OpenAI released a sprawling update to Codex on Thursday. The coding agent can now operate macOS apps on its own, run multiple agents in parallel on the same machine, and connect to more than 90 new plugins including Atlassian's Rovo, GitLab Issues, and Neon by Databricks. The company is also shipping an in-app browser, image generation via a new gpt-image-1.5 model, persistent memory, and a self-scheduling feature that keeps agents working across days or weeks.
The shorthand here is that Codex now looks a lot more like Claude Code. OpenAI isn't subtle about it. Anthropic's coding agent has been eating its lunch in enterprise accounts, and Thursday's announcement reads like a response document.
What "computer use" actually means
The headline feature is background computer use on macOS. Codex can see your screen, move a cursor, click, and type inside any app, not just a code editor. OpenAI pitches this for frontend iteration, native app testing, and anything without an API. Multiple agents can do all this simultaneously while you work in other windows, supposedly without stepping on each other.
I'll believe the "without interfering" claim when it survives a real workday. Agents that take over input focus are notoriously fragile, and anyone who has tried to demo Apple's own automation tools knows how quickly cursor-driven workflows break when windows shift. Still, if it holds together even 70 percent of the time, it's a meaningful capability for the "apps that don't expose an API" category. That's where a lot of actual QA work happens.
There are footnotes. Computer use is Mac-only at launch. It's not available in the EU, the UK, or Switzerland, per the Codex changelog. Some of the more interesting personalization features won't reach Enterprise and Education customers right away. The in-app browser is built on OpenAI's Atlas technology and, for now, is mostly useful for local webpages and game builds. Full web control, per the company, comes later.
Ninety plugins. Or 111, depending who's counting.
The plugin number is slippery. OpenAI's own language is "more than 90." TechCrunch reports 111. Either way, the list pulls in the places engineering teams actually spend their time: Atlassian Rovo for JIRA, CircleCI, GitLab Issues, CodeRabbit, Remotion, Render, Superpowers, and Microsoft Suite. Neon, the Postgres database shop Databricks bought, is in there too.
Plugins bundle three things: skills, app integrations, and MCP servers. That MCP bit matters, because it means Codex can talk to anything that speaks the protocol, not just OpenAI's curated list. In practice, most developers will probably touch two or three of these and ignore the rest. A plugin count is a marketing number. The ones that matter are the ones you already pay for.
Memory, and scheduling yourself future tasks
This is the part I keep coming back to. Codex now has a memory layer (in preview) that holds on to your preferences, your tech stack, and recurring workflows. Fine, everyone has memory now. But OpenAI is also letting the agent schedule future work for itself, waking up on its own to continue long-running projects "potentially across days or weeks."
An agent that sets its own alarm clock. That's the pitch.
Whether you find that exciting or alarming probably depends on how much you've worked with autonomous agents before. The promise is that Codex can triage a backlog over a weekend, keep a migration moving while you're in meetings, or monitor Slack and Gmail for things that need handling. The reality, based on similar features in other tools, is that you'll spend the first month cleaning up the things it decided to do while you weren't looking. Which is fine. That's how hiring a junior engineer goes too. But it's worth calibrating expectations before giving the thing cron-level access to your workflow.
OpenAI says Codex has about 3 million weekly users. The company hasn't broken out how many of those are paying, and it just launched a new $100-a-month Pro tier along with pay-as-you-go pricing for Business and Enterprise accounts. There's clearly room to grow revenue per user.
The rest of the manifest
GitHub pull request review comments now live inside Codex. You can run multiple terminal tabs in one thread. SSH connections to remote devboxes are in alpha. The sidebar previews PDFs, spreadsheets, documents, and presentations before you commit. The gpt-image-1.5 model handles image generation inline, which OpenAI suggests using for product mockups and slide visuals.
That's a lot of surface area for one release, which is itself the story. OpenAI is trying to turn Codex into what reports describe as a super app "in the open," bolting browser, image generation, memory, and desktop control onto a coding tool. Whether that's coherent expansion or an everything-bagel of features depends on how the pieces hold up under real use. The developer docs suggest OpenAI knows it has too much going on, too: they lean heavily on the phrase "command center," which is usually what you say when you're worried users will get lost.
What to watch
Short-term question: is computer use stable enough that developers trust it with real workflows, or does it stay in the demo-it-on-stage category? Medium-term question: when do EU and UK users get access, since those markets matter for enterprise adoption and the regulatory clock isn't getting shorter? Long-term question: do self-scheduling agents actually save time, or do they just create a different kind of work, the kind where you audit what the agent did yesterday?
The update is live today for Codex desktop users signed into ChatGPT. Personalization features reach Enterprise, Education, EU, and UK accounts on an unannounced timeline. Computer use in the EU and UK: no date.
