Codex can now run tasks on your desktop. OpenAI says the tool will work in the background on Macs while you keep working.
What changed
OpenAI has overhauled Codex, its developer assistant, adding features that let the model operate directly on a user’s computer. The firm said the update enables Codex to open applications, click, type and carry out multiple tasks in parallel on a Mac. That allows the assistant to perform auxiliary work while a developer keeps using the machine for other projects.
The company described the new ability as agentic: Codex can spawn several agents that run concurrently without getting in the way of a person’s active workflows. OpenAI listed examples such as iterating on front-end changes, running tests on apps and working in programs that don’t expose an API. The aim is to make Codex more than a code helper — to turn it into a background collaborator for routine and repetitive tasks.
How it works
OpenAI built a background process that mimics mouse and keyboard input to drive apps, clicking and typing like a user would. Instead of asking a developer to paste commands into a terminal or use an API, Codex can now control applications using a cursor that clicks and types. OpenAI says multiple agents can run at once on a Mac; for example, one could run tests while another handles a longer task such as building a project.
Codex now includes an in-app browser that lets agents interact with web interfaces. That browser lets the agents interact with web applications at the user’s direction. The company highlighted front-end and game development as early targets and says it will expand that browser control beyond localhost pages.
Competition with Anthropic
The new Codex features come as a rivalry with Anthropic intensifies over AI tools for developers. Observers of the coding-AI market have highlighted that some of the powers OpenAI is introducing echo moves made by Anthropic recently. Last month Anthropic rolled out capabilities that let Claude and related tools control a user’s Mac remotely when the user isn’t at the keyboard.
OpenAI’s additions appear aimed at closing that gap. By enabling background agents and richer browser control, the company is positioning Codex to serve the same kinds of workflows that made rival products appealing to businesses: remote automation, hands-off testing and help with complex front-end tasks.
Why developers might care
Developers often juggle repetitive chores: building, running tests, checking UI states, updating front-end components. Giving an assistant the ability to do those things without constant human prompts could speed up cycles. OpenAI argued the model could act as a coding partner that handles supporting tasks while a developer focuses on higher-level design and problem solving.
The in-app browser is notable because it lets agents operate inside web interfaces rather than only through APIs. That matters for teams working on tooling where no clean API exists or where the iteration loop is tied to a browser-based admin panel or staging site. In such setups, scripting a browser is often the fastest way to automate verification steps.
Security and control questions
Letting software click and type on your machine opens obvious security and privacy risks, and admins will want ways to restrict and audit those actions. OpenAI said the agents are meant to run without interrupting other work, but the model will still need rights to open and manipulate apps. How those permissions are granted, audited and revoked will shape whether IT teams adopt the feature.
Companies will want concrete controls — access limits, detailed activity logs and an emergency stop button — before they let agents run on corporate Macs. The company hasn't published a full enterprise-control checklist with the announcement, so firms will be looking for clear guardrails before rolling the capability out widely.
Workflows and integration
OpenAI described several practical workflows the expanded Codex could help with. One agent might iterate on UI tweaks while another runs automated tests, then a third could check a staging site via the in-app browser. Because agents run in parallel, long-running tasks wouldn’t block shorter, higher-priority ones.
The approach is intended to make Codex more flexible in corporate toolchains. Organisations that still rely on user interfaces rather than APIs stand to gain first, since the agents can interact with those interfaces directly. Over time, OpenAI expects to broaden the browser’s reach and the kinds of interactions Codex can handle.
Industry reaction and context
People tracking the developer-AI market have framed this release as part of a broader push by major AI companies to own the developer productivity stack. Tools that automate tests, generate code, and manage deployments are attractive to businesses because they promise time savings and fewer errors. That has turned AI-assisted development into a battleground where speed, integrations and safety features decide winners.
Observers note the overlap: Anthropic recently shipped desktop-control features for Claude, and Codex's update mirrors some of that functionality. Both firms are racing to offer the most seamless, hands-off experience for developers. For customers, the choice will likely come down to which product best fits existing toolchains and which vendor provides the security and control companies require.
What to watch next
Two things will matter going forward: adoption and guardrails. If teams adopt background agents in real projects, the pace of everyday development work could change. Releasing the feature to a broad set of users will test whether the permission model and auditing tools are robust enough for business use.
Interoperability matters because teams mix tools from many vendors, so Codex needs to play nicely with existing CI/CD pipelines and toolchains. How well Codex plays with existing CI/CD pipelines, project management tools and authentication systems will shape uptake. OpenAI has flagged plans to expand browser control and other integrations, but concrete partners and enterprise features will determine how quickly organisations move from trial to production.
For now, Codex’s revamp signals an aggressive push by OpenAI into developer tooling that controls the desktop — a space that was until recently more conservative about handing so much control to an automated agent.
Related Articles
OpenAI said Codex can now open apps, click and type on a Mac and run several agents in parallel while the user continues working.
This article was created with AI assistance.
