Using ChatGPT and Codex

Chat, Work or Codex

Choose	Good for	Examples
Chat	Questions and answers, back-and-forth discussion	Asking questions, searching, brainstorming, writing messages, comparing options
Work	Describing the result you want and having it produce a complete, finished piece	Making slides, analyzing files, writing reports, making project plans
Codex	Writing code	Debugging, running tests, reviewing changes, building features

• Desktop app: choose ChatGPT or Codex from the dropdown menu in the top-left corner. After you choose ChatGPT, switch between Chat and Work above the input box. For a quick question, use quick chat: press Cmd+Option+N on macOS or Ctrl+Alt+N on Windows.
• Web: on the new conversation page, switch between Chat and Work above the input box. Codex is at chatgpt.com/codex.

The old agent mode has been retired. Work replaces it. Before important actions, Work stops and waits for your approval. Read carefully before you approve. On the web, Work runs in OpenAI’s cloud. It keeps running after you close the app, and you can follow it on the web or on your phone.

Common ChatGPT features

• Choose a model and thinking level: In the model menu, choose GPT-6, then adjust it with the Thinking slider. Instant is the fastest. The further right you go, the longer it thinks and the more detailed its answers are. Only Pro can use the highest few levels.
• Projects: Put the conversations, files and instructions for one piece of work into a project, so you don’t have to explain the background again when you continue later. Create a project in the sidebar. In the desktop app, a project can also be linked to a local folder.
• Files: You can upload documents, slides, spreadsheets, PDFs and images. ChatGPT can summarize and compare them, extract and organize information, or use them as material for new files.
• Search and deep research: When you ask about recent information, ChatGPT searches the web on its own and includes source links. To research something systematically and write it up as a report, click + in the input box in Work and choose Deep research.
• Personalization and memory: Under Personalization in settings, set the style and tone of answers. You can turn memory off in settings. For conversations you don’t want remembered, use a Temporary chat.
• Voice: On your phone, tap the voice icon next to the input box. In the desktop app, click Start voice chat. Voice in the desktop app needs Plus, Pro or an enterprise-type plan.
• Images: Describe the image you want, or choose the Create image tool. To edit an image, upload it or select an area, then say how to change it.
• Plugins: Install them under Plugins to connect outside tools such as Google Drive. In a conversation, type @ followed by a plugin’s name to use that plugin. If you turn on Computer Use in the desktop app’s settings, you can also install the browser extension, so ChatGPT can read and act on web pages in your browser.

Your first coding task with Codex

The steps below use the desktop app. The IDE extension and the CLI work the same way.

1. In the top-left corner of the ChatGPT desktop app, choose Codex, then open your project folder. The first time you use Codex in a folder, it may ask whether you trust it. Choose to trust it.

2. First have it only read, not change anything, so it gets to know the project:

   Give me an overview of this project: the directory structure, how to run it and how to run the tests. Don't change any files yet.

3. Then give it a small, clear change, and tell it how to verify the result:

   Add parameter validation to the login endpoint. Return 400 when the username is empty. When you're done, run npm test and make sure the tests pass.

4. When it’s done, first look at what it changed (see Check and undo changes). Keep the changes once you’re happy with them.

Leave a checkpoint before you start

Codex relies on Git to keep track of changes. With a checkpoint, you can undo anything you’re not happy with at any time. If the project isn’t a Git repository yet, just tell Codex: “Turn this folder into a Git repository and commit a checkpoint.”

Choose permissions

Permissions decide what Codex can do on its own and what it has to ask you about first. Choose them below the input box. In the CLI, type /permissions:

Mode	What it does
Read Only	Reads but doesn’t change anything. Use it when you’re new to a project
Ask for approval	Works normally inside the current project, and asks you first before going beyond it (for example, editing files outside the project). OpenAI recommends this for most cases
Approve for me	Same scope as above, but requests that go beyond it are sent to automatic review, so you don’t have to approve each one
Full Access	Not limited to the project directory. It may delete files outside the project by mistake. Not recommended

Don’t turn on Full Access just to “get it running first”. When it doesn’t have enough permission, approve that one request when it asks.

Check and undo changes

• Ask Codex to list which files it changed and why, or tell it “Review the changes you just made”.
• If you don’t want the changes, just say “Undo the changes you just made and go back to the checkpoint”.
• When you’re happy with the result, have it commit. That commit becomes the checkpoint for next time.

Set rules with AGENTS.md

Put the instructions you’d otherwise repeat every time into AGENTS.md in the project root. Codex reads it before it starts working. You can have Codex generate one (in the CLI, run /init), then edit it to fit your project:

AGENTS.md

# Project conventions
- Use pnpm, not npm.
- Run pnpm test after changing JavaScript files.
- Ask me before adding a new production dependency.
- Reply in English.

Put instructions that apply to all your projects in ~/.codex/AGENTS.md.

Check your usage

• With a ChatGPT subscription: your usage and reset times are at chatgpt.com/codex/settings/usage.
• With a 2api key: check usage and balance in the 2api console.

Common CLI commands

Only needed if you use the CLI

Type / in Codex to open the command list. It filters as you type.

Command	What it does
/status	Show the current model, permissions and context usage. On a subscription, it also shows your remaining limits
/model	Change the model and reasoning level
/permissions	Change the permission mode
/diff	Show what changed
/review	Have Codex review the current changes
/init	Create AGENTS.md
/compact	Compress the context when the conversation gets too long
/new	Start a new conversation
/resume	Continue an earlier conversation. You can also run codex resume in a terminal
/exit	Quit

codex exec "your task" doesn’t open the interactive interface. It exits when it’s done, which makes it a good fit for scripts. By default it doesn’t change files. To let it change files, add --sandbox workspace-write.

---

This page follows OpenAI’s official documentation, checked in October 2026. OpenAI updates often, so button names and menu locations may differ slightly. Go by what you see on screen.