OpenAI Codex
OpenAI's multi-surface coding agent spanning a desktop app, CLI, IDE extension, cloud, and mobile remote control.
What it is
OpenAI Codex has evolved from a single cloud coding agent into a multi-surface product with five distinct interfaces: a standalone desktop app (macOS and Windows), an open source CLI (Apache 2.0, built in Rust), an IDE extension (VS Code, Cursor, Windsurf, JetBrains), a cloud/web interface at chatgpt.com/codex, and the ChatGPT mobile app for remote control from your phone. Over 4 million people use Codex every week as of May 2026. The desktop app is the richest surface, supporting parallel threads, Automations (scheduled recurring agent tasks), Computer Use (controlling macOS and Windows GUI apps by seeing, clicking, and typing), an in-app browser, and Appshots for capturing context from any app on your screen. The CLI is fully open source under an Apache 2.0 license, rewritten in Rust from the original TypeScript version, and installable via curl, npm, or Homebrew. All surfaces share the same underlying agent, configuration (TOML based config files and AGENTS.md project instructions), MCP servers, skills, rules, and memories. Codex uses OS native sandboxing by default (Seatbelt on macOS, bubblewrap on Linux, Windows Sandbox on Windows), providing strong isolation without relying solely on permission prompts. Current models include GPT-5.5 (newest frontier model), GPT-5.4 (flagship), GPT-5.4-mini, GPT-5.3-Codex (specialized coding model), and GPT-5.3-Codex-Spark (research preview for Pro users). Codex Cloud runs tasks in isolated sandbox environments with optional internet access, and results are delivered as GitHub pull requests. As of May 2026, you can start, steer, and approve Codex tasks from your phone via the ChatGPT mobile app, connected to a host machine through a secure relay.
Who it is for
Developers who want a multi-surface coding agent that works across desktop, terminal, IDE, cloud, and mobile from a single configuration
Teams using GitHub who want AI generated pull requests delivered through the standard code review process
Projects that need parallel cloud task execution, running multiple coding tasks simultaneously in isolated environments
Developers who value OS native sandboxing by default for stronger security without manual configuration
Security conscious teams looking for automated vulnerability scanning and remediation through the Codex Security agent
People who want to start coding tasks from their desk and steer or approve them from their phone while away
Teams that need scheduled, recurring agent tasks through Automations for ongoing maintenance, monitoring, or code health checks
Design to code workflows where Figma designs need to be translated into working implementations
Pros and cons
Pros
- +True multi-surface product: desktop app, CLI, IDE extension, cloud, and mobile remote control all share one configuration, so you can start work on one surface and continue on another.
- +OS native sandboxing by default (Seatbelt, bubblewrap, Windows Sandbox) provides stronger isolation than permission prompts alone, reducing the risk of accidental damage.
- +The CLI is fully open source under Apache 2.0, built in Rust, with over 75,000 GitHub stars. Supports multiple AI providers beyond OpenAI.
- +Mobile remote control via the ChatGPT app is a genuine cross-device capability: start tasks, approve commands, review diffs, and steer work from your phone through a secure relay.
- +Automations allow scheduling recurring agent tasks on a cron schedule, filling a niche that most competing tools require external cron jobs or third-party orchestration to handle.
- +Has a free tier to start with, and the CLI is free to use with any compatible API key. Lower barrier to entry than tools that require a paid subscription from day one.
- +MCP support in both directions: connects to third-party MCP servers as a client and can run as an MCP server itself, letting other agents (including Claude Code) invoke it.
Cons
- −Cloud execution means no access to local tools, databases, or services. If your project depends on a local database or API for testing, Codex Cloud cannot reach it (though the CLI and desktop app can).
- −Computer Use is not available in the EEA, UK, or Switzerland at launch. This is a significant geographic restriction for European users.
- −The desktop app and IDE extension are not open source. Only the CLI, SDK, and App Server are under the Apache 2.0 license.
- −Less extensible than Claude Code overall. While Codex has hooks, skills, MCP, and Automations, it lacks the depth of Claude Code's four-tier CLAUDE.md hierarchy, 27 hook events, managed enterprise MCP allowlists, and markdown-as-infrastructure customization system.
- −GPT-5.5 is available in Codex via both ChatGPT subscription and API key authentication, but API key users still lack access to cloud features (GitHub code review, Slack integration, cloud tasks), limiting the GPT-5.5 experience to local usage only.
- −Cloud tasks are slower than local tools. Each task provisions a container, clones your repo, and installs dependencies before work begins, adding minutes of overhead compared to instant local execution.
Pricing
- Codex CLI access
- Basic exploration and learning
- Limited usage per 5-hour window
- More Codex usage than Free
- Lightweight coding tasks
- Codex CLI access
- GPT-5.5, GPT-5.4, GPT-5.3-Codex, GPT-5.4-mini
- Codex Cloud (parallel background tasks)
- Desktop app with Computer Use and Automations
- Mobile remote control via ChatGPT app
- Everything in Plus
- 5x usage limits across all models
- GPT-5.3-Codex-Spark (research preview)
- Everything in Pro 5x
- 20x usage limits across all models
- Maximum cloud task throughput
- Everything in Plus
- No fixed seat fee, pay as you go based on usage
- Larger virtual machines for faster cloud tasks
- Everything in Business
- Custom usage limits (no fixed rate limits with credits)
- HIPAA compliant use in local environments
Prices change; check the official pricing →