OpenAI Codex CLI

Key Takeaways

So what I'm thinking is maybe I can start introducing codecs from here. Fuad as always is like coding away in the background and then he can join me for the live demos. What do you think? Sounds great. Cool. Let's do it. Take one mark. Hey everyone, I'm Roma and I work on developer experience at OpenAI. What I love most is making sure developers like you feel delighted and inspired by the tools we build. Well, today we're excited to give you a quick look at something new. We've been experimenting with the Codeex CLI. Codeex is a lightweight coding agent that can run directly from your command line. It can read and edit files. It can run commands securely and you can really use it to build features or complete apps from scratch. But enough talk to see that in action. I'm gonna bring Fouad over for some live demos. Hey. Hey. Hey. How's it going? Great. Hey everyone, I'm Philad. I'm on the agents research team and I'm so excited to share codeex with you all. Why don't you uh kick us off with maybe like a project that people might have seen before. Let's say like open.fm that we use as a quick demo lab for our voice models. Yeah, let me go and pull up the website openm I think it's an open source repo, right? So I can just go ahead and clone that locally and fire up codeex. And so here we have codeex just running on my machine. The cool part is that it runs with any of our public models. And since I'm not that familiar with the codebase, I'm just going to go and start off by asking it to explain this codebase to me. And so what I'm looking at here on the screen is that you're using 03 that we just launched today. Yeah, you can actually use anything from 4.1 from a few days ago to 03 and 04 mini today. And I think one really cool thing is as it's actually calling these tools, it's running commands. You can see it run the commands directly on your machine. So now we can see it put together this whole description. I can see that it described what open IFM is. It shows me the code architecture that it's an X.js application. And finally, here's how I can actually run it. So, I'm going to go ahead and run the development server. And um did you have a use case in mind? Yeah, I mean, why don't we do something simple like dark mode for instance? Like I know developers always love dark mode and and maybe we can do something on top of that. Sure. Let me go ahead and codeex. This time I'm going to run it in full auto mode. So, what does that mean exactly? Full auto. Now, we can also edit and run commands automatically. Got it. An important point there to make sure it still stays safe and secure is that when you run it in full auto mode, it's actually running network disabled and sandbox the directory that you ran it in. So, it's perfectly safe for you to just be able to walk away. But, it's really important for us to make sure that our users stay in control when you're actually running this on your own computer. That's great. It sounds like what we're talking it's changing all the tailwind CSS uh and and making the changes we want. Yeah. So one of the nice things is that while I got this high level overview in one step, I can also had to go in and make very specific changes without actually having the context of where it's making those changes. So I think the nice part is that you know you as a developer can actually use this to both understand the code and actually make edits to it pretty seamlessly. Now that it's done, I'm going to go ahead and open up the open FM. I can open it up locally and boom, we can see it in dark mode. That's amazing. You know, that's one example where it's an existing codebase, but maybe try something new. Well, now that we've got we've gotten familiar with with codeex with this one example, I thought why don't we create something a bit more fun uh from scratch this time. A little bit like vibe coding complete app from nothing. Sure. Uh is there one that you like in particular on Mac OS for instance? Okay. Yeah. Um you know when I was a kid I used to go to the Apple store and you know play with photo booth. I don't know I don't often do this but um so yeah let's actually pull out maybe one of the filters in photo booth. Okay. Yeah. How about how about this page of filters? Yeah. So, are you able to like screenshot the app and just tell codeex that you want it like put on the web? Exactly. So, I'm going to grab a screenshot of photo booth and I'm just going to pass it in. So, this time I'm going to put it in full auto mode and it is a first reason about the image. It's going to understand what what is it even looking at. Okay, great. It understands that it's a screenshot of Mac OS. Yeah. Um, and now I'm going to tell it reimplement this in a single page HTML. And maybe I want to say um, use the web camera API and make sure it's in landscape mode. And so now it'll just go off and given that context about the screenshot of what I wanted. You know, this could be, you know, a screenshot of photo booth. It could be a screenshot of uh, Figma design. It could be a lot of different things. I've used it in different contexts where I've drawn very low fidelity mock-ups. Um and then just kind of give in codeex uh here's this context that I want and then go off and make the changes. I don't have to give it any additional context, any additional direction. It'll just go off think for a while. You can see its chain of thought reasoning as it's thinking through the problem. Both what commands it's running and also what its thoughts are. And now finally here we see the page. Now if I go ahead and open that page in my browser, I can go ahead and see that it created this photo booth. And boom, we have Oh my god, look how cool that is. It's exactly the same. Is that amazing? That's so cool. And I'm sure you could have prompted your way to get such a result, but just like one screenshot and the model understood exactly what you wanted to build. You didn't have to open a code editor this entire time. It was just all in our terminal. And sometimes I'll kick off multiple of these in parallel and just let them, you know, one explaining the codebase to me, one making some changes and it's just pretty magical to see it all come together. That's amazing. So it was just a quick look at some of the features that we have in Codeex. We've seen how it can read and edit files directly. It can run commands very securely and you have different uh knobs in order to to pick the mode that you're the most comfortable with. But my personal favorite feature is really that one like the multimodal reasoning. That's really the true magic of those reasoning models. You just feed it a piece of paper on a sketch or something and boom, you have uh code being written for you. Do you want to share anything about like uh one more thing about this? Yeah, there's always one more thing. We're really excited to actually share codeex with you. Fully open source. You can go to our GitHub, you can see the Codex repo, you can explore it, you can actually use Codex to understand more about the repo. And we're super excited to hear what you think. Yeah. And Codex works with GPT 4.1 launched on Monday and with 03 and4 mini launching today. So we really can't wait to see what you build. Thank you. Thanks. Okay, that was very good. Cool.