Master Codex Desktop in 30 Minutes
Codex is more than a chat box. In this walkthrough I show you the full desktop workflow in four parts: the foundations (local projects, model selection, fast mode, plugins), the chat plus image generation, the agentic use cases (App Shot, computer use, skills, subagents), and finally automation with goals and scheduled tasks. If you want to move from casual AI use to actually operating with an agent on your Mac, this is the foundation.
What you need
- The Codex desktop application, which is currently optimized mainly for macOS
- An OpenAI plan. Fast mode works by default whether you pay $20 or $100
- The plugins for the tools you already use: Gmail, Slack, Notion, Supabase for developers, and so on
1. What Codex is compared with ChatGPT (0:39)
Think of Codex as the big brother of ChatGPT. It uses the same brain, but it is a lot more powerful because it has access to many more tools. Think of it as OpenAI's answer to Claude Code or Claude Cowork.
2. Create a local Codex project (1:11)
Inside Codex you can work in the cloud, like in ChatGPT, or inside a local project. Select New project, click start from scratch, and give it a name. I called mine Codex YouTube. That creates a real folder on your computer with that name, and you can ask Codex questions about that folder from then on.
3. Choose the right model and turn on fast mode (1:32)
On the upper right you have the model selection. Honestly, always use the latest model. It is the best bang for your buck, and GPT 5.5 is super efficient, with GPT 5.6 to be released soon.
Then there is something Anthropic does not give you: fast mode by default on your plan. You can change the speed at which the model replies and gives tokens back. I always work in fast mode and I have no problem with limitations. The $100 plan here is still superior to Anthropic's $100 plan for Claude.
4. Connect Codex to plugins and apps (2:43)
Plugins are the app store for Codex. Some come pre-installed, and you can add more: HeyGen, Fal AI, Canva, Figma, Supabase. My recommendation is simple: install the plugins for the tools you already use, Gmail, Slack, Notion, Supabase if you develop, and you are ready to go.
5. Basic chat and local file workflows (3:48)
Codex does everything ChatGPT does, and more. Start simple:
Hey, Codex, can you tell me how the weather is looking like for tomorrow in Paris?
Because it is connected to your local machine, you can then ask it to do something ChatGPT cannot:
Can you please leave a report inside of this folder regarding the weather?
It searches the web, writes a Markdown report into your project folder, and shows you where the file lives. The same pattern works with Excel documents, and if a folder is full of junk, Codex can go over the documents and structure them for you.
6. Generate images with Codex (6:15)
Codex has the native image generation of ChatGPT built in, via Image Gen 2. You can go from zero:
Can you please create an image of the Eiffel Tower under these weather conditions for tomorrow?
The real power is combining reasoning, web search, and image generation in one request, and running several chats at once. While the Eiffel Tower image was generating, I opened another chat:
Hey, Codex, I need you to fetch the results of the last three football matches for the World Cup. Then I need you to generate an image using your native image generation capacities, not a skill nor anything, and generate a board with the three results.
A spinning wheel means a chat is still working; a blue marker means it is finished, so you can hop between conversations and let Codex work in parallel.
7. Use App Shot for screenshot context (9:06)
App Shot is a feature most people do not know about, and it is incredibly powerful. Hit the two command keys at the same time and Codex takes a screenshot of the application you are on and sends it into your chat as context. Working on a presentation or a report? One shortcut and Codex has everything it needs, and you can ask it anything about that page.
8. Control your computer with computer use (10:52)
Install the computer use plugin from the plugins tab and Codex can take screenshots of your screen and click the right buttons for you, in the browser or in applications. I had it open Spotify, pick a playlist to lock me in, and play it.
Where this shines is anything that cannot be automated through an API connection. Managing your LinkedIn, checking your messages, maintaining your CRM. It is a bit slow, yes, but incredibly powerful and useful.
9. Turn repeated workflows into skills (12:52)
A skill is a reusable workflow. Whenever you are happy with one of Codex's outputs, ask it to turn the process into a skill and it will do the same thing over and over again. This is your first step toward building automated systems.
Can you please turn this into a skill? I want it to be a skill that actually checks the results for the World Cup matches of the day, and then generates an image with the results like this.
The result was a skill called World Cup Results Board: a skill file with the full step-by-step workflow (verify the results, filter matches, generate images, final response) plus the prompt pattern it will reuse. Skills are a standard, so you can share them with friends and colleagues and they will do the exact same thing on their machine.
10. Use subagents for parallel work (15:08)
Subagents are agents that have to be invoked by the main Codex you are talking to. Checking the weather took longer than I liked, so instead of doing three countries one by one:
Hey, Codex, can you please use three subagents to check for the weather in Sweden, Spain, and Portugal tomorrow? Each location gets one subagent.
Three subagents spin up, each with its own conversation, visible in the Subagents tab. When they finish, they report back to the main Codex, which presents the results. If you master subagents you will increase your productivity by at least 2x, because they let you tackle huge tasks in parallel: due diligence with one subagent per geography, reactivating CRM leads, generating ten outputs of the same thing without losing quality.
There is also a second form, background subagents, which let you keep talking with Codex while they run. I use both types; Codex chooses whichever is best automatically.
11. Personalize Codex with AGENTS.md (19:11)
Every local project can have an agents.md file that stores how Codex should work with you: your name, your business, the way you operate, your clients. You can even ask Codex to build it with you:
Hey, can you help me create an agents.md file to personalize how we actually interact? It's gonna be simple. Just ask me for my name.
From the next conversation on, that context is already in its memory. This is something you must do whenever you set up your Codex.
12. Run long tasks with goals (21:33)
Sometimes AI stops before reaching the objective. Hit the slash key, pick goal, and Codex will not stop working until the task is actually finished.
Hey Codex, I need you to check the weather for tomorrow for every single European country, one at a time, and then I need you to generate an image of the entire European continent with the predicted weather for tomorrow.
My longest loop has been 10 hours of generating videos. Get creative here: you can give AI a lot of freedom, and as long as the output can be verified, it will get to it.
13. Create scheduled tasks (23:04)
Scheduled tasks behave the same as in Cowork: click Create via Chat or pick a template, and the task runs every day at the time you set. I used the template that creates a classic game every day at 2 PM. You can inspect the prompt it sends on every run and change the model it uses. Combine everything from this video, image generation, skills, subagents, goals, with daily automations and you can automate a lot through natural language.
Pitfalls and tips
- Scheduled tasks run on your machine, so your computer needs to be on when the automation triggers. That is not something you can fix, which makes them great for personal automations but limited beyond that.
- Computer use is a bit slow. Use it for things with no API path, not for everything.
- Always pick the latest model and leave fast mode on. There is no downside on either the $20 or the $100 plan.
Where to go next
For cloud automations that do not need your laptop awake, see Automate Daily Work with Claude Routines. And to know what is coming to the model picker, check GPT 5.6 Drops Tomorrow: Everything We Know.
