OpenAI presents the Codex app for macOS, a control center to work with multiple AI agents at the same time. The promise is clear: delegate long, complex tasks, run work in parallel, and collaborate with agents that can accompany you from design to software deployment.
Qué es la app Codex
Think of Codex as a dedicated window where your agents work in separate threads organized by project. Can you imagine switching tasks without losing context? That’s exactly what it aims for: read the session history, review changes in a thread, comment on diffs, and open files in your editor for manual tweaks.
Also, the app syncs with what you already use: it takes your Codex CLI and IDE extension settings and session, so you can jump straight into existing projects.
Cómo cambia la forma de construir software
So far, many people used a single agent inside the IDE or the terminal. Codex proposes orchestrating teams of agents: delegating subtasks, running processes in parallel, and letting agents continue projects that last hours or days.
To avoid conflicts, the app supports worktrees: each agent works on an isolated copy of the repository. You can explore different solutions without touching your local git state, and when something convinces you, bring it into your branch. You can also review and comment on diffs before accepting changes.
Codex doesn't just write code. With the so-called skills it extends its capabilities to tasks like gathering information, synthesizing results, writing documentation, or manipulating images. Skills package instructions, resources, and scripts so Codex can connect with tools and complete workflows according to your team’s rules.
A concrete example: OpenAI asked Codex to create a full racing game. Using an image-generation skill and a web-development skill, Codex designed, coded, and tested the game with more than 7 million tokens from a single initial prompt. It assumed roles of designer, developer, and tester to validate the experience.
Automatizaciones, seguridad y personalización
The app lets you create Automations that run in the background on a cron schedule you define. Repetitive tasks like daily triage, CI failure summaries, or brief generation? Automations can handle those and leave results in a review queue.
On security, Codex applies configurable sandboxing: by default, agents can only edit files in the folder or branch where they work and use cached web searches. For commands that require elevated permissions, they’ll ask you. You can also create project rules to authorize specific actions.
And if you prefer a different collaboration style, Codex offers two personalities: one more direct and pragmatic, and another more conversational and empathetic. You switch between them with the command /personality without affecting capabilities.
Quién puede usar Codex hoy y límites temporales
The app is available starting today on macOS. If you have ChatGPT Plus, Pro, Business, Enterprise, or Edu subscriptions, Codex is included and works with your ChatGPT account in the app, CLI, IDE extension, and the cloud. Also, for a limited time Codex will be available to ChatGPT Free and Go users, and OpenAI has doubled rate limits on paid plans during this period.
OpenAI also notes that Codex usage accelerated after the launch of GPT-5.2-Codex and that they plan to bring the app to Windows, optimize latencies, and expand cloud-based automated flows triggered by events.
Consejos prácticos para empezar
- Start small: ask an agent for scoped tasks and review the diffs before accepting. It’s the safest way to build trust.
- Use existing skills to connect with Figma, project managers, hosting, or image generation instead of building integrations from scratch.
- Configure sandboxing rules for your team and define which commands can run automatically.
- Automate the repetitive with
Automationsand leave the final review for when you need it.
Reflexión final
Codex isn’t just another programming tool; it tries to close the gap between what advanced models can do and how easy it is to use them in real work. If everything ends up controlled by code, tools like this determine who can build and how fast. Ready to try a workflow where agents aren’t one-off assistants, but coordinated team members?