OpenAI’s Codex desktop app is all about managing agents

OpenAI is betting that coding agents are outgrowing the terminal.

The company launched a macOS desktop app on Monday for its Codex software development agent, which focuses less on the code itself and more on managing multiple software development agents in parallel.

Until now, the main way to interact with Codex was through the terminal, similar to Anthropic’s Claude Code or Google’s Gemini CLI. While the new desktop app still allows developers to drop into an IDE like VS Code or the terminal for hands-on work, OpenAI’s mission with this app is to make agentic coding more accessible to a wider range of users — even those who may never want to touch the code itself.

As OpenAI CEO Sam Altman pointed out during a press briefing ahead of Monday’s launch, over the last few months, software agents have “crossed a threshold of real utility.”

More than a million people already use Codex regularly, he said, but because these agents are now so capable, the team also wanted to look into ways to make them more accessible to a wider audience.

“5.2 [Codex] in particular is a model that many of us have found can do extremely complex things. We realized we started to feel limited by the interface. And so as we have gotten to these incredible reasoning models with this incredible capability, and we’ve been thinking about what we need to do to make it easy for people to get use of that capability. We have made this Codex app, which is a way to work with the model and manage agents as they do these like long-horizon tasks running on your laptop,” Altman explains.

The Codex team also argues that as developers begin trusting their agents with more complex work, an app like Codex is necessary to help them orchestrate and supervise them.

“Existing IDEs and terminal-based tools are not built to support this way of working,” OpenAI argues in its announcement.

Developers can point the Codex app at existing repositories, and it integrates directly with GitHub for creating pull requests. Multiple agents can run in parallel, and developers can switch back and forth between different tasks just like they would switch between files in an IDE, because the agents all run in separate threads.

Since Codex supports Git worktrees, multiple agents can work on the same repo in parallel, each using an isolated copy of that code.

From vibe coding to software engineering

As Altman noted, one question the company has been asking itself is whether these agents can transition from vibe coding to serious software engineering. Altman believes so.

“I think we’re over the bar for that,” he says. “I think this will be the way that most serious coders do their job in very rapidly from now.”

Codex beyond code

One interesting aspect of Codex is that it isn’t just about writing code. As Anthropic quickly realized with Claude Code, the core agentic loop that powers these tools is also useful for automating other kinds of workflows. Anthropic launched Cowork as a separate app for this use case, but OpenAI notes that Codex users can use Agent Skills to extend Codex to do this.

With skills, users can “extend Codex beyond code generation to tasks that require gathering and synthesizing information, problem-solving, writing, and more,” the team writes.

That’s not something OpenAI is emphasizing in this release, but given that the company has set its sights on enterprise use cases, we’ll likely hear much more about this in the future.

Availability

Codex is now available to anyone with a ChatGPT account, including, for a limited time, ChatGPT free and Go users. For those on paid ChatGPT Plus, Pro, Business, Enterprise, and Edu plans, OpenAI is doubling the rate limits during this time.