OpenAI launches macOS Codex app for supervising AI agents

OpenAI is making its Codex coding agent available as a standalone desktop app.

The company said existing terminal- and IDE-based tools struggle to support long-running, parallel agent tasks, making it difficult for developers to supervise and coordinate work once agents move beyond short, synchronous jobs.

Available initially for macOS, with a waitlist for Windows and Linux users, the new Codex app is designed as a control surface for managing multiple agents at once. Developers can spin up parallel tasks against the same repository, monitor progress over time, and review outputs as they complete. Under the hood, Codex uses Git worktrees to isolate changes, allowing agents to work concurrently without colliding. The app also integrates directly with GitHub, enabling agents to open pull requests as work finishes rather than handing results back as raw output.

Codex moves from CLI to agent command center

Codex’s public origins can be traced back to 2021, when OpenAI introduced Codex as a code-focused language model, exposed primarily through its API and later used to power tools like GitHub Copilot. At that stage, Codex was a model for generating code in response to prompts, rather than a system that could execute or manage work.

Fast-forward to April 2025, and OpenAI repurposed the Codex name for something different: a command-line tool designed to let developers hand off specific, well-scoped coding jobs from the terminal. That version of Codex focused on execution — issuing instructions, running tasks against a codebase, and returning results in a familiar, scriptable environment.

Shortly after, OpenAI expanded Codex’s reach with cloud support, allowing tasks to persist beyond a single local session and making it possible to run larger jobs without tying up a developer’s machine. However, with Codex’s capabilities growing, the scope of the work developers were handing off was growing too — from quick edits, to multi-step changes spanning hours or days.

According to OpenAI, this evolution exposed a new bottleneck. While Codex had become fairly capable at longer, more complex tasks, developers lacked an effective way to oversee multiple agents at once, switch between ongoing jobs, and review progress without losing context. The terminal, OpenAI argues, is optimized for issuing commands, not for supervising parallel, stateful work over extended periods.

“The core challenge has shifted from what agents can do to how people can direct, supervise, and collaborate with them at scale,” OpenAI wrote in a blog post. “Existing IDEs and terminal-based tools are not built to support this way of working.”

And this is where the new Codex desktop app enters the fray. It’s not positioned as an IDE replacement, which are typically designed around files, edits, and interactive debugging, assuming a human is doing most of the work. Codex, by contrast, treats agents and long-running tasks as the primary unit of activity, with developers stepping in to supervise, review, and decide when to intervene.

Whether this approach becomes the standard way of working with agents remains to be seen, but the direction is evident: as agents take on larger portions of software development, OpenAI is betting that the challenge has shifted from generating code to overseeing it.

Skills and automations for the people

In addition to rethinking how developers supervise agents, OpenAI is also using the Codex app to broaden what those agents can do over time.

Central to that effort are Skills, which are basically reusable capabilities that let Codex agents operate beyond narrowly scoped coding tasks.

Anthropic has taken a similar approach in Claude Code, where Skills are used to give agents more contextual awareness and continuity across sessions. OpenAI, meanwhile, has been quietly moving in the same direction inside ChatGPT and the Codex CLI, where Skills have been used to extend agent behavior beyond simple prompt–response interactions.

With skills baked into the new Codex desktop app, OpenAI is effectively making those reusable agent capabilities first-class objects, rather than features hidden behind prompts or configuration. Skills are surfaced as named, repeatable actions that developers can invoke deliberately, giving agents a clearer role and scope when they’re asked to take on work that extends beyond a single instruction.

For example, you can invoke a Skill designed to handle GitHub pull request feedback: identifying the open PR for the current branch, inspecting review comments that need attention, asking for clarification where required, and then applying fixes selectively. Rather than prompting an agent to reason through that process from scratch, the Skill packages the context, permissions, and execution steps into a single, reusable operation.

The app also introduces support for Automations, which is a way to run agent tasks on a schedule.

Automations allow developers to define recurring jobs — such as maintenance checks or ongoing updates — that run without manual prompts.

These features reinforce the notion that OpenAI isn’t just trying to help developers ask agents for help more efficiently; it’s building toward a model where agents carry forward responsibilities over time, with humans stepping in to guide, review, and redirect work as needed.

Community reactions to Codex macOS app

Simon Willison, a blogger, open source developer, and long-time observer of developer tooling, has had early preview access to the Codex macOS app ahead of its public launch.

In an early assessment, Willison treats the desktop app as an extension of OpenAI’s existing agent tooling rather than a wholesale rethink, focusing on how it packages and surfaces those capabilities for day-to-day use.

“It’s a solid app that provides a nice UI over the capabilities of the Codex CLI agent and adds some interesting new features,” Willison wrote, highlighting both Skills and Automations as two such features.

However, Willison does highlight some of the restrictions of Automations, which currently only run while a developer’s laptop is powered on – a huge constraint for truly unattended or long-running background work. OpenAI does address this limitation, though, confirming that it is working on cloud-based Automations.

The desktop-first approach also shapes how broadly Codex can be used today. OpenAI built the app using Electron, which OpenAI’s head of developer experience Romain Huet said was to support future expansion to other operating systems. Windows support is expected soon, though the company acknowledged that achieving robust sandboxing outside macOS has proven more complex. Responding to “Mac only” criticisms on Hacker News, OpenAI engineer Alexander Embiricos pointed to the lack of comparable operating-system primitives on Windows as a key challenge.

The new Codex desktop app is available on macOS today, with Windows and Linux versions in the pipeline.

Alongside the launch, OpenAI is also temporarily including Codex access for ChatGPT Free and Go users, while doubling rate limits for Plus, Pro, Business, Enterprise, and Edu plans across all access points.