Big Idea

The argument in one line.

Codex becomes a working AI operating system once you understand 18 building blocks, projects, AGENTS.md, the agent loop, goals, environments, models, skills, plugins, and automation, not by learning to code.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You've used ChatGPT or Claude in the browser and want to understand what an agentic coding tool like Codex actually adds beyond chat.
You run a solo business or content operation and want one AI system that holds your rules, files, and history instead of re-explaining context every session.
You're curious about Codex but intimidated by terms like agents.md, work trees, or sub-agents and want them defined in plain language.
You want a practical map of what each Codex setting (model, effort, permissions) actually controls before you start paying for a subscription.

SKIP IF…

You're a working developer who already uses Codex or Claude Code daily, this is a beginner-level glossary, not a workflow deep dive.
You're looking for a step-by-step build tutorial, this video explains concepts using Nate's existing setup rather than building one from scratch on screen.

TL;DR

The full version, fast.

Codex is an agentic AI tool that turns into a working operating system once a few pieces are in place: a project folder, a rules file called AGENTS.md, and an understanding of the agent loop, the think-act-respond cycle that lets it use tools on its own. From there the video walks through environments (local vs. cloud, work trees), control settings (models, tokens, effort, permissions), and scaling tools (skills, plugins, the browser, sites, sub-agents, scheduled tasks, voice mode). The through-line is that Codex gets smarter the more context and structure you give it, not through more complicated prompts, but through folders, rules files, and reusable skills it can reference every time you open a new chat.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 03:58

01 · Projects and AGENTS.md

Defines a Codex 'project' as just a folder of files and rules, shows Nate's own Herc 2 project, and introduces AGENTS.md as the standing rules file Codex reads before every message.

03:58 – 09:54

02 · Agent Loop and Goals

Explains the think-tool-respond agent loop with a live example (pulling YouTube comments), then covers the slash goal command for open-ended, self-retrying objectives.

09:54 – 13:25

03 · Local, Cloud, and Remote

Distinguishes local files/localhost from cloud-hosted work, and shows controlling a local Codex session remotely from a phone.

13:25 – 16:10

04 · Work Trees and .codex Settings

Covers work trees for isolated testing and the .codex folder's project-scope vs. user-scope settings.

16:10 – 19:18

05 · Models, Tokens, and Usage

Walks through the Astra/Sol/Terra/Luna model tiers, their API pricing, how subscription usage limits relate to that pricing, and what a token actually is.

19:18 – 21:38

06 · Effort and Permissions

Explains the effort slider (low to ultra) as a separate cost/quality dial from model choice, fast mode, and the three permission levels up to full access.

21:38 – 24:58

07 · Skills and .agents

Defines a skill as a reusable, improvable recipe, shows Nate's actual skill files, and covers the .agents folder that stores them at project or user scope.

24:58 – 27:05

08 · Plugins and Browser Use

Shows the plugin library for connecting external apps and demonstrates Codex's browser mode logging into and operating a website with no plugin available.

27:05 – 28:18

09 · Hosting With Sites

Shows the Sites feature for turning a Codex-built project into a shareable, permissioned live URL with analytics and optional database backing.

28:18 – 31:41

10 · Sub-agents and Scheduled Tasks

Demonstrates a main session delegating parallel research to multiple sub-agents on cheaper models, then covers scheduled tasks as timer-triggered prompts into a normal session.

31:41 – 34:19

11 · Voice Mode

Shows controlling Codex entirely by voice, starting a thread, generating a thumbnail, and routing the result to another thread, from a phone.

34:19 – 34:33

12 · Final Thoughts

Sign-off and pointer to a Codex content playlist.

Atomic Insights

Lines worth screenshotting.

A Codex project is just a folder of files and rules, the same folder that lives in your file explorer, so there's nothing mystical about giving an AI a 'memory' of your business.
AGENTS.md is read before every single message in a project, so anything you want the AI to always know about you belongs there, not repeated in each chat.
The agent loop, think, use a tool, respond, think again, is what separates an agentic tool like Codex from a plain chatbot that only answers once per message.
An objective goal ('pull in 257 sources and write the report') produces a more effective agent run than an emotional one ('until you feel good about it'), because the agent can prove it's actually done.
'Local' doesn't mean limited: as long as the machine stays on, a local Codex session is reachable from a phone over remote access, so local and cloud blur in daily use.
A localhost URL (127.0.0.1:8008) only works on the machine that generated it, which is why screenshots of 'look what I built' sites with a localhost link in them can't actually be opened by anyone else.
Work trees create a disposable copy of a project to test risky changes in, without touching the main files, but Nate says he almost never uses them for non-coding knowledge work.
Codex settings split into two scopes: user-level (global, under your account, holds memories and sessions) and project-level (specific to one folder, holds project rules and custom skills).
Model pricing scales roughly in a straight line: in this video's on-screen pricing table, Sol costs about 40% of Astra's input/output rate, Terra about 20%, and Luna about 2%, so picking the cheapest model that's good enough for the task matters.
A $200/month Codex subscription was shown delivering roughly $14,000 worth of API-equivalent inference if the weekly usage limit is maxed every week, subscription pricing is far cheaper per token than paying the API rate directly.
Effort level (low to ultra) is a separate dial from model choice, and using a high-effort frontier model for a simple task like writing an email is paying for intelligence the task doesn't need.
Fast mode trades 1.5x speed for faster consumption of your weekly usage limit, it's a cost lever, not a free upgrade.
A skill is a written-down, reusable recipe, feedback after a run ('cook it 30 seconds less') updates the skill file so the next run is better, the same way you'd refine an actual recipe.
Skills can reference other skills, other agents, and external style or context files, so a single invoked skill can chain several specialized instructions together automatically.
Plugins let Codex connect directly to tools like Gmail, ClickUp, Google Drive, and GitHub so it doesn't need a developer-built API integration or a .env file of credentials for every app.
Codex's browser mode lets it log into and operate web apps that have no plugin or API, by reusing saved logins the same way a human session would.
A Codex 'site' can host something built in a chat session on the open web with sharing permissions and analytics, functioning as a lightweight alternative to spinning up hosting on Vercel or Squarespace for a quick internal tool.
Sub-agents let a single expensive, high-intelligence model session delegate research or parallel tasks out to cheaper models, so the main thread doesn't burn its best model's budget on grunt work.
A scheduled task is just a prompt injected into a normal Codex session on a timer, it still uses the same files, skills, and model as a manual chat, it isn't a separate automation system.
Voice mode controls the entire system, starting threads, delegating sub-agents, queueing scheduled tasks, by voice from a phone, so none of the other 17 concepts require sitting at a keyboard to use.

Takeaway

Codex is 18 small pieces, not one big black box

WHAT TO LEARN

Every advanced Codex behavior traces back to a small set of plain-language pieces, folders, a rules file, a loop, and a goal, that combine, so understanding the pieces individually makes the whole system legible.

01Projects and AGENTS.md

A project is just a folder of files and rules, there's no special 'AI project' format, which means anyone who can organize folders can organize an AI's context.
A rules file like AGENTS.md is read before every single message, so standing instructions belong there once instead of being retyped into every new chat.

02Agent Loop and Goals

The agent loop, think, use a tool, respond, repeat, is what makes a tool 'agentic' instead of a simple back-and-forth chatbot, and watching that loop run is the fastest way to learn how an agent reasons.
Giving an agent an objective, provable goal and then stepping back produces better results than trying to micromanage every step, because current models retry and self-correct until the goal is actually met.

03Local, Cloud, and Remote

A 'local' AI session isn't limited to one device, if the machine stays running, the session is reachable remotely, which collapses the usual local-vs-cloud tradeoff for solo users.
A localhost link only works on the machine that generated it, so sharing one with someone else is a dead end, not a working demo.

04Work Trees and .codex Settings

A work tree is a disposable copy of a project for testing risky changes without touching the main files, useful mainly for software work rather than everyday knowledge work.
Settings and skills exist at two scopes, global (applies everywhere) and project-specific (applies to one folder), and picking the right scope avoids duplicating rules across every project.

05Models, Tokens, and Usage

Model pricing scales roughly in a straight line across tiers, so matching task complexity to model cost beats defaulting to the most expensive option for everything.
A flat-rate subscription can deliver far more inference than its dollar cost would buy at API rates, which changes how 'expensive' a top-tier model actually feels day to day.

06Effort and Permissions

Model and effort are two separate dials: a cheaper model or a lower effort setting is often enough for routine tasks, reserving the most expensive, highest-effort combination for work that actually needs it.
Permission levels (ask-always, ask-if-unsafe, full access) are a trust dial you set deliberately based on how much autonomy a given task warrants, not a fixed setting you leave alone.

07Skills and .agents

A skill is a written, reusable instruction set that improves with feedback over repeated use, the same discipline as refining a recipe, which turns one good result into a repeatable one.
Skills can chain together by referencing other skills, agents, or style files, so one invoked skill can pull in several specialized instructions at once.

08Plugins and Browser Use

Plugins remove the need for custom API integrations to connect an agent to everyday tools like Gmail or GitHub.
Browser mode lets an agent log into and operate any site that has no plugin or API, using saved logins the same way a human session would.

09Hosting With Sites

A built project can be turned into a shareable, permissioned live URL directly from the same tool, without a separate hosting step.

10Sub-agents and Scheduled Tasks

Sub-agents let a single expensive, capable model delegate parallel grunt work to cheaper models, which is a cost and speed lever most people don't use until they see it modeled.
A scheduled task is nothing more than a timed prompt into an ordinary session using the same files and skills already set up, it isn't a separate automation system to learn.

11Voice Mode

Voice mode operates every other concept in the system, projects, skills, sub-agents, scheduling, which means none of this requires being at a keyboard to use day to day.

Glossary

Terms worth knowing.

AGENTS.md: A plain markdown file inside a project that lists standing rules, context, and a routing map so the AI knows how to behave and where to find things every time a new chat starts.
Agent loop: The repeating think, use-a-tool, respond cycle that an agentic AI harness runs through to complete a task, as opposed to a single-shot chatbot answer.
Agent harness: The application wrapped around an AI model that gives it tools, memory, and an agent loop, examples include Codex, Claude Code, and other coding/automation assistants.
Slash goal: A Codex command that sets an open-ended objective for the agent to keep working toward, retrying approaches, until the goal is met rather than stopping after one response.
Work tree: A duplicate copy of a project's files that lets you test risky changes in isolation, then merge the results back into the main project folder later.
.codex folder: A hidden settings folder, either project-level or user-level (global), that stores configuration like config.toml, memories, sessions, and automations.
.agents folder: A hidden folder, project- or user-scoped, that stores reusable skill files the agent can invoke by name or natural language.
Skill: A written, reusable recipe (a markdown file with metadata plus instructions) that teaches the agent to perform a specific repeat task the same way every time, refined over use with feedback.
Plugin: A pre-built connector inside Codex that lets the agent log into and use an external app (Gmail, ClickUp, GitHub, etc.) without a custom API integration.
Sub-agent: A secondary Codex agent spun up by a main session to handle a delegated piece of work, often on a cheaper model, in parallel with the main thread.
Scheduled task: A prompt that Codex automatically injects into a session on a timer (hourly, daily, custom), running with the same files, skills, and model as a manual chat.
Token: The unit AI models bill by, roughly four characters or three-quarters of a word; model pricing is quoted per one million input or output tokens.
Full access / YOLO mode: A Codex permission setting where the agent acts without pausing to ask approval for individual actions, as opposed to modes that ask before editing files or using the internet.

Resources

Things they pointed at.

17:00toolGPT-6 Astra / GPT-5.6 Sol / Terra / Luna (Codex model tiers)

00:00productAI Automation Society (Skool community) ↗

00:00productAI Automation Society free resources (Skool) ↗

00:00toolGlaido voice-to-text ↗

00:00toolHostinger VPS (Claude Code hosting) ↗

25:29toolCodex plugin library (Gmail, ClickUp, Clay, Google Drive, GitHub, Canva, Composio)

27:37toolCodex Sites (hosting feature)

18:22toolOpenAI tokenizer

14:10channelschool.com (Skool platform)

Quotables

Lines you could clip.

01:52

“So this is no longer just an AI chatbot. This is now a co-founder. This is a personal assistant.”

crisp reframe of what an agentic tool actually is, strong hook line→ TikTok hook↗ Tweet quote

07:02

“They are going to be so goal-oriented that they're going to keep retrying different things and they're going to just basically keep going until they're able to succeed and hit that goal for you.”

concrete description of agent persistence that's easy to visualize→ IG reel cold open↗ Tweet quote

17:49

“I'm getting about $14,000 worth of inference every month, and I'm only paying 200 bucks for that.”

specific, surprising number that reframes subscription pricing→ newsletter pull-quote↗ Tweet quote

21:54

“A skill is basically just a recipe... if you ever burnt these pancakes, you make a quick update in this recipe... and the next time you run the skill, you basically see, okay, are these better?”

plain-English analogy that explains an otherwise technical concept instantly→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

analogystory

Today, I'm going over the 18 core codex concepts that you actually need to know in order to start using it right away and getting value from it. It doesn't matter if you're not technical at all or you've never used Codex before, by the end of this video, you'll understand exactly how this thing actually works so that you can start using it right away.

So let's not waste any time and just get straight into this one. All right, so I've split these 18 core concepts into four different parts and they get cooler as we continue to go on. So part one, we're gonna start off here is foundations.

So we're gonna kick off with concept number one, which is projects. Now, a lot of you guys, I'm assuming, have used ChatGPT or Cloud in the web before where you can talk to an LLM and it can give you an answer. But the problem you face is that every time you're talking to it or you start...

new chat, you have to kind of familiarize it with you and your business and what you're working on that week or that quarter. So with projects, you can basically keep everything organized into one project so that as you keep working inside of it. it only gets smarter about knowing who you are.

So right here on the left -hand side, you can see that I've got a few projects going on. I've got Herc 2, I've got Hyperframes, I've got Trading Challenge, I've got AIS Demo. And I typically am always just working inside of my Herc 2 project.

My Herc 2 project is basically what I call my AI operating system. And what a project is, is actually very simple. It is just a collection of folders and files.

So right here, this is inside of my File Explorer on my computer, and it's called Herc 2. And this is exactly the project that I have open in Codex when I say right here, I'm inside of... Codex, right?

This is the file path that I'm working in, which is, if I open this up, my Herc 2 project. So what that means is this project has rules about me, it has credentials, it has all my projects, it has all my YouTube videos. This thing has so much knowledge about me.

So now when I come into the chat and I say, hey, can you just give me a rundown of what we've been working on? this past month. This thing basically gets familiarized with my project and then it looks at everything that we've been working on together.

It even looks through all of my previous chats that I've had with Codex. So here you can see that it looked back from the past month and it said, here's what we've worked on. We've been improving your HERC 2 and your AIS OS.

We've been working on YouTube content and strategy, AI model testing and demos, apps and product experiments, websites, branding, business operations. So this is no longer just an AI chatbot. This is now a co -founder.

This is a personal assistant. Brings us on really nicely into concept number two, which is something called an agents .md. And don't let that .md intimidate you here.

If I go over to the right -hand side and I open up my files, you'll notice if I scroll down, I have something right here called an agents .md. And when I open this up, it's literally just a markdown file. That's what the .md means.

And it's just rules. Like this is a very simple thing that you all can read and it's not technical at all. So this is basically like the rules for your project.

So you remember inside of my Herc 2 project, I have this agents .md. Well, if I was inside of my Hyperframes editor project, there would be a different agents .md. So this thing basically just sets the ground rules for the project that you're working in.

So right here, you can see that this one says, hey, you are Nate Herc's AI operating system. Your job is to help him spend less time on operations so he can focus on making YouTube videos. And then I give it things like core operating rules for how to work with me.

I give it some stuff about how to be safe on my desktop. And then the real bulk of my agents .md is what I call routing. So this is a routing map.

It basically means, okay, so yeah, Nate's project is huge. There are hundreds, probably thousands of files and folders inside of this project. So let me just explain to you how this actually works.

If you need business stuff, you go to the wiki. If you need corporate structure, you go to the corporate structure section inside the wiki. If you need his voice, you go here.

If you need course knowledge, you go here. If you want active projects, you go here. So everything that I'm doing.

We're routing it back somewhere so that my agents .md file tells this project where everything is so that every time I open up a new chat, like I said, I don't have to re -explain things. And basically the way that this works, if we picture this being a chat thread, you know, you come over here and you say, hey, you say, hey Codex, I need you to help me do this.

And before Codex actually even reads this message, what it does is it first reads the agents .md. So I'm just gonna put A here, and that stands for agents .md. But Codex will basically read this, and then it will read your message, and then it will respond with something like, hey, how can I help you today?

So the kind of stuff that you wanna put inside your agents .md is the type of stuff that you want Codex to know about you every single time you're talking to it in that project. And now moving on to concept number three, we have something called the agent loop. So you'll hear about something like Codex or Claude Code or Hermes Agent.

These are all called agent harnesses. And these harnesses have their own agent loops. And real quick, if you're wondering about Codex versus like...

work, ChatGPT work, I always use Codex. The majority of my work is not building products or building software, but I only use Codex because it's just the most powerful. And once you get through this video, you'll realize how easy it is to use.

But anyways, the reason why this is called a harness is because they have this thing called the agentic loop, which at the highest level basically just means when you ask the agent a question, it has a bunch of tools at its disposal and it basically thinks and it reasons and it uses tools and then it responds to you. And then it thinks and then it reasons and it takes action and it uses tools, responds to you.

And that's basically the agent loop. And what's really awesome about that is you can see it in action. So right here you can see that this set worked for 25 seconds.

And if I open this up we can basically see what it did, right? So like it hid this response and it gave us our output. but it had a reasoning loop inside.

It first said, okay, I'll check our recent tasks and project nodes, then I'll pull together the main themes from the past month. And when it did that, it ran all these commands. It ran this, which you don't have to know what this means, but basically what this did is it looked inside something right here called a memory .md.

So very similar to the agents .md, but this is about memories rather than just like rules. It then looked at other chats. And so we can basically watch it do things.

So to show you guys a real example of this, hey, could you just... check my most recent YouTube video that I uploaded and give me three comments that you thought were funny from that video. Now as I shoot this off, what I want you guys to pay attention to is what we're watching.

So first it said thinking, now it says I'll find your latest upload, read through the comments and pick three that made me laugh. And now we can see what it's doing. So it's looking through the memory, it's looking at these things, it's running these commands.

And these commands are basically, right here, it's reading files to get established and to get oriented with what it needs to do. And so you can actually learn a lot about how Codex works and how to best work with it by when you ask it a question, you just kind of watch what it does. So you can see what it had to do was it had to run a command in order to actually go talk to YouTube.

look at the data and pull it back and now we can see three comments we have this one never seen anyone so excited to lose money trading with ai is very easy i just prompt claude only take profitable trades and if i want to lose money i can do it myself faster and better and what you'll notice is how quick that all happened now i do have my astra on fast mode but i will talk about that later because that is a core concept that's coming up all right and moving on to the last concept of part one it's a slash goal and this is honestly one of my favorite things about AI ever.

It's the ability to do something called a slash goal. And right here, you can see that that says set a goal to keep pursuing. Now, AI agents are super goal oriented, which is super cool.

So let's say you have Mr. AI agent right here. And you basically say, hey, here is your goal.

You know, this is what you need to do. And you give it this objective goal. What happens is it will basically just keep working and keep working and keep working until it actually hits that goal.

And it's cool because even if you kind of like give them a roadblock, they are going to be so goal -oriented that they're going to keep retrying different things and they're going to just basically keep going until they're able to succeed and hit that goal for you. Now, obviously, the more objective a goal, the easier it is to prove that.

So if you say, hey, can you work on this until you've pulled in 257 sources and then write the report? That's an objective goal, right? 257 sources and written report.

But you can also set goals that are a little bit more emotional. Like your goal could maybe be more like, you know, until you're fully satisfied or until you've verified this over and over and you feel good about it.

And in that case, you're still going to get there and it will still decide when it's done. But as you get more objective with your goal, the goal prompts are just more effective. But let me show you one goal that I actually did that you guys have already been witnessing in this specific example is I went to my Hyperframes project and Hyperframes basically lets your coding agents build HTML and then render it as video.

Essentially, it creates video. But what I did here is I shot it off a slash goal prompt. You can see this one says sent as goal.

So basically what I did is I said, I need you to help me create some motion graphics for a YouTube video. So all the motion graphics you've seen so far were from this exact prompt right here. I'm going over 18 different codex concepts.

I want an intro. I want transition cards. I want all this.

So I basically defined that I wanted one intro scene and that I wanted 18 cards. And then you can see that it said I finished and I visually checked, meaning it actually took screenshots to make sure everything looked good before it decided to be done. And now we have all of these animations.

But what's cool about the goal prompts is that models nowadays are getting so good, the AI models are getting so good, that it's actually better to give them a goal and then step out of their way. If you really just try to like micromanage them, you're kind of just like muzzling their capabilities. So if the situation isn't super, super risky, give them a goal and you can get out of the way.

Now there's one quick thing that I felt the need to address, which is basically the question, but Nate, what if we've been using cloud code for a long time? Well, I'm really glad that you asked. It's actually really simple.

Really the main difference is the agents .md, the .codecs, and the .agents. Because in Cloud Code, you basically have like your cloud .md and you have your .cloud. So when I first started switching over from Cloud Code to Codex, all I did was I said, hey, Codex, take a look at this HERC 2 project.

I've been building this up using Cloud Code for the past couple months. I need you to help me get this Codex ready. And basically what that means is I made a copy of my cloud .md and called it agents .md.

So I have two of those now, one agents .md, one cloud, but they're basically the same file. And then I made a copy of all my cloud skills and put them in the .agents folder. And I made a copy of all of my like...

cloud settings files and I put them in my .agents. And the cool thing is you don't have to do any of that manually. You'd literally just say, hey Codex, look at the documentation, analyze my project and make this Codex accessible, make it Codex ready.

So anyways, if you were feeling a little bit of doubt about all that kind of stuff, that's the way I did it. And now you have a bunch of local files and folders that you can use on any agent harness whenever. It's very flexible.

It's very tool agnostic. And that's what you want to be building at the end of the day. So anyways.

Had to adjust that real quick. Let's get back to the video. All right, awesome.

So let's move on to part two of these core concepts, which is environments. So the first piece of environments is number five, which is local versus cloud. So there's a couple of things to talk about here.

The first thing is if someone says, oh. is this running locally or are these local files? All that means is, does that exist only on the machine that you're currently using?

Whether you're on a MacBook right now or you're on a desktop PC like I am right now, when you say local, it just means it's only accessible by you. So right here, if I open up my files, these are basically local files. Now they're syncing with OneDrive when I want them to, so that would kind of push them into the cloud, but otherwise it's completely local.

So like if you're working on a Word doc on your computer and you forgot to save it and then email it to yourself so you could work on it on a different laptop, that was a problem because it was locally on that other machine. But something like Google Drive is always in the cloud, obviously.

But this also goes beyond just files. This goes to things that are produced. So right here you can see that it actually created me this little site for all these motion graphics.

So if I open up this site, you can see that I can view all of these videos, I can click into the different ones, and it threw this together for me really nice, and it served this to me on a local host. So right here, you can see that this URL is 127 .0 .0 .1 and the port of this local host is 8008. Now, realistically, that's not super important.

You don't really need to understand what exactly this means, but this is just a local host. But this is a little confusing because it looks like this is a web address that if I emailed to you, you could open this up and have the site. But what would happen is this would show on your screen or it would show on your computer as nothing actually lives here.

So if I copied this and I pasted this into a new browser but I changed the number to seven, I don't think I have anything running on port 8087 on my local machine, even though you might. So nothing actually loads up. And it's quite funny, I remember seeing these tweets where, I think it was a joke, but it was like someone was pretending they were a beginner and they were like, wow, Cloud Code and Codex, it's so cool, look what I built in one day.

And then they attached a bunch of local host addresses. And obviously no one else could open that up because they're not on that local device. So anyways, that is the difference between local and cloud.

Now, the other piece here is that when you're spinning up these new chats, you can kind of have these be either local, meaning right here, you know, I'm working in my local files and folders. And that's great because Codex can actually like edit these files, right? It can delete things.

It can make new ones. It can edit them. It can move them around.

It can go to my downloads. It can go to my desktop. It can do anything on my computer, really.

Or you could actually have these work in the cloud. And what you'll notice here is when you choose cloud, you... no longer have the ability down here to actually choose the model you were using and the permissions because this sets up sort of like a cloud sandbox environment.

Now, I'll be honest, I hardly ever do this because when I'm working locally, as long as I keep my machine on, as long as I keep my PC on, it feels pretty cloud -like. And here's what I mean by that. Right here, you can see that I'm mirroring my phone and I'm in the ChatGPT app.

And if I go over here and I click on this right here, which says remote, this lets me see all of my actual codecs. Chats that are right here. So if I go to her to and I go to summarize past month This is exactly what we were just looking at right here.

You can see this is literally the same conversation It has those three chats and I could come in here and I could say hi And this gets sent in my actual codex right here on my laptop or sorry on my desktop as well as my phone So as long as my machines on even though this is running locally and i'm out on a walk i'm out at dinner I could even be in a different state and I could control my desktop from my phone right here So let's say for some reason I left a file locally on my computer, I could just say, hey, can you email that file to me?

And then I could use it on my laptop. So anyways, just wanted to show off that remote capability. And that's going to bring us into concept number six, which is work trees.

Now, honestly, this is something where if you're not really building software or production apps, this isn't going to matter to you too much. Like I don't hardly ever use work trees, but it is good to know about since you will see these in the interface. So right here, you can see that if I was to go to the local setting, it says that you can work in a new local work tree, which says create a copy of Herc 2 to work in parallel.

So basically, because I'm working in Herc 2, if I was going to do something like risky, or if I wanted to try and experiment with something, or for example, if I was building software, And I wanted to try to change the way that the onboarding flow worked, but I wanted to like not interrupt the main branch, especially if other people on my team are working on that main branch, I could use a work tree to basically create a copy and I could test in there.

And then what you can do later is you can merge it back into the main. So it's basically just a duplicate that lets you test. But like I said, there's hardly any times for knowledge work that I ever use work trees.

And if I ever am building software and stuff, a lot of times Codex will say, hey, by the way, you might want to use this. You might want to do this in a new work tree. So I wanted to call it out because it's important because you might see things like this master branch and you might see other work trees that might have been spun up automatically.

but I will say it's not something that I'm actively thinking about or think that you need to be actively thinking about in order to get the most out of Codex. So I wanted to bring that up, but let's move on to number seven, which is .codex. Now, this is really interesting.

We talked about how when you're inside of your project, you have things like an agents .md. Now, if I go to my files, there's also something in here. called a .codex.

So if I open this up, what you'll see in here is I have some agents, and I have some workflows, but I also have a config .toml file, which is basically like some local settings. And there are basically going to be two different places where a .codex file lives, and that's either gonna be inside of your project, so like in this example, this is my .codex inside of my Herk 2.

But there's also going to be one which is user or global, which means this lives under your actual user. And these are settings that hold things like memories and sessions and automations. These are things that will apply to your global codex.

So whenever you're in codex, regardless of what your project you're actually in, the .codex will have personal settings and app data and stuff like that. Now, once again, I think this is something it's important to understand where that lives and what it does, but this isn't a folder that I'm actively thinking about and actively having to maintain.

But I did want to call that out because sometimes if you need to have a certain setting applied or something like that, that is where it lives. Now, the cool thing about this is if you're ever confused, you say, hey, Codex, what did you do there? Or do we need to add this to my .codex?

Or do we need to move this to a different folder? You can just ask it questions because it can search through its own documentation and understand how to do things correctly. But I do think that understanding the difference between project scope things and user scope things, that's definitely important, especially as we get into the .agents, which is gonna come up later.

And that is a great segue into part three of the core concepts, which is control and customization. So number eight here is about different AI models. So when you are starting a new chat, you can choose the model you wanna use.

You can also change the model in between messages. So I could shoot off a message on Astra and then I could, you know, change this to something else and then I could shoot off another message and I can keep changing if I want. Now, if you're watching this video way later, maybe the model names look different, but essentially the theory is the same.

You have different models to choose from and they all have different prices and strengths. So like right now, GBD6 Astra is just the absolute strongest, but it's gonna cost the most. It's gonna eat the most of your weekly limit.

Right here, you can see that I'm at 100 % because it just got reset. But as you chat with Astra, it will drain your weekly usage faster than 5 .6 Sol would, which would drain your usage limit faster than 5 .6 Terra would. Now right here, I just want you to look at the API pricing so I can explain to you how these models are actually billing you.

So they bill by tokens and they usually bill by 1 million tokens. Meaning on Astra, for every 1 million input tokens, it will cost you $10. And for every 1 million output tokens, it will cost you 50 bucks.

And here you can see the different pricing for these different models. And you can see how it gets cheaper as you go down. Now, this is API pricing, which means if you were building an automation that was programmatically using these models, that's how much it would cost you.

But this usage inside of Codex, that's not programmatic. This is usage limit or subscription limit. So when you pay 20 or 100 or 200 bucks a month for Codex, this is what your usage limit is eating up, is that subscription.

And you're getting way more inference. This is a $200 a month subscription. And if I use all of my weekly usage every single week, I'm getting about $14 ,000 worth of inference every month, and I'm only paying 200 bucks for that.

So the subscriptions are much cheaper than if you were using the API pricing. But if you ever do go over your weekly limit, you'd have to buy extra Codex credits. I don't exactly know why they price it like this, but you can see they just give you Codex credits instead.

So let's say I went over my weekly limit. I could go here and I could click on buy credits. This would bring up this screen, which lets you buy credits by like the thousands or something like that.

And that is what I wanted to show you guys. So like 2 ,500 credits, 5 ,000 credits, 25 ,000 credits. And this is how it would charge you based on those tokens.

Now, if you've never heard of tokens before, I know that might be a little confusing. Think about it as roughly four characters or roughly three -fourths of a word. And there's an OpenAI tokenizer right here where if I paste in some text, it will show you how many tokens this would be.

So this paragraph, for example, was about 240 characters, or sorry, 274 characters and about 56 tokens. And it's interesting because sometimes punctuation's a token, right? Like this period is a token, this comma's a token.

This word's a token, this word's a token. And this is basically just kind of breaking down what it might look like for these tokens. But anyways, you can see this was 56.

And what else you might notice is that I earlier said input tokens or output tokens. Input tokens are whatever you're feeding in. So your prompts, or if it's reading like a PDF or a Word doc, those are input because they're going into the model.

And output tokens are anything that it outputs. So even its reasoning here is output tokens. This entire thread is output tokens.

And that's why the output tokens are more expensive. Okay. So let's move on to number nine, which is about effort.

So within each model, not only do you choose the model, but you can also choose the effort level. So right here, I can move Asher down to medium, or I can move it down to low, or I can move it up to extra high. Or if I go like ultra, I have to allow full access because that can get pretty autonomous.

But anyways, the point being, as you move up on these effort levels, you get more quality, like you get more intelligence, but you also get more cost. It's more expensive to run it at higher effort levels. And that's why when you see these benchmarks, when there are new models released, you'll see the model, like for example, CloudFable 5, you can see this.

mark was on low and this one was on medium and this one was on high. So it shows how one model can behave differently and also cost a different amount based on the effort level that it was being used and tested at. Now, it actually is pretty nice to play with these because sometimes Astra High is just so much more powerful than what you need to do.

Like if you were using GBD6 Astra High to help you write an email, that's overkill and you're paying way more than you should. Realistically, to write an email, I could probably come down here and use 5 .6 Tera and be just fine with that and it's a lot cheaper. So definitely try to play with the model.

play with the effort based on your task. But a lot of times, if you're just doing a lot of knowledge work, I would just chuck it on sole and I'd maybe just go to medium and just call it a day there. Now, the other thing about effort is you can do something called fast mode.

This obviously makes the model respond quicker and do things quicker. And it's 1 .5 speed, but it costs you more usage. So it will eat your weekly subscription faster as well.

All right, and moving on to number 10, we have permissions. Basically right down here, you can see that this is orange and it says full access. This is basically Codex's YOLO mode, which means that it's not going to stop and ask you a bunch of permissions and questions.

Can I do this? Can I do this? If you go on ask for approval, it's going to always ask to edit external files and always ask to use the internet.

So that's if you want to sit here and really sort of like manage it and you're wanting to make sure everything's safe, you know, maybe start off like this. So you can just understand, oh, okay, it's asking me if it can run this command or it's asking me if it can open up my browser. you just get familiar with what it looks like for Codex to actually run that agentic loop.

You could also go on a proof for me. So it's only gonna ask you if actions are detected as unsafe, like maybe deletes or maybe certain sorts of like API calls or things like that. So that's how these permissions work.

And if you really wanna get granular about, oh, I wanna like make sure these certain actions never happen, you can ask Codex about that and it can work things into those config files like we talked about earlier with the .codex that it can certainly like block things out. But now let's move on to number 11, which I think is my favorite one, which is skills.

Skills are so, so important. They're basically reusable workflows. recipes that let you do something once and then teach Codex how to do it the same way every single time.

So now you can just run all these skills. So like I said, a skill is basically just a recipe. So let's say you come in here and you make a chocolate chip pancake and you followed a recipe to make that pancake.

You would basically say, okay, if I ever wanna make chocolate chip pancakes again, I'm gonna use this recipe because I know that I'm gonna get a good result. But let's say that you burnt these pancakes and you said, okay, well, what I need to do is make a quick update in this recipe to say, hey, cook it up 30 seconds less on each side.

And the next time you run the skill, you basically see, okay, are these better? Are these good? Do I need to make any more feedback in here?

Or now is this... just the way I like it. And then your agents are able to use these skills.

So here's an example where I basically gave it a YouTube video, so you can see that this is literally just a local file path of a YouTube video, and I said that I needed a video description, timestamps, and a LinkedIn post. It went ahead and it ran its agentic loop, and it gave me a description, timestamps, and a LinkedIn post.

And then I said, okay, now I need you to turn this into an X article and put the draft into my X as a draft. And it knew exactly how to do this with a thumbnail, with a title. It also put screenshots from my actual video and it sprinkled them throughout my whole X article.

And this is because it used a bunch of skills to be able to know how to do this. So right here, if I ask it, what skills did you use? It used my YouTube description skill, my timestamp skill, my LinkedIn post skill, my X article skill, my X article from video skill, my format X article skill.

It used all of these skills because I've taught it how to do something good once. And then every time I run these skills, I just give feedback. Hey, that was good.

You know, I liked this, but I didn't like this. update the skill. And now what's cool is you can just invoke these either with natural language, it will be smart enough to do it with natural language, or you can do it as a slash command.

So I have one in here that I can call like the audit, and you can see this will run my audit. Or I have one in here called the 3D brain, and this runs the 3D brain. Man, I think I planned this out so good.

This is gonna bring us on to number 12, which is .agents. So we talked about .codex, and .agents is basically the exact same. It just holds different things.

So right here you can see that the .agents holds your skills. And these, once again, can either be project level skills or they can be user scope skills. So they can be global skills.

So if I come into this chat right here, which is inside of my, you know, HERC 2 project, and I go to open up the files and I go to my .agents folder, which is right here, we have a folder called skills. And this is a bunch of skills, as you can see. Now, if I open up an actual skill, so for example, let me go to my YouTube description skill.

This is an actual skill .md. So once again, md just means markdown file. So what we have here is some metadata.

And this is what Codex will read to understand, okay, do I need to invoke this skill or not? So it has the name and it has basically when to use the skill. And then all the skill is, is just instructions.

It's hey. For Nate's YouTube descriptions, keep them short, two to four sentences, paragraph form, no bullet points, no em dashes. Here's the process.

You read the video outline, you identify the core topic, you write a short paragraph, and that's the skill. It's a very, very simple prompt. But if I go to my X article one, they can also be way more in depth.

So this one is way longer. There's a lot of other rules and a lot of other criteria. There's way more steps.

There's way more things to do. And also what's cool is skills can reference other things. Skills can reference other skills.

They can reference other agents. They can reference Python files or context files. So in this example, this skill references my voice guidelines, like my LinkedIn style guide, as well as my X article style guide.

But that is where your skills live inside of the dot agents folder. And then moving on to number 13 here, we have plugins. And this is where a lot of the magic actually happens because you can super easily come over here in the Codex desktop app, you can go to plugins, and you can connect a bunch of things.

You can see all the stuff that I have installed, like Alpaca, Gmail, ClickUp, Clay, Google Drive, GitHub, Canva, and the list goes on and on. You can also search for different plugins. So let's say you were really interested in connecting something like, I don't know, let's see, do they have a LinkedIn plugin?

They have a LinkedIn plugin where we can find the right profession. We can grow our business with ads. We have all these other sorts of plugins too.

So if you want to try to connect to your different apps, just come in here and see if you can connect them using these plugins. Maybe they have Composio, they don't have Composio. And if they don't have a plugin here, you can still connect it using a .env, using an API key, but plugins are so much easier because you can just basically sign in once.

They've got ones for creativity, for developer tools, for business and operations, for data analytics, communication, and they're basically always growing this plugin library. All right, so that now brings us on to part four, which is tools and scale. Okay, we've got five left here.

Number 14, we have the browser. Now this is definitely one of my favorite things about Codex. If I open up the tab over here, we can open up a browser, which is really cool because I can basically just control this.

So like I could go to school .com. And what you'll notice is when I log into school .com over here, I'm just going to zoom out a little bit.

You can see right here that I'm already logged into my school account so I could manage my different communities, which also means that Codex can help me manage my communities from here. Because when you save these logins inside of the Codex browser, it then saves them. So if you say, hey, could you go to that account and pull the report for me?

Or hey, could you go into school and could you make this post for me? It can do it through the browser if there's not a plugin or if there's not like an API or MCP server. And the browser use of Codex and specifically with GBT6 Astra is the best that I've ever used ever.

I'll play a clip right here where I asked it to basically open up Canva and paint me and draw me using the tools in there. And I gave it just a picture of me and it was able to replicate it in a way that I thought was really, really good. So silly example, but it definitely demonstrates the vision and the browser use capability of this agent.

All right, so number 15, we have sites. So if you go over here to explore and you click on sites, this is very, very cool because it basically lets you put things out there. onto the cloud.

Remember earlier we talked about local versus cloud? Codex can build something for you and it can be a local host, but then you say, hey, can you just put that on a site real quick so that you could have your team log in or so that you could share it with other people and make it publicly accessible? Like right here, you can see this London and Paris one is everyone, whoever has this URL could go ahead and open this up.

Whereas this one is just me. So it would literally have to be logged into my account in order to be able to open up this site. And what else is amazing about that is you can open up the analytics for the site.

So you can basically host things, you can connect to your own domain, and this is basically gonna replace something like a Vercel or like a, I don't know, a Squarespace, wherever you wanted to host your sites normally. You can now just do it right here from the Codex desktop app, which is pretty cool. You can also even connect a database.

So on the backend, if you had to store like user login data or permissions or settings or a database of... customer records, whatever it is, you could actually store that here too. And it's really cool.

All you need to do to actually use a site is you just basically have to do either say, hey, can you turn that into a site? Or you could do slash site. And then it will basically know that you can host a site or build a site.

So great touch here from Codex. All right, so number 16, we have sub -agents. So sub -agents are really cool because it lets you delegate work out to a bunch of different Codex agents.

So right here in this main chat, you can see that I'm talking to GPT -6 Astra, which is a really, really smart and intelligent, but also expensive model. So let's say I wanted to do a bunch of research, but I didn't want to waste Astra's brain on that. Or let's say I wanted to do like a bunch of different testing and just do a lot of stuff at once.

Astra here that I'm talking to can delegate work out to tons of sub -agents. Hey, I need you to search through X and search through YouTube and also dig through my own comments. So just find what are people talking about right now in the AI space?

Is there any big news or drama? I don't want you to do this research though. I want you to basically just delegate a bunch of sub -agents out to do the work.

And I want all of these sub -agents to be using 5 .6. Terra as the model. So shoot off those agents and let me know what they find.

So this is really cool because I can specify that I want those sub -agents, those little like researchers and workers to use different models. So right here, it's using its agentic loop and it's going to spin up these agents. And whenever you see these little colorful things, these are different agents that it spun up.

So right here, this one's called X. And when I open up X, it is actually this sub -agent working right here. I can watch.

this sub -agent go through its agentic loop and I can watch it like pull in data and everything like that. So this is the same thing as over here, but this one is just a different model and it's not the main session that you're currently talking to. You can see we have two others that started working.

We have this one called YouTube Pulse. And once again, it's going through the same exact agentic loop. We have this one called X Drama.

We have this one called Audience Comments. And we can basically watch all of these work, but we don't have to manage them because once these agents are done, The main session here will say, okay, cool.

All five are done. They reported back to me. Here's what they found.

And this is really good for gathering different perspectives or researching out a bunch of things in parallel. And these intelligent models are really, really good at delegation. So sub -agents are definitely something that you need to be using.

So all these sub -agents are fanned out and they're working. Let's move on to number 17, which is scheduled tasks. Scheduled tasks are awesome because this basically lets you have your Codex working autonomously for you.

So on the left -hand side, if you click on scheduled, you can basically ask Codex to schedule things for you, reminders, updates, or run full sessions. Because all a scheduled task basically is, is it injects a prompt into a Codex session like this. So it uses your same local files, it uses your same skills, it uses your model, it uses everything, the agentic loop, it just does it automatically for you.

So in here, you see that I could create one and I could set it up manually. I can have it run locally or I could have it run on the cloud. I could also choose what product to work in.

So Herc 2 or, you know, Hyperframes Editor. I can also choose what chat. So every one of these routines could be a new chat or it could be a new chat for this task or it could be even in an existing thread that we already have going.

And then you can basically choose when to schedule it. So it could be daily, it could be hourly, it could be, you know, weekly, or you could get it really, really custom in here. And the cool thing is you can describe everything in here manually or you could basically just ask.

codex to make one for you. So for example, I have these like seven or eight different ones that are running for trading. They're literally trading $10 ,000 of my real money.

And I have this set up as scheduled tasks so that they can basically all work together, talk to each other, check the market, make trades, things like that. And of course, you guessed it, these run in a different project right here called Trading Challenge. And they all run inside of this challenge thread.

And you can tell that this has scheduled tasks in it because there's a little clock icon next to that thread. And moving on to the very last concept we have for today, number 18, this is voice mode. Now, everything that you've learned so far, skills, sub -agents, projects, agents, all of this can be controlled by using your voice to have a real conversation with Codex.

With your voice, you can manage different Codex threads. You can delegate different agents. You can run skills.

You can set up scheduled tasks. You can build sites. You can do anything here with your voice.

And all you have to do is come down to the bottom left and click on start new voice chat. So forgive me here as I'm awkwardly like hunching down because my headphones don't reach far enough, but watch this. Hey Codex, so I need you to help me out real quick.

Inside of my HERC 2 project, can you start a new thread to make a thumbnail? I'm making this YouTube video about 18 Codex concepts. So if you could make me a thumbnail for my YouTube channel where you see me sitting at a desk, working on a laptop, and I'm pointing to the right and there's like 18 different little AI agents.

Yeah, just start with that. And then once you have finished that, Thumbnail could you send it to the thread that is currently working inside of my heart too as well?

and it is doing research on YouTube and X and Looking at my channel and stuff like that. Could you just send that thumbnail to that thread? All right.

Let me check on that Okay, I've got your thumbnail brief queued up with you with the laptop pointing toward the 18 ages Using your existing photos for reference and that found your research thread. So I'll send the finished on mail there And do you guys see how this is all coming together?

It started a work tree chat. It was doing the agentic loop. It also created a goal down here.

It literally created its own goal because it knew that we had this task for it and it started working. And the voice chat's still going, so I don't know what it's gonna think of this, but I just wanted to call that out because all of what we just talked about is all coming together. Because even you can see over here, it started making that actual prompt.

Like it shot off this prompt right here. and it started off this new thread. And now this is working in a work tree and it's gonna create that thumbnail for us and then it's gonna send it over into this actual chat and we'll see that actually happen.

But as you keep navigating, like you can go to the web and you can go to other things and you can keep this voice working. So as you're just working on your laptop, it'll show you what's going on and you can keep talking to it. And you can also do all this on your phone as well because of the whole remote thing, you can use voice mode when you're on your phone too.

Okay. So that finished up and you can see what happens is it actually can send tasks or messages to each other, right? So what happened was this was the original thread that it created to make the image and it made this, and then it sent it to this thread and said, Hey, Nate wanted me to send this to you.

Here it is. So with just your voice, I know that was a very simple example, but you're able to coordinate a bunch of different codex threads, manage them, multitask while you're on the go, while you're flipping between different apps. Very cool.

But anyways, guys, that is going to do it. I hope you enjoyed. I hope you learned something new.

But if you want to keep diving into more stuff with Codex, and I've made so many videos, I'll tag a playlist right up here that has just a bunch of Codex content. So I hope to see you guys over there. Thanks for sticking to the end, and I'll see you in the next one.

Thanks, everyone.

The Hook

The bait, then the rug-pull.

Nate Herk opens by promising a complete, non-technical map of Codex, all 18 concepts he says you actually need, delivered by walking through his own live project rather than a slide deck.

Frameworks

Named ideas worth stealing.

00:20list

18 Core Codex Concepts (4 parts)

Part 1 Foundations: Projects, AGENTS.md, Agent Loop, /goal
Part 2 Environments: Local vs. Cloud, Work Trees, .codex
Part 3 Control and Customization: Models, Effort, Permissions, Skills, .agents, Plugins
Part 4 Tools and Scale: Browser, Sites, Sub-agents, Scheduled Tasks, Voice Mode

Nate's own structure for the video, 18 Codex features grouped into four progressively more advanced buckets, shown on screen as a title card at 9:53.

Steal forany product-explainer video that needs to make a long feature list feel structured instead of overwhelming

04:25concept

The Agent Loop

An agentic harness reads context (AGENTS.md), thinks, selects and runs a tool, responds, then repeats thinking and tool use until the task is done, instead of answering once like a chatbot.

Steal forexplaining to a non-technical team why an 'agent' behaves differently than typing into ChatGPT

16:02model

Two Scopes: User vs. Project

User scope: C:/Users/Name/.codex (global settings, sessions, memories, plugins) and .agents (reusable skills across all projects)
Project scope: your-project/.codex (project overrides, config.toml) and .agents (project-specific skills)

Every Codex settings and skills folder exists at two levels, global (applies everywhere) and project-specific (applies to one folder only), shown in the 'Where Codex files live' diagram.

Steal fordeciding whether a rule, credential, or skill belongs in a shared system file or a single client/project folder

17:00list

Model Tiers: Astra, Sol, Terra, Luna

GPT-6 Astra, $10/$50 per million input/output tokens, strongest and most expensive
GPT-5.6 Sol, $4/$20
GPT-5.6 Terra, $2/$12
GPT-5.6 Luna, $0.20/$1.20, cheapest

Four model tiers trade intelligence for cost on a roughly linear scale; Nate states Astra costs about 2.5x Sol's rate.

Steal formatching task complexity to model cost instead of defaulting to the most expensive model for everything

20:43list

Permission Levels

Ask for approval: always asks before editing external files or using the internet
Ask only if unsafe (approve for me): only interrupts for actions flagged as risky, like deletes
Full access (YOLO mode): never stops to ask

Three escalating levels of how much an agent can do without stopping to check in first.

Steal forsetting guardrails appropriate to how much you trust a given agent/task combination

CTA Breakdown

How they asked for the click.

VERBAL ASK

33:59next-video

“I'll tag a playlist right up here that has just a bunch of Codex content.”

Soft end-card pointer to a Codex playlist rather than a hard sales pitch inside the video itself; the two paid offers (AI Automation Society and its paid tier) only appear in the written description, not spoken on camera.

MENTIONED ON CAMERA

00:00productAI Automation Society (Skool community) ↗

00:00productAI Automation Society free resources (Skool) ↗

00:00toolGlaido voice-to-text ↗

00:00toolHostinger VPS (Claude Code hosting) ↗

Storyboard