OpenAI explains how to set up Codex: personalization, detail level, and permissions

OpenAI has published a short guide to configuring Codex, its AI assistant for getting work done on your computer. The company lays out which settings to check in the first few minutes so the agent behaves predictably, doesn't fall asleep halfway through a task, and doesn't clutter the interface with unnecessary detail.

Where to start

OpenAI advises against trying to learn the full list of settings at once. The settings menu sits in the lower-left corner of the Codex interface, and for your first runs it's enough to go through a few basic options. The approach is practical: first remove anything that could derail a long task, then pick a display style that suits you, and only after that tune the agent's behavior to your own working rhythm. That order saves time and reduces the risk of accidental failures.

• Prevent sleep while running — keeps your computer from going to sleep while Codex is running a task.
• Detail level — controls how much behind-the-scenes information the agent shows as it works.
• Personalization — lets you choose a friendlier or more direct tone.
• Custom instructions — adds persistent instructions, just like in ChatGPT.
• Avatar — displays a floating avatar that helps you keep track of background tasks.

The point of this set is that it answers three key questions at once: how Codex behaves during long-running work, how much technical detail the user sees on screen, and how well the interface fits into an existing workflow. For a new user, that matters more than the rarer, fine-grained options you'll only need after a few real runs. OpenAI suggests focusing on these first, rather than on cosmetics.

Controlling the process

OpenAI puts particular emphasis on the Detail level setting. If you want maximum control, you can choose Coding mode, in which Codex shows the specific commands it runs as it works. That's useful for developers and anyone who wants to follow the agent's logic almost step by step. If the task is routine and the extra technical noise gets in the way, the company recommends switching to Default to keep the conversation cleaner and calmer.

In effect, OpenAI is separating two usage scenarios. The first is an observation mode, where the user wants to check every action and understand exactly what the agent is running. The second is a delegation mode, where the result matters more than a full stream of internal telemetry. For a product like Codex, that split is critical: some users want near-terminal-level transparency, while others want a simple assistant that doesn't distract them with unnecessary detail.

Personalization, background work, and permissions

The second key block is Personalization. It works the same way as personalization in ChatGPT: you can choose a tone and add your own instructions. It only looks like a minor detail. In practice, these settings determine whether Codex answers briefly and to the point or, conversely, explains more and acts as a more conversational assistant throughout the session.

OpenAI also issues a separate reminder about permissions. The company explicitly recommends starting conservatively: don't grant the agent extra rights up front, and expand access as it becomes clear which tasks you actually trust it with.

The same block includes a more mundane but useful feature: Prevent sleep while running. If your computer goes to sleep during a long operation, Codex may simply stop. And under Appearance > Avatar, you can turn on a floating avatar, drag it around the screen, and return to the task while the agent works in the background.

What it means

OpenAI's new guide shows that Codex is being pitched not only as a powerful tool but also as an everyday work interface. The core message is simple: to work properly, an agent needs not just models but clear user settings, from detail level to carefully granted permissions.