Codex CLI

OpenAI's terminal-based coding agent. Lighter-weight alternative to claude-code, runs on gpt-5-4. Open-source under the Apache-2.0 license (88.3k GitHub stars), sandboxed execution by default, good for scripting into CI or automated pipelines.

How It Differs From Claude Code

Sandboxed Execution

Codex CLI runs commands in a sandboxed environment by default — important distinction from Claude Code, which runs in your live shell. On macOS, sandboxing uses Apple Seatbelt (sandbox-exec); on Linux it uses bwrap (bubblewrap) + seccomp. This makes it safer for untrusted task execution but means it can't directly interact with your running services, databases, or filesystem outside the sandbox without configuration.

Good for: evaluation pipelines, CI tasks, situations where you want a guardrail around what the agent can touch.

Strengths

• Open-source — you can audit and modify the agent loop
• Sandboxed execution reduces blast radius for mistakes
• Lower startup overhead than Claude Code for short tasks
• Easy to script into pipelines (codex run "task description")
• Good integration with OpenAI function calling ecosystem

Weaknesses

• No hook system — less customizable agent behavior
• No context compaction — long sessions degrade
• GPT-5.4 currently trails claude-sonnet-4-6 on SWE-bench for autonomous coding
• Less battle-tested for overnight-runs type workloads
• Sandboxing limits integration with live development environments

Use Cases

Best for: short, well-defined coding tasks, CI/CD integration, automated code review steps, situations where safety guardrails matter, teams already on the OpenAI platform.

For longer autonomous sessions, claude-code with its hook system and context compaction is more suited.

Installation

npm install -g @openai/codex
codex "describe your task"

Also installable via Homebrew, or through the VS Code, Cursor, and Windsurf extensions.

Related

claude-code · gpt-5-4 · agentic-workflows · vibe-coding

Sources

• Codex CLI GitHub
• Codex Sandboxing Concepts