Name: Getting started with the Codex app
Uploaded: 2026-03-26T12:16:12.391Z
Duration: 1 h 12 s
Description: Getting started with the Codex app

Transcript for "Getting started with the Codex app": Hi, everyone, and welcome to the Getting Started with the Codex App webinar. I'm Vibi, and I'm here with Pab. We both work on Codex at OpenAI. We're looking forward to showing you around the app, helping you get set up, and answering some questions along the way. Cool. Let's wait a few minutes while all the people get online. In the meantime, Pap, what's the feature that you're most excited about in Codex recently? There's been, so many recently that that we've shipped. I think one one I'm really excited about is the, Windows Sandbox in Codex. So Codex is actually one of the first, if not the first coding agent that has a native Windows sandbox. So that means you can actually run PowerShell on Codex. You can build Windows application. And so that's very unique. And I work with a lot of enterprise and it unlocks so many use cases for them so that I pick this one differently. And it's fully open source. Anyone can do whatever they want to do with it. It's open source, so you can like go on the Codex repo and actually like check it out. But I'm curious, like, what's yours? I'm personally quite passionate about sub agents, which we released quite recently, which allows Codex to break a task into individual, vandalizable sub tasks and spawn different sub agents which can execute on those tasks. This not only allows you to do a particular faster, but also break them into atomic units and go after it. Amazing. Should we get started? Let's get started. So we're gonna demonstrate a lot of things right after. Gonna like go through the Codex app, sub agents, a lot of interesting things. The Spark model is something I'm super excited about. We're gonna show skills, image generation, so a lot of things. But let's get started with the Codex integration. Right, so when we talk about AI coding, there's multiple different approaches to it. First and foremost is code complete, wherein you give the model a prompt and the model tries its best finish it and to give you like the most likely prediction of that particular prompt. Second is pair programming, wherein you delegate certain tasks to the model and while working on HSN or parallel tasks together with it. Last but not the least is agentic delegation, which is more autonomous, wherein you give a task and the model itself figures out how to plan for it, what specs to write, and then continue building over it over a longer duration of time. And codecs as a surface means developers where they are. Just to share some stats, since the start of the year, we've seen three times user growth across all the codec surfaces. The usage itself has increased five times over. And as today, we have over 2,000,000 weekly active users. What I love about this is like the usage is growing faster than the users. So that also means that like we're really leaning in that agent delegation and seeing people use more and more of the tool and getting more value, which is amazing. Right. And Qualix as an agent is available where you are. So it's available as a CLI, as the app that we're going to be talking about today and across IDEs via extension. On the CLI itself, you can just npm install the CLI and then get started by asking it questions about your own repositories dataset, ask it to open pull requests for you, and so on and so forth. Similarly, we've got the Codex app, which we're going to be demonstrating shortly. It allows you to run multiple tasks in parallel across multiple workspaces, projects, and so on and so forth, truly in like an agentic fashion. And last but not the least is the IDE, wherein Codex meets you where your code's at. So that could be on Versus Code, on JetBrains, on Xcode, and you can edit your code, write your code and have like a nice chat style extension views for Codex. Great way to get started. We'd love to get your experience on the Codex app, which is one of our big bets in terms of like new interface too. Pull up. So let's enroll the Codex Identity Loop. When we think about Codex, it's basically like an interface where you can send messages, right? So if we look at that diagram that we have here, the first thing is going be user input. Like you as a user, you're going send a message, right? So Codex, can we do this? Can we implement that feature? How is that application working? And then what happens? Well, you need some intelligence, right? So you're going to use a model inference, and we're going to use our API here to send that message. So when we say, hey, Codex, can you do that feature? That goes to the model intelligence and then Codex is going to think about it. And we're going to have either one or another thing, which is like an agent response or some tool calls. The agent response is going to be, okay, great. So if we want to do that feature, here's what we should do. Right? The other thing, which is interesting is like, we get started in that feature, Codex needs to understand, is it a Node. Js application? Is the rest application? Like, what are we actually talking about? And that wasn't in the initial message. So that's where tool calls come in. Codex will actually say, run this command, like list the files, what's actually in there. And that data will actually be understood by Codex to actually go into that next step. So that's something that's like really interesting in terms of like how to build the loop because we need to assemble all of that conversation history. And what's great is it then like makes sense for the whole conversation. So you can have multiple messages back and forth. And at the end, it's just like to summarize, Codex can actually do that. It also comes with some problems, right? A context window is something that we know in LLMs or something to take care about. So it's basically like the maximum number of tokens or like text that you can actually put in an LLM. So you're in Codex. And that means that you have finite contact that you cannot choose, right? And when you assemble that conversation history, that can be a problem. And so that's where compaction is something that's actually amazing that we've built at OpenAI. When you're having that conversation, we will actually compact that history. And so we keep the conversation that you had, we keep the most important information so that this context window eventually becomes virtually infinite and you can have multiple message. So Codex in a nutshell is like an interface where you can send messages, Codex will think about it, it will run command on your laptop to understand what it should do and it will do that indefinitely until it reaches, the objective of your task. And this is especially useful for really large code bases, monorepos, IO, and when you're sort of working on a project which spans across multiple workspaces and so on. So you do want your model to be really good at understanding what's important to keep from the context so far, then like continue sort of going forward, chugging forward and so on. Yeah, exactly. Cool. Let's get started. Let's get started. Hey, Vibi. I think you've prepared an application that we wanna have a look at in the Codex app. Right? Sure. Yeah. Instead of me explaining to you what this, dashboard does, let's ask Codex to, tell us what's in the repo and also, queue us up with the dev server. So Cool, so as you see, like, as soon as I put the prompt in, you can see that the model is currently thinking and it's also telling us what exactly is is happening. So right now, the model is saying, I'm checking the repo layout and the runtime and so on. And these are the tool calls that you were talking about before, which is listing these files, exploring files and, you know, reading the said files and trying to sort of figure out what is happening in the repo. And what we'll see all across is Codex is trying to figure out what this project does. We're also seeing an approval right there that you can actually prove where it's actually installing the dependency. It shows both the intelligence of Codex knowing that there's no dependencies and indeed to install it, but whose dependencies are actually you need to access the internet to download them. Right? So that's great to see that codecs. So that's like that sandbox that you talked about on Windows. Have it on all the OS. And so on your Mac, we can also see that users network access. We actually need to ask you for an approval or even like to run an application as we're seeing right now, which is great and more safe to use all of these agents. Yes. And in a scenario, wherein you're developing a project which involves, you know, running different servers, installing things and so on, do want your agent to ask you for approval and let you know what is safe and what's not. So right now what's happening that Codex has installed the entire project and is now setting up a local server. It's also using Playwright to sort of figure out what is there on the front end of the project as well. So the task is finished. Can you show us the dashboard that you built? Perfect. Like, once I open the dashboard, we can also see the summary that Codex has provided us. So Codex says this is a React 19 plus white single page sales dashboard with, you know, a bunch of information about what the dashboard has, what kind of components it has and so on and so forth. Let's open dashboard. So as you can see, it's like a sales plus dashboard where we can see revenue per month, what's in the pipeline, average deal size and so on and so forth. This by the way was also developed by Codex and we can see how the coverage is, what needs attention and so on. And this will be our sandbox for trying out different things and showcasing different Codex features throughout the webinar. That's great. Love the data there and the 57% win rate. Something I'd love to change about this application is the design. I'm not so much of a fan on this. Could you maybe like, generate a new dashboard, a new design for this? Sure. What I enjoy doing quite a bit is to just discuss how how to go about things, for a particular project with Codex. So, for us to build a better front end or or over a better design, let's ask Codex what we can do. What I'm going to do is I'm going to ask critics to build a better design, all whilst using our image and scale and we'll talk about our image and scale in just a second, but in a nutshell, what it does is it calls GPD image 1.5 and would create a bunch of ideas around how the dashboard can look like and then we can take it from there. I see it's creating a work tree. Can you explain in like maybe one sentence what the work tree is and what it enables us to do in the Codex app? That's a good point. Like typically when you're running with like a a large project, which is kit controlled, at any given point of time, you can only work on one task, right? And Worktrees allow you to have like a managed automatically synced copy, wherein you can work on different tasks at the same time. So, right now, for example, here, we're working on the design. I can just spin up another task and be like, hey, what's the, what's the overall vibe of this project and you can see that colex will automatically spin up another copy of the project itself and try and answer the question on that without really affecting what we're doing in the pilot task itself. So we could potentially start even a new one and implement some new features, right? Yes. So should we try to do that? And something I saw on the GoX app is that small icon with the microphone that I would love to use. Can I try to send a feature request? Sure. Go for it. Hey, Codex. Can you, implement a filter on the dashboard so we can actually see, the revenue per different regions so that, like, each sales leader can actually see the revenue in their own regions. Okay, cool. And so if I understand correctly, like we can see the different thread running and this will be in a different copy so I can actually like work on different features that I have in my backlog. Right? Yeah, and you can also ask questions about the project. For example, in our previous slide, we asked Codex what's the overall vibe of the project and it worked for thirty four seconds and we've got an understanding of it. And you can, just going back to the point of flow trees, you can see that on one of the threads right now, it is adding a region filter, which you just compared it to. And on the second one, it is still working on improving the overall design of the project. Cool. I'd love to go check out the image that it's eventually generating. Actually, speaking of skills and image, can you like explain what skills are, how they work in the Codex app, how you can install them? I'd be curious to know how I could actually use that skill on my actual Codex app. That's a great question. Skills are a way to give Codex superpowers. So each skill is essentially a markdown file paired with an optional script and at the time that you spin up the Codex app itself, Codex would load all the skills that it has access to, and more specifically, it would load, the metadata of each of these skills. And, you can have skills for anything that you, any repeatable workflow that you go through while developing with Connex. So it could be stuff like code review, it could be stuff like generating images, it could be, you know, referencing certain specific docs, interacting with PDFs and so on. So you can see on my screen, I have a few. So, I've got one for editing docs, I've got one for interacting with PDFs. Similarly, I have a few skills to interact with Figma, Jupyter notebooks and so on and so forth. And the best part about this is if you're using Codex, what you can also do is use the skill creator to create skills for yourself as well. And usually for each of these skills, let's look at imageN. There is an example prompt wherein you say like, hey, generate or edit images for task X and so on and so forth. And there would be some sort of guidance that you can follow that allows Codex to follow whilst it's executing that particular skill. And you can install, try, enable it, disable it and so on and so forth. Cool. Let's take that approval that we have here for that image. Yeah, in this case, for example, we were using the Imagen skill and Codex right now is asking us whether it is okay for us to, for it to run the image in, Python script. How long is it usually taking to generate an image with the GPT image generation model? Typically, takes about ten to twenty seconds depending on what kind of quality you want for the image itself. And in this case, we're going for as high quality as we can. We've talked about skills. They're great so that you can share it with teams. They're just a file, a skill. Md file that you can share with your teammates. They're a great thing to, like, for repeatable workflow that you have even on your own machine. They can actually access the internet too. Like we've seen it with some agent, so that's really great. So you've got a lot of optional folders. Script is one of them for image and documentation and assets is good for like templates and documentation. And we're seeing a lot of adoption on Skills. This is something that the industry and the community has built a lot on. I think like some of the things I'm really excited with Skills is the fact that it doesn't include everything in the context directly. It's doing that just in time at inference time. So it helps a lot on the context and the attention, everything we've talked about enrolling the agentic group. I see that image and skill generation is actually finished. Should we ask Codex to show the results? Sure. Can you please open the dev server and show what it there. Whilst Codex spins up a dev server, what's one of your favorite skill that you go back to every day while you're working? There's one that I use every day that I should share more. Codex is open source. We're gonna see that after. And like, I'm asking questions to Codex every day about do we support this? Can you help me debug the logs? And so I have a skills that that that I have a skill that can actually, like, through all of the use cases that I might have, which may be combating log copying logs, asking a a feed about a feature. And so I just use that skill and ask question about Codecs in the Codecs repo. That's like my favorite one that I use daily. What about you? Sweet. I have one, which essentially looks at the GitHub PR and debugs the CI for me. Most of the times whenever I'm building something, I couldn't be bothered to live through CI issues and so on. So what's quite nice about this particular setup is that it allows colleagues to look at the CI, capture the logs from the GitHub actions, iterate on it and it could continue iterating until it has fixed the CI itself. So I quite like that as like an autonomous sort of babysitter of my Okay, I see has finished this task. So what did Codex do exactly here? So in this case, what we asked Codex was to use the image at scale, uses GPD image 1.5 to ideate through and find different ways that the dashboard could look like. And then what it did is it created certain mocks, with mock ups and then continued to implement that, in the repo itself. And post that what we asked you to do was to take some screenshots, send us like how it looks like, as well as, put together a dev server where we can interact with it. So you can see that it gave us this nice screenshot and it also went ahead and gave us a nice mobile screenshot as well, which I'm unable to open right now, but let's look at it on our webpage itself. So you can see, from before to now the dashboard looks much more warmer and much nicer. All the information still is the same. You can still like select a date range, you can select owners, segments and so on and so forth and, you can see the overall trends still, recap and so on. That's great. And we can see that, well, you selected the knowner, but the winner rate was still 57% when we had known, which is great. So that hasn't moved from the end change. And this was great to like, able to even have Codex, like check on the changes with that screenshot. Right. And I believe you asked Codex to make certain changes, before and add like a region filter and we could see that Codex both for seven minutes and seventeen seconds and has like a reasonable output there. And just to reiterate, all of this has been worked on on WorkTrees. So what we can do is go ahead and open pull requests for all of these individual tasks. So you and your team can then review all of these as individual piecemeals and, you know, if you have to make changes, you can. You can even see on the top right, you could actually like see the GIF, you could like eventually create the branch and like do the workflow that you've just talked about and so that's like all isolated in each of the work tree that we've created. Yes and speaking of, one thing that I quite like about the Codex app is that as you look through the diff, you can ask it to make some changes. For example, in this particular setup, I can just ask it all regions should be regions. Anything. And then I can make this comment and I can just ask Codex, hey, look through my comments and fix them. And and it will it will automatically look through all the comments that I've made on the diff interactively and try and fix them. Cool. That that's actually, like, so good for a kind of view. Which I know is something, like, we wanted to talk about. So should we try to have Codex, like, do a CAD review on top of, like so that was great seeing you doing CAD review with Codex, but, like, delegating that CAD review to Codex, I guess, like, that's something that we do a lot internally. Could you maybe show how it would work here? Sure. So let's go back to our improvised design, branch and, let's ask Codex to review this. So as soon as I click on review, it will ask me certain specific questions like should it be against the base branch or should we review uncommitted changes. In this case, let's just review it against the main branch and as soon as I do it, what Codex will do is it will contextualize the diff itself alongside like all the other files that it has. So instead of just looking at the diff, it would look at all the files within the repo in context with the diff And then it will start reviewing file by file, look through all the files that have touched and come up with recommendations. Typically, it gives you three recommendations. It's like, what are some of the P zeros that you should fix immediately? What are some good to have with P one and P two? And this is something which I hear developer community as well as clients say that this is something that has helped them time and again make better code. Yeah, absolutely. And I can see like Codex launching Git command, is something I've never been like, at least as good as Codex is. So I just like to delegate all of this stuff to Codex. So, Papp, whilst this is going on, I see that I have a lightning symbol right next to gpt 5.4. Do you want to queue us up about what this is? Yeah, do believe this is actually something that we've shipped recently that has seemed like very good adoption. This is the fast mode that we have on GPT 5.4, where the tokens basically the inference is faster, I think two times, right? It is 1.5 times faster. 1.5 times faster. So bit less than two, but still faster. And that's something that people, adapt a lot because it just like make the usage of Codex faster. And when we spun like multiple thread as we've, as we've done it, we just like to have them done early. Yeah, and just to emphasize, this is the same model, exact GBD 5.4 that you get in normal mode, just 1.5 times faster so you get the exact same tokens as you would get. Okay so what should we ask to Spark about our repo? Cool! Let's use the voice dictation option and hey, Spark, can you look at the overall repo and summarize how many number of files do we have overall in the repo and, also write a song about it. Getting creative. Okay, so let's see how GoodRx is getting on with that. Yeah, and while it is working, just something to know, oh, it's done. So we have the replies and here's the song. Let's try something else. Hey, Can you translate this to French? And boom. This is so fast. The the interface even that did that itself. Yeah. Once once you get used to it, it's just it's just so hard to go back to slower models. Just to queue up what Codex Spark is, it's a smaller model hosted on three plus in conjunction with our partnership with them, and it just allows you to do so much more in like so much more faster tasks. So I wouldn't use Spark for, you know, intensive tasks which require quite a lot of reasoning, but for tasks like running tests very quickly, asking questions about what is in the repo and so on and so forth, Spark is quite good. Should we go back to the review? I can see that it's actually done now. Sure. Do you wanna queue us up on what That's good. Did. Yeah. It works for, like, two minutes. I can see that we got KeyZero p two. You mentioned that there's some, like, things that we aren't tracking or importing. So that's all great. I get I could have a look at all of that and fix it. But what I find really good with that is I can see the two findings in the composer, and you could actually say, hey. Just fix it. And it would just go ahead and and do that for me. Perfect. While it's been on it, let's just ask you to, hey, Konex, please fix these findings. Thank you. That's real agentic delegation there. Something that I love about CADREVIO and Konex, you mentioned it in the intro, SubAgents, they're quite good for CAD review. Could you maybe like show how it would work and do a CAD review also on this repo? Sure. Let's do something meta. Let's switch to 5.4 Mini, which is one of our very recent models as of this week. It's a smaller model compared to GPD 5.4, but has the same features, is multimodal and is built specifically for running, sub agents. So we just mentioned sub agents. Do you want to kick start that with, Creatum? Perfect. So just a refresher, what sub agents are is the ability for Codex to kick off child processes, which can break down a task given to it into individual parallel tasks, which are independent of each other, thereby helping you finish a task much, much faster. So, hey, can you go through all the booktrees in this particular repo and summarize what they do individually? Whilst you're at it, please spin up eight different sub agents and sort the task. So something to note is that sub agents within Codex are explicit by nature, which means that you need to prompt them, explicitly for Codex to use them. So it's been a thing for a while, look at, like, all the work trees, and I guess like as well actually, so can you create like up to eight sub agents at once? If not mistaken, we have an upper limit of maybe 35 or 40, but you can create as many as you want. It is configured via a configuration, so you can really do anything that you want in this case. Amazing. How are SubAgents like communicating and working altogether? So the way it works is that GPT 5.4 or whichever is the model that you've selected, what it would do is it would queue up individual sub agents with a prompt and the prompt would be, hey, like do X for like these these these files and so on and so forth. And then these sub agents would work in parallel and provide whatever it is that they have understood from their task, execute those tasks and then get back to the main agent And so this way you don't really pollute the context of your main agent and the agent delegates a task and once the task is completed, it gets the output of the task and then can continue going about its chores. Something to note here, Pat, is that once I asked Codex as to what it should do, it of course spun up these sub agents that you can see. These are all individual agents which we can also see over here we can actually go into each and every sub agent and see what they're doing, and what kind of work trees in this particular case they're looking at. I just love the name of those agents, girl, niche, Sartre, which I will take as a French speaking. Yes. And they will continue working on it and they would then send their output back. And something to note is that you see this like task one out of three over here. This is powered by Codex's ply node. This allows Codex to create a plan of what exactly is it that it wants to do. So in this case it was like, it's going to enumerate all the repo work trees, it's going to look at their paths, pass those paths over to the sub agents and then, you know, collect all those results and give me a summary of what's happening in all of those. Great. Okay, so I can see like all the agents who are now finished. It checked all the eight work trees, eight inspectors put up with all of the pages we can see on the side. And if you scroll a bit, so what's the summary? We have some UI design paths and some data filter paths, which we've asked. That's like the test that we've asked before. So that's great. We got our review and summarization of, all the work fees from from, Codex and its sub agents. Something that I'm interested about is, like, we've created all those work trees, like, lot of features. How would you think about, like, brainstorming with Codex to create a new feature? That is a brilliant question. So for specifically for brainstorming and planning things, Codex comes comes with an inbuilt plan mode. So let's go to gpt 5.4 and the way you can can activate it is by clicking on the plus sign and activating plan mode. So let's give it a task. Hey, can you create a plan for migrating our entire stack to Next. Js? And what Connex is going to do is, it's going to think about all the things that you need to do to be able to migrate over to Next. Js. And as it goes through this entire process of figuring out what's the best path for this migration, it's also going to ask me questions. And these questions would not only help it make a better plan, but also need to get a better understanding of the whole process. I can see that Kuric asked you a question. Right, so this is in conjunction with our Next. Js migration and it's asking me whether or not it should do a server first approuter. And what I like about Connects in this particular case is that it would also give a recommendation to me. So, in scenarios when I am not as comfortable with a particular stack or I'm not sure whether or not I should follow a particular path X or Y, it would give me a recommendation. In this case, I'm just going to go with Codex's recommendation and that's it. And you know, you can see that as it asks questions it would continue refining its particular path and plan and continue forward. I love this feature because this is something that was like asked a lot by the community that we work with and when we brought that we saw that it was like widely used. That's a great feature to having Codex. Speaking of like feedback in Codex, we've talked about it, like Codex is open source. Should we like try to see what we can do in Codex knowing that it's actually open source? Perfect. So let me let me go into the into the Codex workspace on my Codex app. And what would you like me to ask Codex? Let's maybe ask, like, the latest features that that we've shipped recently. Perfect. Hey, Codex, can you find the most recent features that have been added in the last two days? And what I know about this is that, because, Codex is fully open source and the entire team works on the open source repo itself, all it has to do is to just look through the, git commits and give a summary of that. That's it. Yeah, that's so good. That's actually like some of the things that I use a lot, as I was saying on skills where you can actually automate like all of that and I use it like, more regularly. Perfect. Cool. I can see that Codex is finished. So it's like, I checked the the Git log, which is great. You can look at the Git history as you were mentioning. We've got some, like, plugins that is that are landing in Codex. I can see that we're also, like, getting a new command slash title, or some, great things about the Python SDK. So a lot of new things are shipping on Codex. And again, what's like, what I love about this is it's open source, and so we can use it, but you can also, use it too. That's also, like, something that I think is generally so great for everyone using Codex. That's really good to have in automation. Could you like maybe show how you would like you create an automation with Codex for that? Sure. The best part about automations within Codex is that you can just ask Codex to create automations, right? So let's create an automation. Hey Codex, can you actually create this as an automation which runs every day at 9AM. And what Codex is now going to do is it's going to contextualize everything that I've asked it and try and figure out what's the best way to create an automation for this as well. And automations show up in this automations pane. As you can see, I have a few automations of my own, which would look through different information sources like, slack, emails, and also our developer docs and would give me a summary every day at eight or 9AM, and so on. So you can see that over here, Codex gave me this nice pop up, which is a daily feature digest. If I click on create, you can see it's, it's pre filled with prompt which is review the current workspace and summarize all the recent features. I can choose which folder or project do I want to run it in, at what time do I want it to run it every day and a couple other things like what model do I want to use and its associated reasoning effort as well and that's it, I can click on save and it would show up within my automations. Okay. So Vivi, we've seen we've seen a lot of things during the demonstration. We've used the Codex app. We've created, like, tons of work trees. We've used SubAgent, the support model that's, like, super fast. Plan mode was great for brainstorming features. We also use skills and image generation for the front end edits that we've seen. And we just ended on Codex being open source, which is great because you can ask a lot of features, more than that, you can actually automate all of those tasks with the automations in Codex. How would you recommend people to get started with Codecs? I think the simplest thing to do is to just try the app itself. The app is available both on macOS as well as on Windows. So it is available where you work at, along with all the other services we spoke about, the CLI, the extension. Just log in with your ChartGPT account. If you don't have one or your enterprise does not allow it, reach out to us and we'll sort you out. Last but not the least, like, explore our documentation, something which I'm quite proud of because I contribute to the documentation itself. There's a bunch of use cases there, guides there, you can experiment with it and yeah, just build things. Cool. Thank you. Thank you.