Claude Code vs Codex: Which Coding Agent Fits You?

Claude Code and OpenAI’s Codex are now direct competitors: both are coding agents that run in your terminal, in your IDE and in the cloud, and both read files, run commands and edit code for you. The differences are in the models, the safety model, and how you customize them.

This comparison is based on each product’s official documentation as of 24 October 2026 (code.claude.com and OpenAI’s Codex docs). We don’t quote prices or benchmark scores because both change often; see our Claude Code pricing guide and OpenAI’s pricing page for current numbers.

Quick verdict

• Pick Claude Code if you want Claude models, a large ecosystem of skills, subagents, hooks and plugins, and fine-grained permission modes including plan mode and a classifier-reviewed auto mode.
• Pick Codex if you already pay for ChatGPT, want the agent sandboxed at the OS level by default, prefer an open-source CLI, or want OpenAI’s GPT models.
• Use both if your team mixes subscriptions. A shared AGENTS.md makes that easy.

At a glance

	Claude Code	OpenAI Codex
Models	Claude models only	OpenAI models
CLI install	curl -fsSL https://claude.ai/install.sh | bash, Homebrew, WinGet	curl -fsSL https://chatgpt.com/codex/install.sh | sh, npm i -g @openai/codex, Homebrew
CLI source	Proprietary	Open source, Apache-2.0 (Rust)
Surfaces	Terminal, VS Code and Cursor extension, JetBrains plugin, desktop app, web (claude.ai/code), mobile, Slack, Chrome	Terminal, IDE extension (VS Code, Cursor, Windsurf; Xcode and JetBrains integrations), ChatGPT desktop app, Codex Cloud, mobile, Slack
Account	Claude Pro, Max, Team or Enterprise; Claude Console; or Bedrock and other cloud providers	ChatGPT plan (Free through Enterprise) or OpenAI API key
Project instructions	CLAUDE.md (can read AGENTS.md)	AGENTS.md and AGENTS.override.md
Safety controls	Permission modes (default, acceptEdits, plan, auto, dontAsk, bypassPermissions) plus optional OS sandbox via /sandbox	Sandbox modes (read-only, workspace-write, danger-full-access) plus approval policy, set via /permissions
Reusable workflows	Skills (SKILL.md)	Skills (SKILL.md), invoked with $ or /skills
Delegation	Subagents in .claude/agents/	Subagents and custom agents
Lifecycle automation	Hooks	Hooks (hooks.json or [hooks] in config.toml)
External tools	MCP: claude mcp add, .mcp.json	MCP: codex mcp add, mcp_servers in config.toml
Packaging	Plugins and marketplaces (/plugin)	Plugins and marketplaces
Headless	claude -p, --output-format json	codex exec, --json
CI	GitHub Actions (anthropics/claude-code-action), GitLab CI/CD	GitHub Action (openai/codex-action@v1)
PR review	GitHub Code Review	GitHub code review (@codex review)

Getting started

Both install in one line and start with a single command in your project directory.

# Claude Code
curl -fsSL https://claude.ai/install.sh | bash
cd my-project && claude

# Codex
curl -fsSL https://chatgpt.com/codex/install.sh | sh
cd my-project && codex

Claude Code asks you to log in with a Claude account (or uses an approved ANTHROPIC_API_KEY). Codex asks you to Sign in with ChatGPT or use an API key. Our Claude Code getting started guide walks through the first session in detail.

Both tools also have an /init command that drafts the project instruction file: CLAUDE.md in Claude Code, AGENTS.md in Codex.

Models

Claude Code runs Claude models only, switchable with /model. Through the terminal CLI and IDE extensions you can also reach Claude via third-party providers such as Amazon Bedrock.

Codex runs OpenAI models. With ChatGPT sign-in, the models available depend on your plan; with an API key, they follow the models your key can access. /model changes the model and reasoning effort.

If you’ve already decided which model family codes best for you, that usually decides this comparison.

Safety model: permissions vs sandbox

This is the most interesting architectural difference.

Claude Code: permission modes first

Claude Code decides what runs through permission modes and allow/deny rules. Press Shift+Tab to cycle modes:

• default: asks before edits, shell commands and network access.
• acceptEdits: auto-approves file edits and common filesystem commands in the working directory.
• plan: researches and proposes changes, but doesn’t edit until you approve.
• auto: no prompts; a separate classifier model reviews each action and blocks anything beyond your request (on supported models).
• dontAsk (for CI) and bypassPermissions (for isolated containers only) are set by flag.

An OS-level sandbox for Bash is available on top: run /sandbox (Seatbelt on macOS, bubblewrap on Linux and WSL2). See our permission modes guide.

Codex: sandbox first

Codex combines a sandbox mode with an approval policy. Locally, an OS-enforced sandbox limits what commands can touch, with network off by default:

• read-only: inspect without changing anything.
• workspace-write: edit inside the workspace; ask before going outside it or using the network.
• danger-full-access: no sandbox; for isolated environments only.

In the app and IDE these appear as presets such as Ask for approval, Approve for me (requests routed to automatic review) and Full access; in the CLI you pick them with /permissions.

Takeaway: Codex is locked down by default at the OS level; Claude Code gives you more graded interaction modes and opt-in sandboxing. Both reach the same place if you configure them deliberately.

Project instructions: CLAUDE.md vs AGENTS.md

Claude Code reads CLAUDE.md at the start of every session, from your user folder, the project and subdirectories, and can also import other files with @path. If a repo only has AGENTS.md, Claude Code can read that instead. See the CLAUDE.md guide.

Codex builds an instruction chain from ~/.codex/AGENTS.md down through each directory to where you’re working, with AGENTS.override.md taking precedence in a directory. Combined instructions are capped by project_doc_max_bytes (32 KiB by default).

For a repo used with both tools, write conventions once in AGENTS.md and add a one-line CLAUDE.md that imports it:

@AGENTS.md

## Claude Code specifics
- Use the test-runner subagent for long test suites.

Extensibility

Feature lists have converged. Both now support skills, subagents, hooks, MCP and plugins.

Claude Code has the larger third-party ecosystem today: skill packs, subagent collections, plugin marketplaces and full kits are built for it first. Everything is plain files you commit: .claude/skills/, .claude/agents/, .claude/settings.json, .mcp.json. Start with our guides to skills, subagents, hooks and plugins, or browse the best Claude Code skills.

Codex stores configuration in ~/.codex/config.toml or a project .codex/config.toml (trusted projects only). Skills follow the same open Agent Skills format, and OpenAI ships a built-in $skill-creator. Hooks cover events such as PreToolUse, PostToolUse, SessionStart and Stop; non-managed hooks must be reviewed and trusted before they run.

Because skills share a format, a well-written SKILL.md often works in both with little or no change. Some multi-runtime kits, such as Agentkit.best, target both tools; see Agentkit.best vs DIY kits.

MCP

Both connect to the same MCP servers:

# Claude Code
claude mcp add context7 -- npx -y @upstash/context7-mcp

# Codex
codex mcp add context7 -- npx -y @upstash/context7-mcp

Claude Code shares project servers through a committed .mcp.json; Codex through a project .codex/config.toml. See best MCP servers for servers worth adding to either.

Automation, CI and the cloud

Headless. claude -p "task" and codex exec "task" both print the final answer to stdout so you can pipe it. Claude Code offers --output-format json and stream-json; Codex offers --json (JSON Lines events) and --output-schema for structured output. Codex’s exec starts in a read-only sandbox by default. More in our headless mode guide.

CI. Claude Code has anthropics/claude-code-action for GitHub and documented GitLab CI/CD support; Codex has openai/codex-action@v1. See Claude Code GitHub Actions.

Cloud. Claude Code on the web (claude.ai/code) runs tasks in the cloud and lets you pull them into the terminal with claude --teleport. Codex Cloud runs tasks in published cloud environments and can be driven from the CLI with codex cloud. Both offer GitHub PR review and Slack integration.

Choose Claude Code if

• You prefer Claude models.
• You want plan mode and graded permission modes you can switch mid-session.
• You rely on community skills, subagents, plugins or a ready-made Claude kit.
• You already pay for a Claude plan.

Choose Codex if

• You already pay for ChatGPT, or want to try an agent on a free plan.
• You want an OS-level sandbox and no network access by default.
• You value an open-source CLI you can read and build yourself.
• You want OpenAI’s models.

Using Claude Code and Codex together

1. One instruction file. Keep conventions in AGENTS.md; import it from CLAUDE.md.
2. Same MCP servers in both, so they have the same docs and issue access.
3. Cross-review. Let one agent implement and the other review; our code review prompt works in either.
4. Separate worktrees when both run at once, so they never edit the same checkout.

Related comparisons

• Claude Code vs Gemini CLI
• Claude Code vs Cursor
• Claude Code vs GitHub Copilot
• Claude Code vs Windsurf
• Claude vs ChatGPT for coding
• Claude Code tool profile