TL;DR
OpenAI rebuilt the Codex desktop app this week. The headline: Codex can now drive your Mac on its own, click around in other apps, remember context across sessions, schedule work that runs for days, and plug into 90+ external tools. It’s the biggest Codex shake-up since the app shipped, and it pushes the product past the editor onto the whole desktop.
What actually shipped
OpenAI dropped the update on April 16 under the tagline “Codex for (almost) everything”. The blog post is short; the feature list is not. Here’s the compressed version of what changed in a single release.
| Area | Before | After (April 16, 2026) |
|---|---|---|
| Scope | Code editor + terminal | Full-desktop agent on macOS |
| Other apps | None | Sees, clicks, types in any Mac app |
| Browser | External | In-app browser with page comments |
| Image gen | Not available | gpt-image-1.5 inside the app |
| Plugins | smaller prior set | 90+ new, incl. Atlassian Rovo, CircleCI, GitLab, Render |
| Memory | Per-session | Persistent preview across sessions |
| Long tasks | Session-bound | Scheduled, resumable across days |
The short summary from OpenAI’s own announcement: “Codex can now operate your computer alongside you, work with more of the tools and apps you use every day, generate images, remember your preferences, learn from previous actions, and take on ongoing and repeatable work.” That’s a lot in one sentence, so let’s unpack the parts that actually change how you’d use the thing.
Computer use is the centerpiece
The update’s centerpiece is background computer use. Codex now has its own cursor. It can see what’s on screen, click, type, open apps, and run workflows the same way a human would. OpenAI is careful to say this runs “alongside you”: multiple agents can work on your Mac in parallel without stepping on what you’re doing in other apps (MacRumors coverage).
The framing is important. Until this release Codex was something you fed code into; now it’s something you assign a task to and walk away from, coming back to a finished PR, a screenshotted Figma update, or a reply drafted in Outlook. The agent uses apps the way you do, no API required.
flowchart LR
A[You assign task] --> B[Codex opens needed apps]
B --> C{Needs info?}
C -->|Yes| D[Uses in-app browser]
C -->|No| E[Clicks, types, screenshots]
D --> E
E --> F[Runs terminal / commits]
F --> G[Returns output + memory]
G --> H[Schedules follow-up]A few observations from the launch materials. First, this is macOS only on day one. Windows users are not in the announcement. Second, personalization features and computer use are gated out of the EU and UK at launch, with “gradual rollout” promised. Third, OpenAI is drawing a soft line between “web app computer use” (already shipped earlier) and “your-Mac computer use” (new). The new piece is the desktop one.
The risk signal here is louder than OpenAI’s blog admits. “Uses all your apps with its own cursor” is also “can write to any file, send any email, hit any API your session is already authed into.” Help Net Security’s write-up, one of the few that took a security angle, notes OpenAI plans to expand Codex’s browser control “beyond localhost web apps,” a polite way of saying the sandbox will get thinner over time. If you’re running this on a work machine, audit what’s logged in before you hand over the cursor.
Memory, automations, and work that runs for days
The second shift is temporal. Until now, Codex started fresh on every session: you opened it, asked your question, got an answer, closed it. The new version retains state across sessions, days, and weeks.
Memory preview stores “personal preferences, corrections, and information that took time to gather.” In practice, that means you stop re-explaining your tech stack, your lint config, or the one weird thing about your auth flow every time you start a conversation. Memory is described as a preview for now, and personalization is one of the features geoblocked out of the EU and UK, so expect the usual quiet rollout.
Automations now pair with memory in a way that actually changes what you ask for. Codex can schedule future work for itself and wake up automatically to continue a long-running task, which OpenAI describes as “across days or weeks.” That opens the door to things like: “keep trying to get this test green, check back when CI turns over.” Or: “re-run the bench every night, flag when p99 latency drifts more than 10%.” Those were possible before with cron and a bash script. Now they’re one English sentence away.
The part I’m most curious about is reliability. An agent that runs overnight is only useful if it doesn’t wake up, misread its own note, and burn an hour chasing a dead lead. OpenAI hasn’t published reliability numbers for long-running Codex runs, and I’d guess this is where the first round of “my agent burned through API credits doing nothing useful” stories come from.
90+ new plugins, including Jira, CircleCI, and Render
Ninety new plugins sounds like a content-marketing number until you read the list (plugins documentation). The roster pulls Codex into the middle of every dev team’s workflow stack. A few standouts from OpenAI’s own post:
- Atlassian Rovo for Jira management
- CircleCI and GitLab Issues for CI and ticket touch-points
- CodeRabbit for code review
- Microsoft Suite for Outlook, Word, Excel, Teams
- Neon by Databricks for serverless Postgres
- Remotion for programmatic video
- Render for deploy hooks
- Superpowers for reusable agent skills
These aren’t integrations in the “we have an API” sense. They’re MCP servers and skills bundles. Codex can plan across them. You can hand it a Jira ticket and have it read the issue, open the repo, check the CI status on CircleCI, make the change, open a PR, have CodeRabbit review, and redeploy via Render. The individual pieces existed; the end-to-end stitching is what’s new.
For teams already using Cursor or Claude Code for editor work, the plugin list is the argument to at least try Codex for task-level work. Cursor is still sharper inside the editor, but Codex now covers more ground outside it.
Image generation inside the agent
Codex uses gpt-image-1.5 to generate and refine images inside the app. OpenAI pitches it for product concepts, frontend mockups, and game assets. On its own, image-gen-in-an-IDE is a novelty. Paired with computer use, it gets more interesting: Codex can generate a mockup, drop it into a Figma file, compare against the live page in the in-app browser, and iterate, all without leaving the Codex window.
This is the first time I’ve seen OpenAI ship a coding product that treats visual assets as a first-class citizen. Whether it sticks depends on quality. gpt-image-1.5 appears competitive with leading image models in early demos rather than clearly dominant, so the value here is in the workflow integration rather than the raw image quality.
Availability and pricing
OpenAI didn’t announce a price change alongside this update. The new Codex features ship to existing Codex desktop app users signed into ChatGPT, with the usual ChatGPT tier gating. Personalization and memory are rolling out to Enterprise, Edu, EU, and UK users “soon,” per OpenAI’s announcement, with computer use following the same staged path in the EU and UK.
gpt-image-1.5 calls and long-running automations will presumably hit usage limits the same way any other model call does. If you’re running overnight background jobs, the API bill is the thing to watch.
What it means for the AI coding tool fight
For most of 2025, the AI-coding story was: Cursor is winning the editor, Claude Code is winning the terminal, GitHub Copilot is sliding. Codex was present but rarely the first tool devs reached for, even with 3 million weekly users (per OpenAI’s own numbers earlier this year).
This update changes the shape of the competition. Codex is now the first AI coding tool with genuine multi-app, multi-day, background-running capability on a developer’s actual machine. Claude Code is terminal-focused and Cursor is editor-focused; Codex is now competing in a broader category than either of them.
That doesn’t mean it wins. Reliability at these longer horizons is an open question, and the security surface got bigger overnight. But the product category it’s competing in is larger than it was last week, and no one else has shipped a peer.
Risks to weigh before turning it on
Three things worth paying attention to as you decide whether to turn this on.
The blast radius is bigger than the announcement says. Any app Codex can see is an app Codex can mutate. If you use the same machine for work and personal, or if you have production credentials in any menu bar app, the agent’s permissions are now your session’s permissions.
Reliability on long runs is unproven. Overnight tasks that get stuck or misread their own state burn money and time. There’s no public data yet on how often this happens, and the first round of cautionary tales hasn’t landed.
And then there’s geoblocking. If you’re in the EU or UK, the most interesting features are not yours yet. “Gradual rollout” has been a long wait with OpenAI lately. Expect a wait measured in weeks rather than days.
Sources
- OpenAI announcement — “Codex for (almost) everything” — official release notes for the April 16, 2026 update
- Codex plugins documentation — official feature reference for plugin system
- MacRumors coverage — independent reporting on the macOS rollout and EU/UK geoblocking
- Help Net Security analysis — security angle on computer-use expansion
FAQ
What is the new OpenAI Codex computer use feature? It lets the Codex desktop app drive your Mac the same way you would: seeing the screen, clicking, typing, and opening apps. Multiple Codex agents can run in parallel without blocking your own work.
When did OpenAI release the Codex update? April 16, 2026, under the blog post “Codex for (almost) everything.” Rollout to macOS users began the same day; some features are delayed in the EU and UK.
Does Codex computer use work on Windows? Not at launch. OpenAI only announced macOS support. Windows availability has not been confirmed.
How does Codex memory work? Codex now remembers context across sessions: tech stack, preferences, prior corrections, things that took time to gather the first time. It’s labeled a preview, and it’s geoblocked from the EU and UK for now.
Can Codex schedule tasks to run for days? Yes. Automations can resume conversation threads and schedule future work, with OpenAI describing it as work that continues “across days or weeks.”
What plugins ship with the new Codex? Ninety-plus new ones, including Atlassian Rovo, CircleCI, CodeRabbit, GitLab Issues, Microsoft Suite, Neon by Databricks, Remotion, Render, and Superpowers, combining skills, app integrations, and MCP servers.
How does this compare to Cursor and Claude Code? Cursor remains sharper inside the editor and Claude Code remains strong in the terminal. Codex is now in a larger category: multi-app, multi-day, background-capable agents. Incumbents are fine for editor-only work; for multi-app task work, Codex is currently alone in its category.
Bottom line
“Codex for (almost) everything” is the right name for this release. The underlying model didn’t need to change for the scope to expand this much. Last week’s Codex stayed inside your editor; this week’s reads Jira tickets, clicks around in Figma, and schedules its own follow-up runs.
Whether that’s good depends on whether you trust the agent near your actual machine. A throwaway weekend task on a sandboxed account is a fine way to get a feel for it. Production workflows with real credentials are worth waiting on for two weeks while the first round of post-mortems lands. This is the most ambitious thing OpenAI has shipped into a developer’s laptop, and the ambition is also where the failure modes will live. If you want editor-focused tools while you wait, our Claude Code vs Codex CLI breakdown covers the terminal-first option.