OpenAI Codex: The Coding Agent You Already Have Without Knowing It.

There is one tool you are probably already paying for, and most of you are not using: Codex, the coding agent by OpenAI, the same company that brought us ChatGPT. And this is a true mindset shift. Until today, we got used to working with ChatGPT, Gemini, and Claude as language models: you ask a question, you get an answer, sort of like ping pong. Codex does not just answer, it executes. You tell it what you want, and it writes code, edits files on your computer, builds apps, websites, and automations. Without knowing a single line of code. What used to require a development team and a computer science degree can now be done by anyone who knows how to describe their idea. This guide covers everything: what Codex is, what an AI agent actually is, the difference from Claude Code, how to install it, and how to build your first project with no technical background.

The bottom line upfront: if you have a ChatGPT subscription, you have an entry ticket to Codex at no extra cost. On the 100 dollar subscription, you get computational power whose real worth is thousands of dollars, because companies are currently subsidizing the usage. It is simply a shame not to take advantage of it.

Want to watch this being built live?

This guide is just the summary. In the AI Master club, you get the full training, recordings, templates, work files, and an active community of people building with AI every week.

What is Codex

What is an AI agent anyway

Without a memory file, an agent has the worst memory in the world, like a goldfish it does not remember what it did a minute ago. Therefore, an important part of the work is giving it a file containing your history, preferences, and rules.

This is a guide from a live session of the AI Master club. Every such guide comes with a full recording, a practical demo, and community support.

What is a skill and what is the difference from an agent

Claude Code versus Codex: What is the difference

The advantages of Codex

• Token economics: If you run out of your quota in Claude Code, you can continue the exact same project in Codex. Both work on folders
• Agility: Usually faster in workflows, and more token efficient on the same project
• Built-in ChatGPT ecosystem: Image generation (GPT Image 2), transcription, voice comprehension, debugging, all without connecting external tools
• Knows you: If you have already worked with ChatGPT, your history and memory transfer directly to Codex
• Excellent Hebrew transcription: The Whisper model transcribes Hebrew really well, so you can give tasks by speaking instead of typing
• Works easily on Windows: Supports PowerShell without complicated installations, and on Mac as well

Why is speaking better than typing? We speak about four times faster than we type, and when we speak, we give our thoughts freedom. The model receives all this context as background, and the output turns out more accurate and richer. When typing, we filter way too much.

What can you build with Codex

Step by step installation

Installation and launch in seven steps

• Search Google for Codex OpenAI and enter the official OpenAI page
• Click Download and choose Mac or Windows
• Install with a short installation and open the blue icon
• Log in with your existing ChatGPT account
• Choose a project: a new folder for a new idea, or an existing folder if continuing a project
• Choose an approval mode (Ask for approval at first), model, and intelligence level
• Type or speak in Hebrew what you want to build, and Codex starts working

A project in Codex is a folder. Unlike regular ChatGPT, every piece of work takes place on a folder on your computer that contains all the files the agent creates. A new idea from scratch equals New Project (new folder). Continuing work means choosing an existing folder. The exact same folder works in both Claude Code and Codex.

Three approval modes: How much freedom to give Codex

Models, intelligence levels, and token economics

At first it is intimidating: Codex runs lines of code on the screen and stops to ask for approvals, unlike ChatGPT which only answers. That is completely normal. Open it, see what it asks, and approve. Did you not understand what it wants? Just ask it in Hebrew to explain to you in simple language what it is doing. You do not need to understand code.

The app up close: Threads, cloud, and Worktrees

Three runtime modes, and when to use each

• Local: Everything runs on your computer. Simplest, and what most people need
• Cloud: Start a task, close your laptop, and return later to where it stopped. Requires a GitHub connection
• Worktree: Codex clones the project to a parallel environment, so you can run two agents on the same code without conflicts

Live example: A quote system for a gardener

What is needed to turn a mockup into a working system

• OpenAI API: A credit card that connects ChatGPT capabilities (images, text) to the system
• Resend: The leading system for sending emails, so the system can send quotes automatically
• Any additional tool you want: Automatic follow-up, CRM integration, and more, you just connect more tools

Want to build such a system for your business?

In the club, we do not just talk about tools. We build real processes with them: landing pages, automations, agents, quotes, content, and work systems, together, step by step.

How to phrase a request so Codex executes it correctly

Request template based on the four pillars

Goal: {What you want Codex to build or fix} Context: {Which files, documents, or background are relevant. You can describe in simple words} Constraints: {Preferred technology, style, and things that must not be changed} Completion criteria: {How you will know it succeeded. For example: the page opens without errors, the email is sent, all buttons work} Before you start executing, build a short work plan and present it to me for approval.

Want dozens more ready templates for Codex, Claude, and Gemini? In the club, prompts, skills, work templates, and recordings showing exactly how to use them are waiting for you.

agents.md: The project memory file

Tool integration: MCP and connectors

Safety in installing MCP and skills: As long as it is official documentation from a known company, there is no problem letting Codex install it. If it is a source you trust less, stop and ask Codex if what it is about to install is safe before it executes.

Computer control, App Shot, and browser

Code control and Git, all from within the app

Automations: Scheduling the agent to work on its own

Deployment to the web with Vercel

Project sharing and version backup with GitHub

Where to find skills for Codex

Hebrew works better in Codex than in Claude Code. Left to right text is less disruptive, and voice transcription is excellent. Still, as with all tools, there are challenges with Hebrew inside images, and sometimes you need a few prompting rounds until the visualization comes out exactly as you wanted.

Frequently asked questions

No. Codex writes all the code behind the scenes. You just describe in natural language, including Hebrew, what you want. If you did not understand something it is doing, simply ask it to explain in simple language.

Mainly for token economy and the ecosystem. If your quota runs out in Claude Code, you continue the same project in Codex. In addition, if you already pay for ChatGPT, Codex is included, and comes with built in image generation and transcription. Both tools are very similar and it is a force multiplier.

Codex is included in the ChatGPT subscription. With the 20 dollar subscription you can build nice projects, but tokens are limited to a five hour window and a weekly quota. The 100 dollar subscription gives about five times more. The free account also has access, but tokens run out quickly.

Yes, once you understand the limitations. At first, work in Ask for approval mode to approve every action and build confidence. After gaining experience you can switch to Full Access. Always stop and ask Codex if installing an MCP or skill from an unknown source is safe.

Yes, and easily. Unlike some tools, Codex supports PowerShell and works directly on Windows without complicated installations, and of course on Mac as well.

Definitely. You can write to it in Hebrew, and also give tasks by speaking, as the Whisper transcription model recognizes Hebrew really well. This is a big advantage over Claude Code, which understands speech only in English.

No. Every conversation (thread) is saved, even after closing the app completely. You reopen it and everything is there, including the context you gathered. This is an advantage over working in the terminal, where closing the window deletes the context.

Yes, in Cloud mode. The task runs on OpenAI servers, you close your laptop and return later to where it stopped. This requires a connection to a GitHub account. In Local mode, on the other hand, the computer must stay turned on.

This is the project instructions file, the place where you write everything Codex needs to know and cannot guess: working rules, style, and permanent details. If you explain the same thing to Codex over and over, put it into agents.md and it will remember in every conversation. In Claude Code the equivalent is called CLAUDE.md.

Your task for the future: If you have a ChatGPT subscription, download Codex now (a matter of minutes to install), and ask it for an idea for a small and simple project. Let it explain step by step to you. Start small, gain confidence, and add capabilities gradually. Do not think of this as just another AI subscription, think of it as hiring a digital employee for your business.