OpenAI is turning Codex into something closer to a control room for your entire development workflow, and it’s starting on the Mac. What began as “an AI that writes code” is now a full-blown desktop app for macOS that can juggle multiple agents, long-running projects, and even background automations, all tied into your local tools and cloud stack.

At its core, the Codex app feels like a command center: you spin up different agents as if they were teammates, each focused on a specific task or project, and then watch them work in parallel. Threads are organized by project, so instead of one long chaotic chat log, you get clean lanes of activity where you can see what each agent is changing, review diffs, and open edits directly in your editor when you want to jump in. The app ties into your existing Codex CLI and IDE setup, pulling in your configuration and history so you don’t have to start from scratch on day one.​

A big technical win here is how the app handles multiple agents touching the same repo: it leans on worktrees, giving each agent its own isolated copy of your code so they can explore different branches of an idea without stomping on each other—or your local git state. You can choose to periodically check out what an agent has done, or just let it keep grinding away while you work on something else, knowing your main branch stays clean until you’re ready. For teams already running multi-branch workflows and continuous integration, this model maps pretty neatly onto how real-world shops ship software today.​

But Codex is no longer just about cranking out functions on demand; OpenAI is pushing it into “uses code to operate your computer” territory via something it calls skills. Skills are essentially packaged workflows—bundles of instructions, scripts, and integration logic—that let Codex reliably talk to external tools, run jobs, and follow your team’s preferred patterns. In the app, there’s a dedicated interface to create and manage these skills, and you can either explicitly tell Codex which one to use or let it decide based on the task.

If that sounds abstract, OpenAI’s racing game demo is the loudest example of what this looks like in practice. Using a web game development skill and an image generation skill powered by GPT Image, Codex was asked to build a 3D voxel kart racer—Voxel Velocity—with multiple tracks, characters, items, full race flow, and pretty detailed handling and AI behavior, starting from a single initial prompt. Over more than 7 million tokens of work, Codex essentially took on the roles of designer, engineer, and QA tester, iteratively playing the game, adding missing features, and fixing bugs as it went. OpenAI has published the prompt, skills, and multiple iterations of the game so developers can see how it evolves over time.​

The skills library ships with a bunch of practical building blocks that mirror how real teams work today. There’s a Figma implementation skill that pulls designs, screenshots, and assets and turns them into production-ready UI code with 1:1 visual parity. There are skills for managing projects in Linear—triaging bugs, tracking releases, juggling workload—plus deployment skills for pushing web apps to Cloudflare, Netlify, Render, and Vercel, and document-focused skills that handle PDFs, spreadsheets, and docx files with polished formatting. There’s even a skill that keeps Codex aligned with the latest OpenAI API documentation by referencing an up-to-date repo.​

OpenAI itself has gone heavy on skills internally, using them for everything from eval runs and training babysitting to documentation and experiment reporting. That internal dogfooding shows in how the app treats skills as first-class: when you create a new skill in the app, Codex can immediately use it across the desktop app, CLI, and IDE extension, and you can check skills into your repo to share them with your whole team via a “team config” setup. For larger organizations, that means you can slowly build a shared library of battle-tested workflows—almost like internal microservices, but for AI behaviors.

Then there’s Automations, which is where Codex starts to look less like a conversational partner and more like a background worker that never clocks out. In the app, you can define Automations that run on a schedule, combining instructions with optional skills so Codex can, say, triage issues every morning, summarize CI failures, generate release briefs, or flag potential bugs before anyone even opens their laptop. The result of each Automation appears in a review queue, so humans still have the final say, but don’t have to manually kick off every recurring task.​

OpenAI says its own teams already use Automations to offload repetitive but important chores: daily issue triage, CI failure summaries, daily release notes, and bug checks—essentially the kind of maintenance work that keeps software healthy but rarely gets anyone excited. The Codex app even showcases Automations visually, including an example that periodically creates new skills, hinting at a future where Codex is not just executing workflows but actively expanding what it can do over time. The longer-term roadmap includes adding cloud-based triggers so Codex can respond to events in your infrastructure, not just time-based schedules on your laptop.​

On the human side of the interaction, OpenAI is acknowledging that developers don’t all want their tools to talk the same way. Codex now offers two personalities: one that’s terse and pragmatic—perfect if you just want things done with minimal chatter—and another that’s more conversational and empathetic. You can swap between these with a simple / personality command in the app, CLI, or IDE extension, and importantly, the underlying capabilities stay the same; you’re only changing how it talks to you, not what it can do.

Security is doing a lot of heavy lifting under the hood, which matters when you’re basically giving an AI deep access to your code and, potentially, your machine. The Codex app uses native, open-source system-level sandboxing similar to the Codex CLI, and by default, agents are confined to editing files within the folder or branch where they’re working. For anything that requires elevated permissions—like network access—Codex has to ask, and teams can define project- or org-level rules that specify which commands are allowed to run automatically.​

From an availability standpoint, the Codex app is rolling out first on macOS. If you already pay for ChatGPT Plus, Pro, Business, Enterprise, or Edu, you can use Codex across the CLI, web, IDE extension, and the new Mac app with your existing ChatGPT login, and usage is baked into those subscriptions, with extra credits available if you need more. In a notable move to pull more developers into the ecosystem, OpenAI is also making Codex available for a limited time to ChatGPT Free and Go users and doubling rate limits on all paid Codex plans during this period.​

The Windows story is “coming soon,” but it’s clearly on the roadmap. OpenAI says Codex usage has doubled since the launch of GPT-5.2-Codex in mid-December, and over the past month alone, more than a million developers have used Codex in some form. The company is promising to keep pushing on both the model side—faster inference, more capable agents—and the experience inside the app, especially around multi-agent workflows and context management when you’re hopping between agents.

Stepping back, the Codex app is a pretty clear statement about how OpenAI sees the future of “AI for developers.” The premise is simple but ambitious: if everything is ultimately controlled by code, then the better an agent is at reasoning about and generating that code, the more broadly useful it becomes—not just for pure programming, but for a wide spectrum of technical and knowledge work. Codex is OpenAI’s attempt to shrink the gap between what frontier models can theoretically do and what people can actually harness in their day-to-day work by wrapping those capabilities in tools that feel native to modern software teams.​

In that sense, the Mac app is less a standalone launch and more a visible layer on top of a broader Codex push: multi-agent orchestration, reusable skills, background Automations, and a shared security and configuration model that spans app, CLI, IDE, and cloud. If it works as advertised, Codex stops being just an assistant that helps you write code faster and becomes something closer to a programmable collaborator that can own chunks of your workflow end-to-end—while still letting you stay in the driver’s seat.