Getting started with Codex

Hi, welcome to today's getting started with codec session. I'm Derek

>> and, I'm, Charlie >> and, we, work, with, customers, to, help, them get onboarded to Codeex. It's a really exciting time in the AI coding space and it's quickly changing how developers

work. So, Codeex is OpenAI's coding

work. So, Codeex is OpenAI's coding agent that developers can use across different surfaces. And we've seen

different surfaces. And we've seen firsthand how developers are delegating routine and time consuming tasks to codeex and spending more time on complex and novel challenges like design and

architecture. And we've seen a lot of

architecture. And we've seen a lot of excitement from the developer community about codecs. And our goal this session

about codecs. And our goal this session today is to give you a highle introduction to codeex.

>> So,, we're, going, to, cover, a, few, different topics here. Uh first we're going to

topics here. Uh first we're going to walk through just the introduction to Codex, what it is, kind of what the extent of the product is. Then we'll

talk about installation and setup, how to get started with Codex. We'll then

dive into agents MD, a readme format for codeex and how to configure codecs with config.toml. We'll also talk about

config.toml. We'll also talk about prompting with codecs and some of the best practices around that. We'll give

you some CLI and IDE tips and tricks talk about MCP configuration, and then dive into advanced use cases, and give you some follow-up additional resources.

So, you can use codecs in multiple clients today. So, the CLI allows for

clients today. So, the CLI allows for lightweight interaction in the terminal.

It also offers a headless SDK mode for more advanced programmatability. Um, so

Codeex's IDE extension works with any VS Codebased IDE and really allows developers to write code in a more of a rich graphical user interface. And you

can also kick off remote tasks in a cloud environment as needed. And with

Codeex's cloud environments, it really allows you to run multiple parallel tasks in the background even with your laptop closed. So it's really ideal for

laptop closed. So it's really ideal for kicking off async tasks like code review or working while on the go with your mobile phone. So for today's session

mobile phone. So for today's session we'll focus more on the CLI and the ID extension.

>> So, here, are, a, few, of, the, ways, that customers are using Codex today and how it can integrate with your team's workflows. Uh as Derek mentioned, you

workflows. Uh as Derek mentioned, you can kick off code reviews and Codex cloud whenever PR gets opened and Codex can provide comments and find critical bugs and issues before they get merged into production. You can also use it

into production. You can also use it with Slack mentions. So you can integrate codeex into slack at mention it and then it picks up the whole conversation thread that you've maybe been chatting with your team about and

then implements it and gives you a PR and you can also build your own integrations with the codeex SDK. This

allows you to run codecs programmatically in your own containers and so you can get structured output of code out of codeex and we'll cover this kind of in the the back half of the presentation here.

So, OpenAI at its core is a research company and Codex is backed by our state-of-the-art models, most recently GT5.1, Codex Max, which is our best model for aentic coding. So, it's

specifically trained in the codeci harness, and our continuous focus will be to have each model generation be more efficient, uh, be able to run faster and longer, and to generate even better

production ready code. So our models are natively trained um in Linux, Mac OS Windows environments. So it'll be very

Windows environments. So it'll be very reliable in bash and PowerShell and will adhere to the sandboxing rules. We also

train the model to be able to accurately autocompact long conversations. So you

can have codecs work for you on longer running tasks. So for example, really

running tasks. So for example, really large scale refactor tasks.

>> Yeah,, Windows, support, and, those, long ring tasks are like maybe the two of the biggest feature requests we've been getting with Codeex.

>> Definitely., And, we, also, recently published a guide that tells your team how to build an AI native engineering team. Uh it goes through how we can use

team. Uh it goes through how we can use coding agents to accelerate different parts of the software development life cycle. We broke it down into seven

cycle. We broke it down into seven phases that span planning and design all the way to documentation and maintenance. And as we go through

maintenance. And as we go through today's session, I would just encourage you to think about how codecs can help your team in these different areas. Even

though we're starting with just the basics here, all of this is applicable throughout the software life cycle.

So let's get started. Uh as a quick reminder again, the focus of today's session is the CLI and the IDE services.

Um we'll be using an open-source repository. Um so if you're new to

repository. Um so if you're new to Codex, we definitely encourage you to follow along using the same repository or just a repository of your choice. So

let's dive right in. So how do we get started? Let's start with installing

started? Let's start with installing codec cli on your local machine. So you

can install in a couple ways. You can

install using brew or npm which is the suggested method since it's going to be the most upto-date. Um our team ships very very quickly. Uh sometimes multiple times per week. So uh this way you

you're always going to be on top of the latest release. Uh when you run codec

latest release. Uh when you run codec sessions as well, there's going to be a banner on the top that reminds you if you if there's a newer version to update to. Um, if you want, you can look at the

to. Um, if you want, you can look at the latest binary releases on GitHub since Codex is open source. Codex CLI is open source and you can also look at the

change log on our developers website for more specific changes for each version.

So you can install the um codeci extension in VS Code as well. And so to do that, um, you can open VS Code, um

and click on the extensions tab here and search OpenAI codecs. And so you want to make sure you're looking at the official one made by OpenAI as seen here. And

when you download it, you'll see the version that you have. Um, and you'll be able to switch between U pre-release version and the release candidate. And

you can also toggle on auto update or not. I definitely recommend toggling on

not. I definitely recommend toggling on auto update since you want to make sure you're seeing the latest updates um for the IDE extension.

So, let's talk about actually signing in with um your ChatBT enterprise account to use Codeex. So, there's two different ways. Um and if you use either way

ways. Um and if you use either way you'll be signed in on both sides. So

on the IDE extension, when you first open it, you'll be able to see a splash screen that says sign in. And so when you click on that, it'll bring you to a sign-in page similar to how you'll see

on the Codex CLI, which I'll demonstrate now. So in the Codex CLI, all you need

now. So in the Codex CLI, all you need to do is type Codex login, which will take you to a page where you can sign in with your work email. Um, it'll probably route you through your internal SSO to

be able to log in. And once you're you've done with that, you'll be able to start using codecs uh locally on your computer on the CLI and the ID extension.

So, let me go back here. And so, once you're in codeex CLI, um I'll show you how to start running it in a second.

You'll see that there's a couple slash commands that I will note throughout this um session. The first one is slash status. This just gives you a highlevel

status. This just gives you a highlevel overview of um the model that you're using, the directory, the sandbox mode the approval policy, um how much context

window you have left. So, for example um just to go back to the CLI here, um I'm on this specific version. I'm using

this model and currently in the agents.m

MD directory. And so, what you can do is go to slashstatus and you'll see here that I'm using this specific uh set of configurations um with my account and my

session ID as well.

And so let's get started by cloning the repository that we'll be using here today. Um so we're using the agents.md

today. Um so we're using the agents.md

website which you'll see here. Um so

Charlie will talk more about agents MD and why it's very important. Um but for today we'll be cloning this repository that's open source. Um so once you clone that repository through this link below

just cd into that file, install npm and then npm rundev. Um, so you'll be able to run it locally and I will do the same here as well. mpm rundev.

>> It's, kind, of, meta., We're, using, the agents MD micro site >> while, also, talking, about, agents, MD.

>> Exactly., Exactly.

>> But, you, know, >> we, want, to, hammer, in, agents, MD, today.

>> Yeah.

>> Awesome., So, we've, set, up, agents, MD, and I'm running it locally now.

So, I'll hand it over to Charlie to to dig a little deeper on agents MD.

>> Great., All, right., So,, we, have, a, project set up and we want to work on it. Um, we

can use agents MD to help get Codeex clued in to how this project is going to work. Um, why do we need an agent MD?

work. Um, why do we need an agent MD?

Well, when you think about it, coding agents don't really retain any context between sessions. Every time you start

between sessions. Every time you start it up, it's coming in with a fresh context window. Agents MD ensures that

context window. Agents MD ensures that the instructions that you want to give the agent to just give it the background, the TLDDR on your project are always loaded automatically. Codex

will always load this file in if it's in the directory where you're starting Codex.

>> You, can, use, it, basically, as, a lightweight readme um that allows your agent to quickly under understand and get up to speed on your project and just get that context it needs to actually be

able to do its work. So there are a few different ways that you can use agents MD. You can create it with a slashinit

MD. You can create it with a slashinit command in the codex cli and this will autogenerate a uh agentsmd file for you just in the directory where you're working. You can also create an agentmd

working. You can also create an agentmd file in uh your uh codex home folder which is a global agents md. Um this is good for just kind of global configs

that you always want codeex to adhere to things like using specific tools that are you know just things that you use or different kind of ways you want the agent to act. You can also create

different agents MD files within your project and even subdirectories within your project and these are really good for giving repo specific context.

Usually we'll put one at the git root of the project that kind of is kind of the starter place for codeex because that's where we typically start up codeex. We

might also include it in subdirectories for different services that just tell Codex as it goes into that directory okay, here's what you're actually working on here. Here's the context that you need to pay attention to. Great. So

what's included in an agent MD? Common

sections that we often have in here are the project overview and the structure.

Um just giving CEX pointers on where to go for different files. Also how to build the project and running test uh commands. Doing helpful CLI tools that

commands. Doing helpful CLI tools that you might use for yourself like the GitHub CLI for example or different MCP servers that you want to integrate. And

also giving Codex a workflow just talking about how to go about implementing a feature from front to back. Lastly, you can include pointers

back. Lastly, you can include pointers to other task specific guidance in different files for the agent so that the agent can do progressive discovery as it wants to find out more about how

to do things in your project.

So, some of the best practices that we recommend here, the first is to keep it brief and focused. If you include too many instructions, it can confuse the the coding agent and it won't know exactly what to pay attention to or if

they're conflicting, it can spend a lot of time figuring out how to resolve the conflicts. Most of OpenAI's agents MD

conflicts. Most of OpenAI's agents MD files and we looked at the monor repo to figure this out were less than 100 lines. So they really don't need to be

lines. So they really don't need to be that long. The second recommendation is

that long. The second recommendation is to focus on unlocking agentic loops. So

anytime you can give the agent feedback or a way to get feedback from tools like llinters tests etc and verify its work.

It really accelerates how much the agent can do. And a good kind of intuition to

can do. And a good kind of intuition to build here is after you run codeex, if you're finding that you run commands to say say check did the code compile or maybe grab screenshots or something like

that, see if you can add those as different commands and instructions for codeex and have it just do that itself and unlock those loops.

Next, we recommend continually updating the agents MD file um as codeex encounters mistakes that maybe it makes or kind of gotchas. So if you're watching Codex do a roll out and you see

it took a really long time to derive the specific command it needed to run your test suite for example just add that command into your agents MD next time and watch how Codex will pull that into its context and then run a lot faster

and then you can also point it to task specific files and we'll talk about this in the next section but basically reference other MD files from the main agents MD file. So just to look at an

example of this pointing codeex to task specific docs, you can use uh agents MD to tell Codex where else to look. Agents

MD is always loaded into context. So it

has to be kind of general. And what you can do is just add a section here that says task specific documentation and then call out different task specific documents that are located in your uh

repository. For example, we might have

repository. For example, we might have an exec plans.md file for planning large tasks. We might have a front-end.mmd

tasks. We might have a front-end.mmd

file that's specific for doing front-end tasks. We might have an architecture.mmd

tasks. We might have an architecture.mmd

file that outlines specific architectural judgment calls that you need to make when you're doing a feature that requires large architectural changes. Codeex will read this in the

changes. Codeex will read this in the agents MD, understand that it can call on these things and its context and then if it needs to, it'll go and read the file and read it into its context. uh

which allows it again to do progressive discovery on these things and kind of gear up to the different challenges you're giving it. So here's an example of providing a planning template with

plans.mmd. You can use agents.mmd to

plans.mmd. You can use agents.mmd to

point codex to a plans.mmd file and this is basically a template that tells the agent how to build out a plan for larger multi-step kind of projects that we want to take on. Codex will generate a

markdown file using this exec plan as a template and it can update this individual plan as a living document. So

it'll go through and actually check things off a checklist uh and kind of update its progress here. Uh and this helps codeex really run large refactors do large kind of crosscodebased sweeps

or work on multi-hour project uh tasks just uh more effectively by having a place where it can track how it executes and update as it progresses. I'll note

you can also use MCP servers for this but there's nothing like just having a markdown file for the models to read and having a checked in ticket. Uh it, you know, just makes it easy for the models to keep that context up to date.

>> And, you, can, see, an, example, of, using plans.md on the opening cookbook, which

plans.md on the opening cookbook, which I've linked to here as well.

>> Agents, MD, is, really, a, cheat, sheet, at, the end of the day. And uh the Charlie the the example Charlie just mentioned with plans. MD, one of our engineers was able

plans. MD, one of our engineers was able to run a 10 plus hour refactor um with codeex and was it was successful in the end. So this is a great example of what

end. So this is a great example of what you can use Codex for today.

>> Yeah.

>> All, right., Let's, talk, about, how, to further customize codecs to the way you work in addition to agents MD which is the configuration tommo file. So codeex

CLI was written in Rust hence why it's in a tommo file. As you can see here on the right, there's a couple uh different areas that you can customize and have it uh run as default when you have when you

start a codec session. For example, you can specify the default model uh the default model reasoning effort, sandbox mode, approval policy, which I'll talk about in the subsequent slide. You can

also toggle certain features to be default on. For example, web search. Um

default on. For example, web search. Um

you can also specify different profiles.

So for example, sometimes I just want to start a codec session that defaults to uh the fastest model uh with the lowest reasoning effort to get the fastest

responses. So I created a profiles

responses. So I created a profiles called fast. So I can just do codeex-p

called fast. So I can just do codeex-p space fast and it'll create um that specific configuration for the session.

And as Charlie mentioned, you know, MCP is an area that you can also configure codecs to work with. So in this example we, have, uh, one, of the, MCPs, that, we'll cover later in the session as an example

on the right. So this is just the tip of the iceberg. As you can see on um this

the iceberg. As you can see on um this slide here, there's a lot of different configurations that you can do uh with the codec. Um and if you go to this uh

the codec. Um and if you go to this uh website, you will see a lot of different uh configurations for you to go.

>> Yeah,, one, of, my, favorites, to, set, here, is the terminal notifications. Uh, it

allows you to get like a popup when Codex finishes a task, which some people don't like. I actually like it because

don't like. I actually like it because Codex will run in the background and like ring a bell when it's done for me.

So, that's nice. But

>> a, lot, of, bells, going, off.

>> Yeah., Yeah.

>> Um,, to, go, back, to, the, last, lie, here., So,

uh, approval policy and sandbox mode are two areas. Uh, I just want to dive a

two areas. Uh, I just want to dive a little further into. The default for Codeex is to have the approval mode um on requests, which is Codeex will run until it decides it needs to ask you for

any escalated permissions. The other

default is workspace write as the sandbox mode. What that does is uh

sandbox mode. What that does is uh codeex will write to files in your current directory and not anything outside. U but as you can see here

outside. U but as you can see here there's a couple configurations that you can change with the default when you start a codec session.

Awesome. Let's turn it back over to Charlie to talk about prompting with codecs.

>> All, right., Okay,, so, similar, to, writing an agents MD file, your prompt is kind of the ultimate context which you want to give codecs, right? And so writing good prompt really determines kind of what you get out of codeex.

>> There, are, a, few, tips, here, which, we, found to be kind of universally helpful. So

the first is to use appmention uh to just point CEX to a you know file within your codebase that you want to start focusing on. Um, I use this a lot and it

focusing on. Um, I use this a lot and it doesn't necessarily have to be the file that you, you know, end up, uh, wanting to analyze or kind of like pull in. Um

but a starting point to kind of anchor codeex where it can start then traversing all the code paths is really helpful. When we see the agent kind of

helpful. When we see the agent kind of go off the rails, a lot of times it's because it starts looking in a part of the codebase where it's just not really relevant to what you're trying to do.

So, if you can anchor it with this super helpful. The next recommendation

super helpful. The next recommendation is to start with small tasks before you move on to larger ones. So, if you're new to prompting uh with a tool like Codeex,, smaller, tasks, are, going to, be

easier for you to test out and kind of review to start. And as you get more comfortable, ramp up asking Codex to do kind of more and more uh tasks and larger units of work. But as a tip here

you can actually use Codeex to help break tasks down. And it can do this by looking at your codebase. So if you ask it to do kind of just a pure research effort to go look through your codebase and figure out how it would implement a

feature and break it into small tasks it'll actually come up with very small units of work that you can then just prompt it to do and follow up on and follow along with. Um we always recommend including verification steps.

So provide codeex with the steps to check its work. For example, running tests, llinters, etc. If this isn't in your agents MD, it's good to include in your prompt. Another way to do this is

your prompt. Another way to do this is just just tell Codeex here are the requirements that you need to implement in order for this feature to be considered complete and it can use that as, kind, of a, yard, stick, to, measure, its progress as it works along. The next is

if you're debugging which is one of the top use cases of codeex just paste the full stack trace into Codex give Codex the full kind of detail there and it can use that to navigate your codebase and figure out what's generating the error.

Um super super helpful. Um, as we mentioned before, you can customize how Codex does its work by using agents.md.

And we also recommend that you try open-ended props. Uh, this can be great

open-ended props. Uh, this can be great to just brainstorm ideas for feature improvements or how to improve performance. Um, you can ask Codex kind

performance. Um, you can ask Codex kind of after you implement a feature, what would you consider building next? And

Codex is actually really good at kind of looking at your codebase and holistically considering, okay, here's what I would add. So there's more on the developers website uh on this and we encourage you to check out the prompting guide for Codex as well.

>> A, great, thing, Codex, does, as, well, is, um Codex CLI actually gives you next steps suggested next steps when you come up with when you ask it a prompt. So this

is a great way for you to continue the brainstorm conversation as you said.

>> Yeah., Yeah., So, some, uh, great, starter tasks, you know, if you were just try codeex today, explain the codebase. uh

just write a quick readme based on something that you're interested on the codebase or maybe another team's codebase um at your company. Use it for bug fixing. Uh just paste in the stack

bug fixing. Uh just paste in the stack trace and have it uh work on fixing a bug for you. Uh have it expand test coverage. So take a look at a feature

coverage. So take a look at a feature and ask it, you know, where which edge cases do we need to consider? Um are

there other things that we need to add in here to complete our test coverage?

And it can build out those tests for you. And then also doing refactors

you. And then also doing refactors across many files. So maybe we want to make a generic component that's kind of spread across multiple different implementations. Um Codex can look

implementations. Um Codex can look through and identify patterns, extract that component and then make it kind of generic across all those different interfaces. It's really good at those

interfaces. It's really good at those kind of refactoring tasks.

>> And, one, underrated, area, is, writing documentation. Engineers typically don't

documentation. Engineers typically don't like doing that. So I I love uh using codeex CLI to help me write the documentation that's needed.

>> Keeps, it, up, to, date, as, you're, as, you're building things. So let's take a look at

building things. So let's take a look at this in the CLI of how you can send a readonly prompt and change the reasoning effort uh with codecs. Um if you're in the CLI, you can fire it up using codecs

and then use /models to switch between uh medium uh low, medium, and high. Um

let me pull this up and show you what this looks like in VS Code. Um so I'm in VS Code and I'm just going to um pull open codeex in the side panel here. Um I

have a few options down at the bottom. I

can go and select if I want to work locally, which we're going to focus on today. But you can also actually run

today. But you can also actually run tasks in the cloud. Um, so if you've connected uh Codex cloud, uh you can have your uh agent spin up a cloud container, run multiple different

iterations in parallel on those containers. Uh and it's a great way to

containers. Uh and it's a great way to get different versions going. We can

also switch the different modes that Codex has access to here. So for this um I'm just going to ask it about the codebase. I'll put it into chat mode

codebase. I'll put it into chat mode but I can also put it into agent mode or agent full access mode or a custom configuration mode here. And then

lastly, I can pick the different reasoning models or the different models here and also the reasoning level uh that I want codeex to use. This will be a pretty simple question. I'll just say

you know, can can you tell me uh about this project?

You can make this more custom. Um and

it'll go through it will actually research kind of what's going on in the project. You can see it's starting its

project. You can see it's starting its chain of thought here. Um, and once uh it gets going working with this, we can see the steps that it's following. So

it'll give us a summary of what it's trying. Uh, you can actually check out

trying. Uh, you can actually check out uh kind of the files uh that it's read down here. Um, and when it's done, it'll

down here. Um, and when it's done, it'll produce a full summary for us to actually go and review here. One of the things that I love that's kind of a more advanced way of using this is actually asking it to look through the git

history and it can use git uh just like it would on the cl the command line to figure out maybe who committed to this file or what were the recent changes that you know happened in the project.

Um super super helpful way to get on board.

>> Codeex, is, a, know-it-all, historian.

>> Yeah., Yeah., Um, so, that's, sending, like, a readonly you know prompt but obviously we also want to be able to make changes.

So um here's an example uh that I've already kind of uh pre-built here. I

wanted to create a simple HTML page uh called assets html that shows how the files in this agent MD project are related. So we could generate a markdown

related. So we could generate a markdown file or we could generate a whole documentation website. So let's generate

documentation website. So let's generate a documentation website. Um just to show how I did this. If I go back here and actually um view the different uh

prompts that I have um I can see the different conversations uh that I have uh going here. Um, and if I go into this one, here's where we created the uh

simple HTML um, site, I can see that it created this assets HTML file for me. I

can uh, click to open that uh, in the side panel. Um, and I can also go and

side panel. Um, and I can also go and see kind of the summary of the chain of thought. Um, so Codex went and actually

thought. Um, so Codex went and actually implemented this as a file here. Um, if

I pull it up here and grab it, it's in assets.html.

assets.html.

Um, >> just, go, to, Chrome, and, show, it., And, if, we pull it open in Chrome here, um, I can tab over to Oh, if I tab over to Chrome

here. Let me get this going again here.

here. Let me get this going again here.

Um, >> is, it, in >> It's, right, here.

>> Oh,, there, we, go., If, I, tab, over, to, Chrome here, I can pull this open and we have this file that's already kind of generated a really nice looking like documentation page for us. uh here where

it's calling out kind of the different um you know flows in the project, what all the different files do. Again, super

simple example, but just kind of a a quick uh run through of what Codex can generate really quickly for you. Um all

right, so let's go back into uh the presentation here real quick. And um

there we go. Um so we can also send a prompt that makes uh further changes to the site. So not just kind of like

the site. So not just kind of like documenting things but like changing existing functionality. Um so that for

existing functionality. Um so that for this one I can say can you implement you know next to the explore examples view on GitHub buttons and maybe a download button and another one to uh copy the

pages markdown. Um so all I need to do

pages markdown. Um so all I need to do to do this is open up um a new conversation here. We'll just start a

conversation here. We'll just start a new chat and say uh can you implement uh some hero buttons um to download the repo

link to GitHub um and copy the markdown of the site.

Let's kick that off there. And um I'm actually going to pause this real quick.

I'm going to switch it over into agent mode so that it can go uh and implement this for me. And then let me run this again here.

So, this, will, kick, off, codeex, and, it will actually go and uh work on implementing these new buttons for us. It looks like it's exploring the files. Uh it's

reading kind of our hero file to figure out where we want to put that in there.

One of the nice things about the ID, it gives you these pills that kind of allow you to click and open up the file that CEX is referencing so you can follow along with it. um it's come up with a

planning uh uh task list here that kind of shows us how it's going to think through the different steps that it wants to uh work on. And then when it's done, it'll actually have added these buttons to the site for us. So, uh it'll

just go through there a little bit and kind of figure out exactly what it wants to uh what it wants to add. Um

>> yeah., Okay., So,, we, just, wrapped, up, the task there. And then if we go over uh to

task there. And then if we go over uh to check out agents in the browser now, you can see it's added a few different buttons here where we can download the repo, we can view it on GitHub, we can copy the markdown, etc. And Codex has

just implemented that for us based off of a simple prompt there. Um

>> so, pretty, cool, how, how, quickly, we, can just implement things with that.

>> All right,, for, some, more, CI, tips, and tricks.

>> Thanks,, Charlie., Um, let's, talk, about some uh different tips and tricks uh across this two surfaces, the CLI and the IDE. So Charlie already mentioned

the IDE. So Charlie already mentioned this briefly earlier, but you can point codecs to specific files in your codebase. Um for example, in the CLI, um

codebase. Um for example, in the CLI, um I can quickly just at mention any sort of file. So um

I can specify file and ask any question so Codex knows exactly where to look at.

Um, also within the ID extension, I can also do the same. I can pick a specific file. One thing to note here as well is

file. One thing to note here as well is you can actually bind specific keyboard shortcuts um to uh specific functionality. So here there's one that

functionality. So here there's one that you can actually bind and add context by just highlighting specific parts of the text in a file. So I bind it command shift C personally and as you can see

here I added context for the hero tsx file um these specific 10 lines by just binding the hotkey there. So just

another tip in terms of what you can do with the codeex um in the CLI and the IDE extension. So you can also use

IDE extension. So you can also use to-dos in codec IDE and this is a feature that I personally love using.

So, for example, uh if I go back to this uh ID extension, I can actually go to a specific part of the file. In this case I believe I have a todo already here um that asks Codex to do something similar

to what Charlie did, which is to implement a button to download markdown to view on GitHub. And once you have that to-do created, you can actually just click on implement with codeex, and

it'll actually fire off a task on the right here. I won't have it fully run

right here. I won't have it fully run since we've already completed this task.

But as you can see here, you can technically have a lot of different todos within um your code and actually just have codeex implement all of them for you, which is fantastic.

So, let me go back to the slides here.

So, another area that um a lot of people use codecs for is to actually prompt codeex with an image example. So in this case, if I go to the agents.mmd file

you see here there's no background color for uh these three setup commands. What

I can do is actually just take a screenshot of it and ask codeex to make the inline background code chips orange.

So uh in this case, maybe I'll just use the um ID uh CLI here since we've already used the IDE for a couple examples. In this case, I'll just paste

examples. In this case, I'll just paste in the screenshot here. Can you make the inline um chips

orange?

Great. So, Codeex is able to read that um as you can see um it'll explore the codebase and find exactly where it needs to update. Um in a few seconds here

to update. Um in a few seconds here it'll be able to actually change the background colors um of those three setup commands to be orange.

>> Super, helpful., Like, anytime, you, find yourself trying to describe like which button third from the left or whatever like Codex is supposed to modify or like you just give it a an image that's able to understand the image and exactly make

code changes based off of that.

>> As, you, can, see, here, it, added, a, couple lines took out a line here. So if I actually go back great it was able to implement according to what I asked it to do which is fantastic. So um you can

also do this in the IDE just to quickly show how that would work. You would just click on the plus and add image. As you

can see here, I can specify the screenshot.png that I have on my desktop and do the exact same thing on the ID extension.

All right. So, one other great tip is to um know how to resume a codec session.

So, within the CLI, when you quit a session, typically u people want to sometimes go back to the session to continue the conversation, especially when there's still um huge amount of

context remaining. So one thing you can

context remaining. So one thing you can do is in the codec cli interface is um just run codeex resume. And so what this will do is come back with all the

previous conversations that you had with codecs and you can actually just go specifically back to a conversation. In

this case um I went back to this specific conversation and I can just continue it. So this is just an example

continue it. So this is just an example where for example I can just ask codeex to change the chip code to green after I change it to orange. So again this is a

very very simple example of how you can resume conversations. Um and as you can

resume conversations. Um and as you can see within codec cli there is a session ID. So if you want to go back to a very

ID. So if you want to go back to a very specific session and not have to scroll through um the codeex resume UI, you can just run the specific command and it'll go back to that session. And uh Charlie

already um showed this earlier, but you can always go back within the IDE and search for a specific task and and show that uh accordingly and continue the conversation.

>> Yeah,, as, you, can, tell,, context management is just so important to getting the right results out of there.

And that session contains all of the context that you've given the model. So

I I use it a lot as kind of like just almost like a mini project container for a specific aspect of a a project that I'm working on. Maybe like I have a test session that I'm kind of working on or I have like a specific session for

implementing a front-end feature or something like that.

>> Definitely, very, very, helpful.

>> Another, great, tip, is, you, can, actually use codecs in the ID extension to generate mermaid sequence diagram. So

this will show you how you know the processes operate with one another and in what order order in the repo. So to

show this in the IDE extension um I can go back sorry too many windows and actually show you that I ran this earlier. So all I asked is provide me a

earlier. So all I asked is provide me a clean mermaid sequence diagram for this codebase and it came back with a very concise sequence diagram that I I can copy I can download. Um and this is

really great um for complex repos.

Awesome.

So one other area that we briefly touched on earlier is web search is default to be off um for codecs. And so

what you can do in the config tommo file is to have the web search request default to be true. If you don't have that in your config toml you can also just pass in a configuration flag. So in

this case you can just ask codeex with enable web search request to add something like the latest news footer uh box that gives you the current article

of uh of the next next.js um 15. And so

this is just a a quick example of how um you can pass in specific configurations when you start a codec session.

And one great area that a lot of our users use codecs for is to add specific custom commands. So in this CA case um

custom commands. So in this CA case um you can go to your codec home folder and create a prompts file folder and add it at test.mmd um file as an example. So

at test.mmd um file as an example. So

what this file will do is ask codeex to generate unit tests for any of the files that changed um and use the existing test runner and just to keep the diffs minimal. So what you can do and let me

minimal. So what you can do and let me demonstrate very quickly is you can go to your codec home folder. In this case I just opened it within VS VS Code. And

you can create, like I mentioned, a prompts folder. And in here, I added an

prompts folder. And in here, I added an AT@ test.md file. And so what that does

AT@ test.md file. And so what that does is if I go back to codec cli. Um, you

know, if you just added it, you should quit and just restart it. What you'll

see is that if you type in prompts you'll actually see the add test.md file

here. So when you click enter, it'll actually uh pull in the exact uh um instructions that you gave this uh markdown file and actually generate the unit test for for any of the files that

changed. So very very convenient. And

changed. So very very convenient. And

you know, obviously this is just one example of a custom command that you can add for codec.

So let's move on to MCP. So you know MCP is most people know here it's it's really a protocol for connecting you know models to any sort of additional tools and context. So with codecs we

support standard IO and also HTTP transport. And so here are just a couple

transport. And so here are just a couple examples of common you know MCP servers that a lot of our customers use. For

example people connect the Figma MCP server so they can generate any sort of front-end designs based on mockups there. Um the Jira um and linear MCPs

there. Um the Jira um and linear MCPs are both very popular. Uh so for example, you can see what tickets are assigned to you. You can have codecs actually create the changes code changes

for you and update the ticket to done for example. Um you can also use for

for example. Um you can also use for example context 7 to implement any thirdparty API callouts. And also the data dog MCP server is hugely popular to

diagnose any production issues as well.

But you know, as you know, there's a lot of different MCP servers out there. Um

and Codex is able to connect to these servers for um, additional context.

So, here is just one example. Uh, we

called it a cupcake MCP. This is um, a very simple example that we created just for this session. Um, so what this does is just returns um, an example of how

many cupcakes someone has ordered. Um

just a very simple example here. And to

run that command um and to add the MCP to codeex uh what I'll do here um as you can see here this is the the generating

unit tests um that I did earlier. In

this case I I'll just quit um and start it again. And I copied in this command

it again. And I copied in this command to to add this Cupcake MCP server. So

with codeex we made it very simple for you to just type in CEX MCP add and add in the um parameters. And when I click enter, um, it added a global MCP server

called Cupcake MCP. And to show that it actually did add, if I go back to the config tommo file, as you can see here if I reduce the size of this, it did add

um the um MCP server here. So you can also just directly add it in your config tomal, but we've seen a lot of our users just use the simple Codex MCP ad as an

example.

So um to use one example on how to use the cupcake MCP. So we are going to ask codeex to add a section at the bottom of

the agents MD page that just fetches uh Rachel's cupcake order. Um and very simple um I'll just use this to illustrate how um MCP connections work

within codeex. And so I'm starting

within codeex. And so I'm starting another codec session here and I'm going to paste in this request. And so as you'll see in a second, it'll be able to

call the codeex uh cupcake MCP server and actually add this uh piece of code at the bottom of the agents MD page.

>> Cool.

I'm waiting for it to pop up. It's going

to get it's going to call it. I can

tell.

>> Yeah., Yeah., Yeah., It'll, come, in, a, few few seconds here.

It's analyzing the integration options for Rachel's order. It's serious

business here. It's searching for Cupcake.

And there we go. It called Cupcake MCP search. It was able to fetch her order.

search. It was able to fetch her order.

As you can see here, Rachel ordered seven marble cupcakes for pickup. I'm

getting hungry just looking at this. Um

and what it'll do, Codex will add this to the bottom of the agents.mmd file.

>> So,, pretty, simple,, but, you, can, see, how you could extend this to, you know, for example, like production logs, you know maybe you want to pull down to diagnose a performance issue or like

>> grab, like, a, Jira, ticket, and, like, pull, in the context from that ticket for codeex to work on. So, again, super basic example to start, but I think like indicates, you know, how much you can actually do with MCP.

>> Exactly., Exactly., I, will, move, on, to, the next example as it completes this task.

So one other MCP I just want to mention that is hugely popular is contact 7 where it can pull in the most up-to-date documentation for uh any framework out

there. So what you can also do is use a

there. So what you can also do is use a similar codeex MCP add command to add context 7 as an MCP server in your

config. So in this case um I did already

config. So in this case um I did already add um context 7. So you can see here I already have context 7 present in my config tommo file and so the next

example I will use this but I do see that the cupcake example um was finished and so let me actually go back to the file here the agents MD file running

locally on my computer and I'll reload it and it should be able to show me um the cupcakes example at the bottom. Let

me scroll to the bottom here. There we

go. It looks like yes indeed. Rachel has

seven marble cupcakes uh ready for pickup. So I know what I'll be doing

pickup. So I know what I'll be doing after this session.

>> All, right., So, let's, move, on, to, this example which is using context 7. So um

this example I'm going to ask codeex to implement an input on the agents MD page running locally where people can actually generate their own agents MD file using a small prompt um that

actually calls OpenAI's responses API.

Um, and it'll actually take the latest docs from context 7 to get the latest API spec. And so, let me copy this

API spec. And so, let me copy this prompt. And let's go with the IDE

prompt. And let's go with the IDE extension for this example. And so, I'm going to keep it on agent. Uh, maybe in this case, I'll use a different model.

Uh, I'll use GT51 codeex mini. Um, I'll

use medium. And I'll ask it to implement this input for me. So what you'll see in a second, it's able to call the context

7 MCP and actually add this um change to to the agents MD page. So as you can see here, it's able to uh retrieve context

7's library docs. Um it's able to now explore um different files. It's reading

our agents MD that's present in our repository. Um and looks like it's

repository. Um and looks like it's preparing to add the agents MD generator. So um very quickly it's able

generator. So um very quickly it's able to understand what my um request is look at the docs and actually read through the the different files um within this repository.

>> And, this, is, super, helpful,, right?

Because you know models have so much pre-training context, but it kind of has like a knowledge cut off date. And so

like as programming frameworks evolve and documentation gets updated things like context 7 or really any MCP that you want to integrate a lot of customers build their own internal documentation

MCPS just helps codeex stay in sync with actual implementation that you know as it updates in real time. Definitely

definitely awesome. See, it's continuing to work. In this case, I will go back to

to work. In this case, I will go back to the slides and talk about the next um tip that I have, which is actually to add um more custom instructions to your

global agents MD that sits in your codeex home folder that Charlie mentioned earlier. So in this case um if

mentioned earlier. So in this case um if you don't want to have to ask codeex to look at context 7 every time um you can

actually add a specific input um within the codeex home folder which I'll show you here um which is to actually say when implementing new features of any

you know external libraries or APIs always search if there is relevant documentation on context 7. Um, as you can see, I have a I have a couple other lines on my uh personal Codex home

agents MD file. But what this will do is actually allow the user to not even specify context 7 and CEX will be able to read your global agents.md and

actually call the exact same um MCP server uh without you specifying it which is fantastic. So I believe um perfect timing. It looks like it's about

perfect timing. It looks like it's about to be done with all three. It looks like it was able to change three files. It

added almost 300 lines of code. Um, as

Charlie mentioned, you you can also look at the specific diff as well. Um, looks

like it's, you know, doing serious business here. Looking through some of

business here. Looking through some of the files. Um, all right. Let's go back

the files. Um, all right. Let's go back to this um this last tip that I have before this prompt finishes, which is actually to use codeex to review your

code for you. So there's a couple ways to do that within the CLI and the IDE.

Within the CLI, there's four options.

You can have Codeex review against a specific base branch. You can have Codeex review all your uncommitted code locally. You can also have Codex review

locally. You can also have Codex review specific commit. And you can also

specific commit. And you can also provide custom review instructions on top. Um, as Charlie mentioned earlier

top. Um, as Charlie mentioned earlier when we were talking about agents MD you can actually have a specific separate MD file that's not directly in

agents, MD., So,, for example,, what, I've

agents, MD., So,, for example,, what, I've seen a lot of users do is create a code review guidelines MD that's referred to within the agents.mmd file to not overflow the agents MD file with too

many lines. Um, so this is one great way

many lines. Um, so this is one great way to steer codecs to understand how to actually review your specific team's code. And within the IDE, there's also

code. And within the IDE, there's also uh the ability to um review code as I'll show here in a second. There's slash

commands here. So when you click on code review, you can also review it against a specific base branch or review your uncommitted changes. But before we do

uncommitted changes. But before we do that, let's take a look at what Codeex was able to do with adding this input.

So, let's go back to this file and see if it was able to add this responses API section. So, great. Looks like it's able

section. So, great. Looks like it's able to add a section that will actually use OpenAI's responses API to generate an agents.mmd file with your prompt. So, uh

agents.mmd file with your prompt. So, uh

just a great way to do that. So, I will quickly show how to um fire off the code review. So again, you can use codec cli

review. So again, you can use codec cli with the review command. So you can review your uncommitted changes, review a specific commit or add custom review instructions. In this case, I'll quickly

instructions. In this case, I'll quickly show how it works within the ID extension. So I can click on code review

extension. So I can click on code review here and click on review uncommitted changes and it'll be able to actually run through uh and it will come back with a couple um responses. Usually

it'll look to see if there's any P 0 or P1 issues and it'll definitely surface that. It'll also surface sometimes P2 or

that. It'll also surface sometimes P2 or P3 issues. Um, and it's just a great way

P3 issues. Um, and it's just a great way to know what you need to fix and actually ask Codeex to fix it for you.

Uh, because it does come back with um a button within the ID extension where you can just click on it and it'll help you fix the issue, which is pretty awesome.

>> Exactly., What, a, great, way, to, to, do everything in one with the codec cli and the ID extension.

>> I, think, one, of, the, most, important, things we found with code review is that it really can't be too noisy. like if it's commenting on too much, people just tune it out. And so the model has actually

it out. And so the model has actually been trained to kind of really only focus on those P 0 or like P1 issues.

And sometimes it'll find P2s or like P3s, but like I often find usually when it comments, it's like if I don't fix this, it's going to break in production.

>> Exactly., It's, defin, it's, definitely found some um code issues internally that it help fix. Um and it's just a great way to not have an expensive llinter, right? because it always comes

llinter, right? because it always comes back with something that is actually an issue that you can actually have codeex fix for you.

>> Now,, let's, hand, it, back, over, to, Charlie to talk about some of the juicy advanced use cases.

>> Yeah., So,, one, of, the, things, that, I, get really excited about with Coke CLI is that you can use it obviously interactively, but you can also use it programmatically. So you can build it

programmatically. So you can build it into different pipelines or build different use cases with it um that kind of run on your own containers and just do really whatever you want to build. Um

so to kind of show this I wanted to jump into the terminal here uh and just pull up codeex uh and um kind of show what this looks like. So in our agents MD f

uh in our agents MD site here um we have a special file that I've created. I'll

actually show this over here um called codeex output schema. If I close codex on the side here, we can see what this is. It's just a JSON file that adheres

is. It's just a JSON file that adheres to the OpenAI structured output schema format. And this tells Codex precisely

format. And this tells Codex precisely the format that we need to receive a response back in. So for this example, I was working on doing like a code style

analysis kind of checker and um I have a few different uh properties it has to include like the number of files that it analyzed, the total number of issues that it found, the overall score from 0

to 100. We can also define like an array

to 100. We can also define like an array uh so that we can have individual items for every single uh issue that it finds and each of these can have their own required fields. It's a very extensible

required fields. It's a very extensible format. Uh and what we can do is

format. Uh and what we can do is actually run codeex uh with this and do codeex exec which is codex's headless mode. Um analyze this app for code

mode. Um analyze this app for code quality issues and then give it this output schema param and just have it generate JSON. I'll fire this off here

generate JSON. I'll fire this off here and you can see rather than going into an interactive terminal mode, Codex is just streaming kind of the raw logs out which we can consume. But to kind of give you an example of what this looks

like, we can see it already ran through this here. And when it's done, it prints

this here. And when it's done, it prints out, this, valid, JSON, um that, we, parsed with jq. Um it gives you all the total

with jq. Um it gives you all the total files, the total issues, the file citations, the line numbers, the severity of the issue, the description um all in a structured format. And we

could do a ton of stuff with this. For

example, if we want to make an API request for every single issue that it finds, or we want to save each one to a database, we can build this into an app.

We can run it in our CI/CD pipeline. Um

it's a very simple but powerful kind of primitive to be able to do a lot of programmatic work with codecs. So, what

can we use to actually build with this?

Um, you can do things like security triage. You can make a test coverage

triage. You can make a test coverage bot. You can do a refactor and cleanup

bot. You can do a refactor and cleanup automation. You can do a release hygiene

automation. You can do a release hygiene automation that creates change logs and readmes every time you uh mint a new release. All these things are kind of

release. All these things are kind of things you can just build into the software development life cycle here.

And on top of that, we can actually get a bit more advanced. Instead of just doing a singlestep workflow, we can build advanced multi-step workflows with the OpenAI agents SDK and codecs. So

OpenAI's agents SDK um helps us build agent handoffs and scope agent context and tooling to individual agents and codex can actually work as an MCP server

within agents SDK. That basically makes it a tool that can be called by the agent. This is extremely powerful

agent. This is extremely powerful because we can use the agents SDK to say uh create a front-end agent that's just focused on building front end or a PM agent that's just focused on doing

project management or a backend engineer that just focuses on the back end. Give

them each their own MCP tools, their own context and tell them to hand off to one another. And because we're using the

another. And because we're using the agents SDK, we automatically get traces which you can see on the side here that shows how Codex is called or the different agents are handing off to each other before they call CEX. in the

context that they gave Codeex before they asked it to write some code. We

have a whole example in the opening act cookbook, but we think this is a really neat way to build out some of these longer uh controlled rollouts that we want to do for big refactors or big feature work.

>> The, next, one, I'll, call, out, is, on-prem coderview. So, similar to what we saw

coderview. So, similar to what we saw with the example with parsing JSON, um you can build your own coder view the same as what is built into GitHub with Codex Cloud, uh just running in your own

containers. Uh this can be super useful

containers. Uh this can be super useful if you're on prem or if you have another SCM that's not GitHub. Um you can just build this there. Um uses Codex's unique code review capabilities. You can use

the exact same prompt that we include in the CLI and Codex cloud. Um and you can use structured outputs to flag the different issues that it finds and just run uh in your own pipeline. Similar to

that, you can go beyond doing code review and actually autofix CI. So if

tests fail when you push up uh you know your your something to be your pull request, you can have codeex trigger based off of that, check out your branch, make a pull request back on top of that with a fix to make tests pass uh

and then just merge your your PR from there and save a bunch of time that way.

And then one we actually use on the Codex open source repo is triggering codeex whenever issue gets created to autolel the uh issue. This is pretty basic but super super helpful.

Basically, we can use C codeex to categorize the issues based off of a series of predefined labels. And this

doesn't need to match any kind of like specific, you know, text matching or anything like that. Codex can kind of understand the intent of the issue and use that to apply a label, which our team then uses to figure out how we're

going to tackle the issue itself.

So before we wrap up, I just want to share a few additional resources. some

for all users. So for the Codex uh documentation, go to developers.openai.com/codex.

developers.openai.com/codex.

Um lots of uh great tips there and kind of the, full, documentation, available., We

also have a series of codec cookbooks which are guides that we've put together based off of our own kind of experience onboarding customers and experience from our engineers internally at OpenAI. And

then also the Codex change log which captures all of the different features that are being added to Codeex. As we

mentioned, we're shipping pretty rapidly, so that updates pretty frequently. If you're an admin, we have

frequently. If you're an admin, we have an enterprise admin guide for you, as well as a security guide and then a rate card that captures the credits that are used by codeex so you can understand

usage. And with that, that's that's it.

usage. And with that, that's that's it.

Thanks for watching our presentation.

Um, and we hope uh you are able to use codeex and uh spin up on it. Well

>> thank, you, so, much.