Codex CLI

0:29 Codex CLI is OpenAI's coding agent for the terminal: install it with curl -fsSL https://chatgpt.com/codex/install.sh | sh on macOS or Linux (or the PowerShell installer on Windows, npm, or Homebrew), type codex in a project folder, and sign in with your ChatGPT account (Codex CLI docs). It inspects your repository, edits files and runs the tools already installed on your machine, inside a sandbox that limits what its commands can touch. It is open source (github.com/openai/codex), first released on Apr 16, 2025 (OpenAI), and the latest version in OpenAI's changelog is 0.166.0, from Oct 29, 2026.

The CLI is one of the ways to use Codex; it shares its config.toml, MCP servers and usage with the Codex App and the Codex IDE Extension. Step-by-step installation for every system, with troubleshooting, is on How to Install Codex CLI.

Quick facts

How do I install and start Codex CLI?

1. Install it with one of the four methods in Quick facts. The standalone installers update themselves when you run the same command again; with Homebrew, update with brew upgrade --cask codex (docs).
2. Open a project folder in a terminal and run codex.
3. The first time, choose Sign in with ChatGPT (a browser window opens) or another sign-in method.
4. Describe a task, for example "Tell me about this project".

OpenAI suggests making a Git checkpoint before and after each task so you can undo the changes (docs). On a remote or headless machine, codex login --device-auth signs in with a one-time code instead of a browser callback (Authentication).

Codex CLI commands

Source: Developer commands, OpenAI docs, read Nov 6, 2026. Useful flags: --model or -m to pick a model, --image to attach an image, --search for live web search, --cd to start in another folder and --add-dir to make one more folder writable (Codex CLI docs).

Slash commands

1:46 Inside a session, commands that start with / control Codex. The ones most people use:

• /model switches model or reasoning effort; /status shows the session and the limits left; /usage shows account usage.
• /permissions changes when Codex may edit files or run commands without asking.
• /init writes an AGENTS.md scaffold for the project; /review reviews your working tree; /diff shows the changes.
• /plan switches to planning; /goal sets a long-running goal (see Codex App).
• /compact shrinks a long chat's context; /new, /resume, /fork and /rename manage chats.
• /mcp lists connected MCP tools; /skills, /plugins and /apps browse extensions; /import brings in a Claude Code or Cursor setup.
• /fast toggles Fast mode where the model offers it; /voice lets you talk to Codex.
• /app continues the chat in the desktop app; /quit or /exit leaves.

Source: Developer commands, read Nov 6, 2026. Typing @ searches for a file to mention, a line starting with ! runs a shell command, Tab queues a follow-up while Codex works, Enter adds instructions to the turn in progress, and pressing Escape twice edits your last message and forks from there.

Models in the CLI

Pick a model with codex --model gpt-6.1-sol, codex -m gpt-6.1-terra, or /model during a session; the same flag works with codex exec. As of Nov 6, 2026 the recommended models are GPT-6 Astra, GPT-6.1 Sol (the default in config.toml), GPT-6.1 Terra and GPT-6 Luna. GPT-6.1 Terra needs CLI 0.159.0 or later and GPT-6 Astra needs 0.153.0 or newer. Reasoning effort runs from Low (Light elsewhere) to Extra High, plus Max and Ultra, which uses subagents (Models). More on Codex Models.

Approval modes and the sandbox

2:08 Two settings decide how free Codex is. The sandbox limits what its commands can read, write and reach on the network; the approval policy decides when it must ask you first. The default, the Auto preset, is --sandbox workspace-write --ask-for-approval on-request: Codex edits and runs commands inside the project, with the network off, and asks before going outside it (Agent approvals and security).

Source: Agent approvals and security, read Nov 6, 2026.

The sandbox is enforced by the operating system: Seatbelt on macOS, bubblewrap on Linux and WSL2, and a Windows sandbox (MXC where available) in PowerShell, and it covers every command Codex spawns, not only its own file edits (Sandboxing). Even in workspace-write, the .git, .codex, .agents and .aws folders stay read-only. Auto-review makes extra model calls, which count toward usage (Agent approvals and security).

config.toml

Codex reads ~/.codex/config.toml, and a trusted project can override it with its own .codex/config.toml. Command-line flags win over everything, then the project file, then profiles (--profile name), then your user file, then workspace defaults, the system file (/etc/codex/config.toml) and built-in defaults. A minimal file:

model = "gpt-6.1-sol"
approval_policy = "on-request"
sandbox_mode = "workspace-write"

Administrators can enforce a requirements.toml, for example to forbid approval_policy = "never" or full access (Config basics). approval_policy = "untrusted" is no longer supported; mark a project untrusted with trust_level = "untrusted" instead (Agent approvals and security).

codex exec and resume

codex exec "task" runs Codex without the interactive screen: progress goes to stderr and only the final message to stdout, so you can pipe it (codex exec "generate release notes for the last 10 commits" | tee release-notes.md). It starts in a read-only sandbox; add --sandbox workspace-write to let it edit. --json streams every event as JSON Lines, -o writes the final message to a file, and --output-schema asks for an answer in a JSON Schema. --full-auto is no longer supported (Non-interactive mode). codex exec resume --last continues the most recent run in the folder (Developer commands). The CI uses are on Codex SDK.

2:46 codex resume reopens a saved interactive chat from the current repository or searches across all of them, and codex fork copies one so you can try another approach (Codex CLI docs).

MCP servers

Codex connects to Model Context Protocol servers, local (STDIO) or remote (Streamable HTTP with a bearer token or OAuth), and the desktop app, CLI and IDE extension share the same list. Add one with codex mcp add context7 -- npx -y @upstash/context7-mcp or codex mcp add example --url https://mcp.example.com; entries live under [mcp_servers.<name>] in config.toml (MCP). Among its tips for making usage limits last, OpenAI lists limiting the number of MCP servers you use (pricing). The old codex mcp-server command, which ran Codex itself as an MCP server, was removed on Sep 5, 2026; use the App Server instead (changelog, Codex SDK).

What testers found

17:20 In an Aug 27, 2026 video comparing the Codex CLI with Claude Code's CLI, Tim Ruscica of Tech With Tim asked each to split a 5,000-line Python file without breaking anything. Six of the seven models he tried finished the job; the OpenAI models took "significantly less time", GPT-5.6 Sol was the fastest of all, and Claude Opus 5 took 36 minutes, almost five times longer, because it also built a tool to prove nothing had broken.

14:09 Asked to explain an unfamiliar codebase, Opus 5 took over five times as long as the other models and wrote 635 lines of analysis, while he found Codex "a little bit more to the point", giving what was asked. One tester's run, with the GPT-5.6 models of the time.

Frequently asked questions

Is Codex CLI free? Codex CLI is free and open source, and it is included with every ChatGPT plan, including Free and Go, though free use is limited. You can also run it with an OpenAI API key and pay API prices per token.

How do I change the model in Codex CLI? Run codex -m gpt-6.1-sol (or another model name) when you start, or type /model during a session. To change the default, set model in ~/.codex/config.toml.

What does --yolo do in Codex CLI? --yolo is an alias for --dangerously-bypass-approvals-and-sandbox: Codex runs with no sandbox and never asks for approval. OpenAI does not recommend it outside an isolated machine or container.

Does Codex CLI work on Windows? Yes. Codex CLI runs natively in PowerShell with its own Windows sandbox, and also inside WSL2. WSL1 is not supported since version 0.115.

How do I see my Codex usage in the CLI? Type /status for the session and the limits you have left, or /usage for your account. The web dashboard is at chatgpt.com/codex/settings/usage.

Is Codex CLI open source? Yes. The Codex CLI's code is public on GitHub at openai/codex. The models it calls are not open.

Sources

• Codex CLI, OpenAI docs, read Nov 6, 2026
• Developer commands, OpenAI docs, read Nov 6, 2026
• Authentication, OpenAI docs, read Nov 6, 2026
• Models, OpenAI docs, read Nov 6, 2026
• Pricing, OpenAI docs, read Nov 6, 2026
• Changelog, OpenAI docs, read Nov 6, 2026
• Agent approvals and security, OpenAI docs, read Nov 6, 2026
• Sandboxing, OpenAI docs, read Nov 6, 2026
• Config basics, OpenAI docs, read Nov 6, 2026
• Non-interactive mode, OpenAI docs, read Nov 6, 2026
• MCP, OpenAI docs, read Nov 6, 2026
• Codex SDK, OpenAI docs, read Nov 6, 2026
• Windows WSL, OpenAI docs, read Nov 6, 2026
• ChatGPT Work and Codex, OpenAI Help Center, read Nov 6, 2026
• OpenAI Codex, OpenAI, Aug 10, 2021, with later updates
• openai/codex, GitHub, read Nov 6, 2026
• Using OpenAI Codex CLI with GPT-5-Codex, OpenAI, Oct 14, 2025 (video)
• Codex vs Claude - an Honest Comparison, Tech With Tim, Aug 27, 2026 (video)