---
title: "OpenAI Codex review: the coding agent in CLI, cloud and IDE"
description: "Is OpenAI Codex worth it? Open-source CLI, IDE extension, cloud tasks, AGENTS.md, sandboxing and prices checked on the official page."
url: "https://vibe-code.blog/en/openai-codex-review"
lang: "en"
author: "Thierry Gustin"
published: "2026-09-21"
updated: "2026-11-21"
---

# OpenAI Codex review: the coding agent in CLI, cloud and IDE

> **Key takeaways** — OpenAI Codex is OpenAI's coding agent, available as an open-source command-line tool, as an extension for VS Code and its forks, and in the cloud for delegating long tasks. It comes with ChatGPT plans: limited on Free and Go, fully from Plus at $20 a month as of November 21, 2026. Its main strength is one consistent agent everywhere, with a sandbox turned on by default.

OpenAI Codex is OpenAI's coding agent: it reads your repository, edits files, runs commands and can keep working in the cloud while your laptop sleeps. Our OpenAI Codex review in short: it is the most logical choice if you already pay for ChatGPT, with one consistent agent across terminal, editor and cloud. Here is what stands out from the documentation, the pricing and published feedback.

## What is OpenAI Codex in 2026?

Codex is a coding agent, not the old completion model that once carried the same name. It comes in three main surfaces that share the same logic and the same ChatGPT account:

| Surface | What it is | What it's for |
| --- | --- | --- |
| Codex CLI | Terminal agent, open source under Apache-2.0 | Working on your local repo, scripts, CI |
| IDE extension | Codex panel in VS Code, Cursor or Windsurf; integrations in Xcode and JetBrains | Focused edits with the open file as context |
| Codex Cloud | Tasks run in a remote environment, started from the web, desktop app or mobile | Delegating long tasks, in parallel, laptop closed |

On top of that come automatic code review on GitHub pull requests and a Slack integration, both reserved for ChatGPT plans (not API keys).

The GitHub repository for [Codex](https://openai.com/codex/) describes the CLI as a "lightweight coding agent that runs in your terminal". Being open source matters: you can read what runs on your machine.

## What Codex does well

### A sandbox on by default

This is the most reassuring point for a vibe coder. The security docs say the agent runs with network access turned off by default, inside an OS-enforced sandbox typically limited to the current workspace. An approval policy decides when Codex must stop and ask you.

For a first read-only exploration:

```bash
codex --sandbox read-only --ask-for-approval on-request
```

Inside a session, the `/permissions` command shows and adjusts what Codex may do without asking.

> **No network is a feature** — When Codex asks for network access or to write outside the folder, read the command before you approve. It is often legitimate (installing a dependency), sometimes not. The sandbox exists precisely to make you ask the question.

### The cloud for delegation

Codex Cloud runs a task in its own remote workspace, which keeps going while your computer is asleep, according to the docs. You can start from a connected GitHub or GitLab repository, or from a default environment with no repository. When it finishes, you review the changed files, ask for fixes, then open a pull request.

From the terminal, `codex cloud` lists your remote tasks, submits new ones and applies the result to your local repo.

### AGENTS.md: instructions that travel with your project

Codex reads `AGENTS.md` files before doing any work:

1. a global file at `~/.codex/AGENTS.md` for your preferences;
2. then the project's `AGENTS.md` files, from the root down to the current folder, with the closest one having the last word;
3. an `AGENTS.override.md` temporarily replaces the file at the same level.

By default, Codex stops adding files once the combined size reaches 32 KiB (the `project_doc_max_bytes` setting). Keep instructions short. Bonus: Cursor and Claude Code can read `AGENTS.md` too, so one file serves several agents. Our guide to CLAUDE.md, AGENTS.md and rules files covers the method.

A starter template:

```md
# AGENTS.md

## Commands
- Install: npm install
- Check before handing back: npm run lint && npm test

## Rules
- Propose a plan before any change touching more than 2 files.
- Never put a secret key in front-end code.
- Ask me before adding a production dependency.
```

### Automation with codex exec

The `codex exec` command runs Codex without the interactive interface, for a script or a CI pipeline. Progress streams to stderr and only the final message goes to stdout, so you can chain it with other tools:

```bash
codex exec "summarize the repository structure and list the top 5 risky areas"
```

## How much does OpenAI Codex cost?

Codex comes with ChatGPT plans, which share their usage limits with ChatGPT's other agentic features. Prices checked on November 21, 2026 on the [official Codex pricing page](https://learn.chatgpt.com/docs/pricing):

| Plan | Price | What you get for Codex |
| --- | --- | --- |
| Free | $0 | A lightweight model in the desktop app, subject to rollout |
| Go | $8/month | Same scope as Free, for light tasks |
| Plus | $20/month | Codex on the web, CLI, IDE extension and iOS; automatic code review and Slack; extra credits available |
| Pro | From $100/month ($100, $200 or $500) | Everything in Plus with more usage; Ultrafast mode on the $500 tier |
| Business | $20 per user per month billed annually (2+ users), $25 billed monthly | Secure workspace, SSO, no training on your data by default |
| Enterprise and Edu | Contact sales | Advanced controls, audit logs, data residency |
| API key | Pay per use at API rates | CLI, SDK and IDE extension, no cloud features |

On limits, the page gives Plus users very wide ranges of local messages per five-hour window depending on the model, and stresses that these are not fixed caps. Cloud tasks generally use more than local messages. Pro plans had no five-hour limit when we checked, though weekly limits may apply.

> **API key or subscription?** — With an API key you pay per token and lose the cloud features (GitHub code review, Slack). For personal use, a ChatGPT plan is easier to budget. The API key is mostly for CI.

To place this cost in a project's overall budget, read [what a vibe-coded app really costs](https://vibe-code.blog/en/vibe-coding-cost).

## Strengths and weaknesses

### Strengths

- **One agent, three surfaces.** Terminal, editor, cloud: same account, same AGENTS.md instructions.
- **Open-source CLI.** Readable code, Apache-2.0 licence.
- **Sandbox by default.** Network off and writes limited to the workspace, with no setup.
- **Included in ChatGPT.** If you already pay for Plus, trying it costs nothing extra.
- **GitHub code review.** An `@codex review` comment on a pull request triggers a review.

### Weaknesses

- **Not an all-in-one builder.** No built-in hosting or database: you need to know how to deploy, for example with [Vercel or Netlify](https://vibe-code.blog/en/deploy-vibe-coded-app).
- **Hard-to-predict limits.** Published ranges vary tenfold depending on model and task.
- **A fast-moving offer.** Model names, surfaces and plans change often: check the docs before committing to an annual plan.
- **The vocabulary.** Between ChatGPT, Codex, Codex Cloud and the desktop app, it takes a while to know where to start what.

## Is Codex right for you?

- **Already paying for ChatGPT Plus?** Install the CLI or the extension and try Codex on a small project. Extra cost: zero.
- **Complete beginner?** An all-in-one builder like [Lovable](https://lovable.dev/), or the tools in [our Lovable vs Bolt vs Replit comparison](https://vibe-code.blog/en/lovable-vs-bolt-vs-replit), will feel more comfortable.
- **Want a full editor with a built-in agent?** See [our Cursor review](https://vibe-code.blog/en/cursor-review).
- **Torn with Claude Code?** Read [our Claude Code guide](https://vibe-code.blog/en/claude-code-beginners-guide), then the [Cursor vs Claude Code vs Codex comparison](https://vibe-code.blog/en/cursor-vs-claude-code-vs-codex).
- **Want the whole landscape first?** Start with [the best vibe coding tools](https://vibe-code.blog/en/best-vibe-coding-tools).

## Get started with the Codex CLI in five steps

1. Install the CLI:

```bash
# macOS / Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# or with npm
npm install -g @openai/codex
```

1. Move into your project folder and type `codex`. Choose "Sign in with ChatGPT".
2. First message: "Tell me about this project".
3. Create an `AGENTS.md` at the root (the template above).
4. Make a Git commit before and after each task, as the docs recommend, so you can roll back. If Git feels opaque, [our Git guide for vibe coders](https://vibe-code.blog/en/git-for-vibe-coders) eases you in.

Before shipping what Codex built, run through [the security checklist](https://vibe-code.blog/en/vibe-coding-security).

## Frequently asked questions

## Frequently asked questions

### Is OpenAI Codex free?

Codex is available on ChatGPT Free, but in a restricted way: a lightweight model in the desktop app, subject to rollout. To use the CLI, the IDE extension, the cloud and code review, you need at least ChatGPT Plus, at $20 a month as of November 21, 2026. The CLI itself is open source and free to install.

### What is the difference between Codex CLI and Codex Cloud?

Codex CLI runs on your computer, in your local repository, inside a sandbox limited to the workspace. Codex Cloud runs the task in a remote OpenAI environment that keeps going while your computer sleeps. Use the CLI to iterate quickly and the cloud to delegate long tasks or run several in parallel.

### Does Codex work in Cursor or VS Code?

Yes. The official extension installs in Visual Studio Code, Cursor and Windsurf, and Codex is also integrated into Xcode and JetBrains IDEs. It uses the open file and selection as context, shows changes next to the code and can send a longer task to Codex Cloud.

### Can Codex access the internet or my personal files?

Not by default: the docs say the agent runs without network access and inside a sandbox typically limited to the current workspace. Codex asks for your approval when it wants to go beyond those limits, depending on the approval policy you chose. Read each request before approving, especially on a project that holds secrets.

### Do I need an OpenAI API key to use Codex?

No. The recommended method is signing in with your ChatGPT account, and usage then counts against your plan. An API key is still an option for the CLI, the SDK and the IDE extension, billed at API rates, but without cloud features such as GitHub code review or the Slack integration.

## Sources

- [Codex CLI — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/codex/cli)
- [Codex IDE extension — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/codex/ide)
- [Codex Cloud — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/cloud)
- [Custom instructions with AGENTS.md — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/agent-configuration/agents-md)
- [Agent approvals & security — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/agent-approvals-security)
- [Non-interactive mode — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/non-interactive-mode)
- [Code review — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/code-review)
- [Pricing — ChatGPT docs (OpenAI)](https://learn.chatgpt.com/docs/pricing)
- [openai/codex — GitHub](https://github.com/openai/codex)

— Thierry Gustin, Formateur en intelligence artificielle (GustinAI) · November 21, 2026
