OpenAI Codex CLI is an open-source agentic coding tool that runs in the developer’s terminal. Released in April 2025, it connects to OpenAI’s GPT and o-series models and provides them with access to the local file system and shell, enabling multi-step coding tasks. Code execution runs inside an optional network-disabled sandbox for safety.
Key capabilities
Sandboxed execution — By default, shell commands run inside a macOS sandbox (or Docker on Linux/Windows) that prevents network access and restricts file system writes to the working directory. This lets the agent run untrusted build steps and tests without risk to the developer’s machine.
File system and shell access — The agent reads, creates, edits, and deletes files and runs shell commands within the project. Changes are shown as diffs before application in interactive mode.
Multiple approval modes — Three operating modes: suggest (review every action), auto-edit (auto-apply file edits, confirm commands), and full-auto (approve everything). Developers choose the right trust level per task.
Model flexibility within OpenAI — Supports GPT-4o, o3, o4-mini, and other OpenAI models. The o-series reasoning models are particularly effective for multi-step debugging tasks.
Open source — Released under the Apache 2.0 license; the full source is auditable and self-hostable.
Autonomy level
Level 3 (supervised agent): interactive mode shows each proposed action for approval. full-auto mode raises the effective autonomy to 4 within the sandbox boundary.
Strengths
- Sandboxed execution provides a meaningful safety boundary for unattended runs
- Open source and auditable
- Multiple approval modes let developers dial the trust level per task
- Works well with o-series reasoning models for complex debugging
Limitations
- OpenAI models only (no model flexibility outside the OpenAI ecosystem)
- Sandbox setup on Windows requires Docker, adding friction
- Newer tool with a smaller ecosystem than Cursor or Aider