OpenAI has significantly expanded what Codex can do on a Mac, and the timing is not subtle. Just weeks after Anthropic announced desktop control capabilities for Claude Code, OpenAI pushed an update that lets Codex run multiple agents in the background, opening applications, clicking through interfaces, and typing, all while you continue working in other windows. The race to own the developer's desktop is now explicitly underway.
The core of the update is parallel background execution. Previously, agentic coding tools required the user to hand off control of their machine. The new Codex architecture runs agents as separate processes that operate independently, meaning a developer can set Codex to work on frontend iterations or API testing while they remain active in their own editor. OpenAI says the agents will not interfere with the user's own work in other applications, though how that holds up under real workloads remains to be tested.
Alongside background execution, Codex gains an in-app browser that lets the agent interact with web applications directly, useful for frontend and game development workflows where the target environment is a live browser rather than a local file. A memory feature, currently in preview, allows Codex to retain context from previous work sessions, building up a model of how a particular developer works over time. Image generation has also been added, aimed at producing mockups, placeholder assets, and product concept visuals without switching to a separate tool.
The integration story is broad: OpenAI announced 111 plugin connections covering tools like CodeRabbit, GitLab Issues, Slack, and Google Calendar. The practical effect is that Codex can now reach into the surrounding workflow, not just the code editor. A new pay-as-you-go pricing tier for enterprise and business ChatGPT customers makes the expanded feature set accessible without a fixed seat commitment.
The competitive framing is hard to miss. TechCrunch's headline put it directly: OpenAI is "taking aim at Anthropic" with this update. Anthropic's Claude Code had moved into Mac remote control territory last month, and Codex's expanded desktop capabilities appear to be a direct response. Both companies are converging on the same vision: an AI agent that is not just a tool you invoke but a persistent presence on your machine that handles parallel workstreams while you focus on higher-level decisions.
What neither company has fully resolved is the trust question. Giving an AI agent persistent background access to your desktop, your browser, and your workflow history is a significant commitment. Memory features that build a profile of how you work are useful precisely because they are accumulating something sensitive. The developer tools market will likely move fast on capability; how carefully it moves on the permissions model is less certain.
For now, the scoreboard between Codex and Claude Code is being updated roughly weekly. The next moves from both sides will probably come before the month is out.