OpenAI has released the Codex app for macOS, a new desktop application for working with AI software agents that can carry out long-running tasks. The app lets Mac users run several agents at once, review their work as it progresses, and step in as and when changes need approval or tweaking.
Codex is a software agent rather than a chat tool. It can edit code, run workflows, and continue working across extended periods with limited human input. The macOS app offers a central place to manage that work when multiple agents are involved.
Codex app for macOS
For a limited time, access to Codex will be included with ChatGPT Free and the new Go plan. Usage limits have been increased on Plus, Pro, Business, Enterprise, and Edu plans, and those higher limits apply wherever Codex is used, including the desktop app.
Each agent runs in a separate thread organized by project, which makes it possible to move between tasks without losing context. Different pieces of work can run in parallel while users review progress or work elsewhere.
As agents make progress, their changes appear directly in their threads. Users can review diffs, leave comments, or open the work in their editor to make manual edits.
Each agent works on an isolated copy of a repository through built-in support for git worktrees. Multiple approaches can be explored at the same time without conflicts and users can pull an agent’s work into their local environment or let it continue without touching their current git state.
Session history and configuration carry over from the Codex CLI and IDE extension and existing projects and preferences appear automatically when the app is opened.
Skills extend Codex beyond editing code by combining instructions, scripts, and resources, and these can be used for tasks such as gathering information, running workflows, or producing written output. The app includes an interface for creating and managing skills, and Codex can either be directed to use specific ones or select them based on the task.
In one demonstration cited by OpenAI, Codex was asked to build a racing game from a single prompt. It created multiple racers, eight tracks, and in-game items activated with the space bar. The agent handled design, development, and testing, including validating its work by playing the game itself.
The task continued autonomously as Codex identified missing features, implemented changes, and fixed bugs. It processed more than 7 million tokens while iterating on the project over time.
Agents are restricted by default to specific folders or branches and have to request permission for actions that require wider system access. These rules can be configured at the project or team level.
The Codex app is available on macOS starting today. Windows and Linux support is planned for the future and you can register your interest here.
What do you think about the Codex app? Let us know in the comments.
