in the plan mode and want to make sure that we get the plan right because once you have a really good plan, these systems are good enough that they can really validate the implementation. I am happy with the plan. We're going to just say implement this plan. And you can see that it exited out of the plan mode. And now it's simply going to the implementation mode.
Now we can actually go back and look at the other app that it created. Okay. So by default, it runs all in a sandbox environment. But if you want, you can give it full access on your computer. I don't think I want that. Now let's see. Okay. So I want to actually make use of some of these skills. Use the image generation and speech generation skill to create a storybook. Every storybook needs to have a storyline comprising of multiple different images, and they need to be narrated using the voice. The user will be able to provide a theme for the story. All right. Okay.
So, I'm happy with the transcription. We're going to send this in. Seems like the implementation here is complete. Can you use the Playwright skill to run this in the browser? Okay. So we're going to send that in. Usually for testing and automation, I like to use either the Playwright browser skill or Cloud Code comes with its own browser automation. And Gravity also has its own browser. So usually you're able to ask these agents to test the apps that it's building, which is pretty amazing.
Okay. So right now it's asking me for my OpenAI API key. Codex with GPT 5.2.2 is on the slower end compared to some of the other products. Even Opus 4.5 seems to be faster, but it usually works. It overall, I like the experience because you can just switch between different projects that you are working on, which is pretty great.
Okay, let's try to create an automation. Now, the way this works is that you provide a name, you provide a folder where you want the system to run, then a specific prompt, and essentially it's a cron job. So you set the time and when to run that automation, which is pretty awesome. So for non-developers, they can set different automation tasks and will be able to run them. This is kind of geared towards non-developers who can [clears throat] use these tools to do knowledge work.
Okay, so I actually failed to run the test because it's in sandbox environment. I'll have to give it full access. Okay, seems like it has completed our story book. So we're going to just ask it to run it. Let's see what happens.
Now, if you want to see where the files are located, you can open them in IDE of your choice. So let's open this project in VS Code. Here's how it looks like. All right. You can actually also provide custom instructions and settings. So, for example, if you click on this play button, we can just copy the instructions that we want to use in order to run our app. Let's see what it does. Okay. So I think we need to install the Playwright CLI.
Now I can't seem to be able to run multiple terminals. Also, you can see the diff. Although for this, I'll probably recommend using a proper IDE. Now, at the same time, we have the other app running for testing, which is pretty neat. You can switch between multiple different agents. It's really a command center for agents that you are working on.
Seems like you can also connect MCP servers. So, some of them are listed here, or you can bring in your own MCP servers, which is neat. We also looked at some of the skills. This is get work trees. Okay, so it ran that scene for us. It generated all the scenes. It looks pretty neat. Okay, it also created individual audio files now.
But kind of the overall idea of this app is that you can work on multiple different projects simultaneously. They don't have to belong to exactly the same source code, or if you are trying to build multiple features within the same source code, so you can use something like work trees. I kind of like this minimalist interface, which is pretty clean. Although I'll have to spend a lot more time to see whether this is something I would be using on a daily basis or not.
Now, one thing which I don't really like about GPT 5.2.2 is it's really slow. So hopefully these new collaboration between OpenAI and Cerebras will make GPT 5 extremely fast and more usable. >> [snorts] >> OpenAI definitely needs to hire somebody to name these apps or tools correctly. It's just becoming too confusing now. Anyways, I hope you found this video useful. Thanks for watching, and as always, see you in the next one.