Codex

Codex CLI: how it works and differs from App and Web

Learn how Codex CLI works in a local terminal, accesses repositories, runs commands, applies project rules, and differs from Codex App and Codex Web workflows.

Codex CLI is OpenAI’s local command-line coding agent. It runs in your terminal, works with the selected workspace, reads and edits files, runs commands, and applies your sandbox and approval settings. Unlike Codex App, its primary interface is the terminal. Unlike Codex Web, it runs in your local environment rather than a browser or remote runtime.

Short answer

It is useful when you want an agent to help with code changes, code review, test fixes, scripts, and multi-step engineering work inside a local repository.

Codex CLI is different from a normal ChatGPT conversation. It does not only suggest code. It works with the current directory, project instructions, tool results, and local execution permissions.

CLI, App, and Web at a glance

Option	Where it runs	Choose it when
Codex CLI	Local terminal and workspace	You need commands, files, and `config.toml` setup
Codex App	Desktop application	You need graphical sessions, plugins, or Remote Control
Codex Web	Browser or remote environment	The task should run in a cloud or remote environment

See Codex CLI vs Codex App vs Codex Web for a detailed comparison.

When this matters

You are installing @openai/codex for the first time
You are comparing Codex CLI, Codex App, and Codex Web
You want an agent to edit files in a local project
You need to configure a custom API provider, Base URL, or model
You want to understand sandbox, approval, and AGENTS.md

Core concepts

Concept	Meaning
Codex CLI	The Codex coding agent running in your local terminal
Workspace	The directory Codex can read, search, and modify
`config.toml`	The local Codex config file for models, providers, sandbox, approvals, and other behavior
`AGENTS.md`	A project instruction file that tells Codex repository rules and collaboration preferences
Sandbox	The local security boundary for filesystem, network, and command access
Approval	The policy that decides which actions need confirmation

Recommended workflow

Start Codex CLI inside the project directory you want to change.
Ask it to inspect relevant files or instructions before making broad edits.
For complex work, ask for a plan before letting it change files.
Pick sandbox and approval settings based on the risk of the task.
Put durable project rules in AGENTS.md and one-time requirements in the current conversation.

If you use a third-party API provider, focus on model_provider, base_url, wire_api, and model fields in config.toml. Codex and other OpenAI-compatible tools usually use an OpenAI-compatible endpoint, which is https://bettertoken.ai/v1 in this documentation.

Common mistakes

Treating Codex CLI as a chat-only assistant.
Assuming Codex CLI, Codex App, and Codex Web run in the same environment.
Changing only model without configuring the matching provider and Base URL.
Relaxing sandbox and approval settings without tests, review, or command boundaries.
Mixing OpenAI-compatible and Anthropic-compatible Base URLs.

About BetterToken

When you use a custom OpenAI-compatible provider, BetterToken can act as the model API layer for Codex CLI. It handles the API Key, Base URL, model routing, balance, and usage records.

BetterToken does not replace Codex CLI’s local sandbox, approvals, AGENTS.md, or command execution policy.

For general Codex API access and product use cases, see the BetterToken Codex API page. This FAQ focuses on Codex CLI and its local workflow.