The tool itself is free and open-source (Apache-2.0). But you need a ChatGPT Plus subscription or an OpenAI API key to use it. API usage is billed per token.

Does Codex CLI work on Windows?

Yes, via WSL. Native Windows is not supported. Install WSL first, then follow the Linux installation steps.

What models does Codex CLI support?

GPT-4o, o3, and other OpenAI models. Switch models during a session with the /model command.

Codex CLI vs Claude Code — which is better?

They're complementary. Codex CLI is open-source with a plugin ecosystem and works well for quick tasks. Claude Code has deeper code understanding and handles complex refactoring better. Install both.

Can I use Codex CLI without a ChatGPT subscription?

Yes. You can use any OpenAI-compatible API endpoint by configuring a custom model provider in ~/.codex/config.toml.

How to Install OpenAI Codex CLI on Linux, macOS & WSL

May 25, 2026 · 4 min read

#AI #Tools #Tutorial #Codex #OpenAI

What is Codex CLI?

Codex CLI is OpenAI’s open-source terminal coding assistant. It reads your codebase, edits files, and runs commands — similar to Claude Code but fully open-source (Apache-2.0) with 85k+ GitHub stars.

I use it alongside Claude Code. For quick questions and small edits, Codex is faster. For deep refactoring across many files, Claude Code wins.

Installation Demo

Full installation on a fresh machine. Scrub the timeline or adjust speed:

How to Install Codex CLI

Prerequisites

Linux, macOS, or WSL
A ChatGPT Plus account OR an OpenAI API key

Step 1: Run the Installer

# Recommended: official installer
curl -fsSL https://chatgpt.com/codex/install.sh | sh

# Alternative: macOS via Homebrew
brew install --cask codex

# Alternative: npm
npm install -g @openai/codex

Done when: codex --version prints something like codex-cli 0.133.0.

Step 2: Authenticate

Two options:

Option A: ChatGPT Account (simplest)

codex

Select “Sign in with ChatGPT” on first run. Browser opens for OAuth.

Option B: API Key

Edit ~/.codex/config.toml:

model = "gpt-4o"
approval_policy = "on-request"

[model_providers.openai]
name = "OpenAI"
base_url = "https://api.openai.com/v1"
experimental_bearer_token = "sk-your-key-here"
wire_api = "responses"

Step 3: Verify

codex "Hello, what model are you running?"

Done when: You get a response with the model name.

How to Configure a Custom API Endpoint

For self-hosted or third-party OpenAI-compatible APIs:

model = "gpt-4o"
model_provider = "custom"

[model_providers.custom]
name = "My Proxy"
base_url = "https://your-proxy.example.com/v1"
experimental_bearer_token = "sk-your-key"
wire_api = "responses"
supports_websockets = false

Any endpoint that implements the OpenAI /v1/responses or /v1/chat/completions API works.

What Are the Three Approval Modes?

Codex CLI has a unique safety system — you choose how much autonomy to give it:

Mode	What It Does	When to Use
Read Only	Can only read files, no modifications	Code review, understanding a new codebase
On Request	Asks before each write/execute action	Daily development (recommended default)
Full Access	Fully autonomous, no confirmations	Trusted automation, CI pipelines

Set at startup:

codex --approval-mode full-access

Or in config:

approval_policy = "on-request"

What Can Codex CLI Do?

Task	Command	Notes
Explain code	`codex "What does this function do?"`	Reads surrounding context
Generate code	`codex "Write a REST API for user management"`	Creates files, installs deps
Fix errors	`codex "This test is failing, fix it"`	Reads error output, patches
Refactor	`codex "Convert this to TypeScript"`	Multi-file changes
Run tasks	`codex "Set up ESLint and fix all warnings"`	Installs, configures, runs

How Does Codex CLI Compare to Claude Code?

	Codex CLI	Claude Code
License	Apache-2.0 (open source)	Proprietary
Models	GPT-4o, o3	Claude Opus, Sonnet
Context window	200K tokens	200K tokens
Safety modes	3-tier approval system	Permission config file
Plugins	Community ecosystem	None
Strength	Quick tasks, extensible	Deep understanding, complex refactoring
Cost	ChatGPT Plus ($20/mo) or API	Claude Pro ($20/mo) or API

My recommendation: Install both. Use Codex for quick edits and questions. Use Claude Code for multi-file refactoring and architecture work. They complement each other.

Tips From Daily Use

/model to switch models mid-session. GPT-4o for speed, o3 for complex reasoning.
/help shows all slash commands. There are more than you’d expect.
Paste screenshots. Codex CLI accepts image input — paste error screenshots directly.
Start from project root. Same as Claude Code — context matters.

Troubleshooting

”Authentication failed” with ChatGPT account

Clear the cached session:

rm -rf ~/.codex/auth*
codex

Then re-authenticate.

Slow or no response

Check your API connectivity:

curl -I https://api.openai.com/v1/models

If this fails, check firewall/proxy settings.

”Model not found” error

Your account may not have access to the requested model. Try switching:

codex --model gpt-4o