# OpenAI Codex CLI — OpenAI's open-source (Apache-2.0) terminal coding agent, bundled into ChatGPT plans from Free through Pro — Business seats are now grandfathered-only.

> Source: The Agents Index — https://theagentsindex.com/codex-cli (structured, researched, re-verified)
> Facts last verified: 2026-07-23

OpenAI Codex CLI is OpenAI's terminal-native coding agent: an open-source (Apache-2.0), locally-run command-line tool that inspects code, edits files, runs commands and automates repeatable engineering work without leaving the shell. It installs as a standalone binary (curl script, npm, Homebrew) with IDE extensions for VS Code, Cursor and Windsurf, plus a desktop app and a cloud-hosted agent at chatgpt.com/codex. It draws on the same GPT-5.6 model family and credit pool as ChatGPT, so it is bundled into Free, Go, Plus and Pro ChatGPT plans rather than sold as its own subscription — though since June 24, 2026, new Business workspaces (or ones that never previously added a Codex seat) can no longer get one, and Enterprise remains a custom, sales-led add-on.

| Fact | Value |
| --- | --- |
| Website | https://developers.openai.com/codex/cli |
| API | Yes |
| Best for | Developers who already pay for a ChatGPT plan and want a terminal-native agent with granular, inspectable sandbox/permission controls across CLI, desktop and IDE surfaces. |
| Not for | Developers who want to freely swap in Claude, Gemini or an arbitrary third-party model with zero config, or teams that need flat, predictable per-seat pricing rather than token-metered credits that can spike. |

## Verdict

Codex CLI’s biggest advantage isn’t a feature — it’s bundling: because it draws on the same plan and credit pool as ChatGPT, a developer who already pays for Plus, Pro or Business gets a capable, actively-maintained terminal coding agent with no separate subscription decision to make, and even the Free plan gets limited access to try it. The tool itself is well-built and open-source, with granular permission controls, subagents, MCP support and a genuinely useful built-in code-review mode running across terminal, desktop and IDE surfaces. The honest catches are real: switching to a non-OpenAI model takes manual config beyond the built-in Ollama/LM Studio/Bedrock options, OpenAI’s own numbers put typical cost at $100-$200/developer per month for regular use, and a large, currently-open GitHub issue documents a real rate-limit consumption regression from June 2026 that was still unresolved as of this listing’s last check. Pick Codex CLI if you already live in the ChatGPT ecosystem and want a terminal agent without a new bill; pick Claude Code instead if you want a lower, more predictable entry price, or Windsurf/Juggler if free multi-model switching matters more to you than OpenAI’s specific model lineup.

## Strengths and weaknesses

- ✓ The only major CLI coding agent bundled into a plan with a genuine free tier — Free gives "Limited Codex access," so you can try it before paying anything.
- ✓ Apache-2.0 licensed and very actively developed: 100,960 GitHub stars, 15,132 forks and a release (rust-v0.145.0, shipped 2026-07-21) landed days before this listing was re-verified.
- ✓ Ships built-in support for local/self-hosted models (Ollama, LM Studio) and Amazon Bedrock alongside OpenAI’s own models, not just a single vendor.
- ✓ Real, current demand: "openai codex cli" gets an estimated 1,900 US searches/month with a low keyword-difficulty score (12/100) and rising interest, and zero AI-agent directories currently rank for it.
- ✗ An open, still-unresolved GitHub issue (208 comments, filed 2026-06-18, last activity 2026-07-20) documents Plus-plan users seeing rate-limit credit consumption jump roughly 10-20x since June 16, 2026, draining a 5-hour usage budget in as few as 2-3 prompts.
- ✗ Out of the box it only natively switches between OpenAI, Ollama, LM Studio and Amazon Bedrock — pointing it at OpenRouter, Azure or another third-party provider needs manual config.toml setup, and that endpoint must speak OpenAI’s "responses" wire protocol specifically, a real integration hurdle multi-model rivals like Windsurf don’t have.
- ✗ OpenAI’s own rate-card page puts typical real-world cost at roughly $100-$200 per developer per month for regular use — well above Claude Code’s $17/month entry point, once you move past the Free tier’s limited access.
- ✗ A large, high-traffic open-source repo with a correspondingly large backlog — 10,626 open issues against 100,960 stars — so an individual bug report can sit a while before triage.
- ⚠ Free-tier access is explicitly "limited", regular use effectively requires at least a $20/month Plus plan or a separate API key.
- ⚠ Rate limits are credit/token-based and have shifted materially (both the April 2026 pricing model change and the June 2026 consumption-rate regression), so the effective cost of a workflow can change under you without any action on your part.

## Key features

- **Configurable sandbox + permissions** — `/permissions` sets exactly when Codex may edit files or run commands without asking, and which directories stay writable, inspectable before every run.
- **Subagents, skills and MCP** — Delegate focused sub-tasks to subagents, package team instructions as reusable skills/plugins, and connect external tools over MCP.
- **Built-in code review** — A dedicated review mode (running GPT-5.3-Codex) scores uncommitted changes and surfaces prioritized findings before you commit.
- **Local, desktop and cloud surfaces** — The same agent runs in the terminal, as a native desktop app, inside VS Code/Cursor/Windsurf, or asynchronously in the cloud via `codex cloud`.
- **Token-metered credits** — Since April 2026, usage bills as credits mapped to input/cached-input/output tokens per task, replacing the older flat per-message estimate.

## Use cases

- **Terminal-driven feature work** — A developer already paying for ChatGPT Plus hands Codex a scoped feature ticket from the shell and reviews the diff before committing.
- **Pre-commit code review** — Before opening a PR, a developer runs Codex’s review mode against their uncommitted changes to catch issues a human reviewer would flag.
- **Async cloud delegation** — A developer kicks off a longer task via `codex cloud` and comes back to a finished diff instead of babysitting a local terminal session.

## Integrations

VS Code, Cursor and Windsurf IDE extensions · MCP servers (external tool connections) · Ollama, LM Studio and Amazon Bedrock (built-in local/alternate providers) · OpenAI API key auth (bypasses ChatGPT billing entirely) · GitHub (via `codex cloud` PR workflows)

## Sources

- https://developers.openai.com/codex/cli
- https://github.com/openai/codex
- https://chatgpt.com/pricing/
- https://help.openai.com/en/articles/6950777-what-is-chatgpt-plus
- https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers
- https://help.openai.com/en/articles/20001106-codex-rate-card
- https://learn.chatgpt.com/docs/config-file/config-reference
- https://github.com/openai/codex/issues/28879
- https://www.npmjs.com/package/@openai/codex