OpenAI Codex Computer Use Review: Background Agents That Control Your Entire Mac

OpenAI Codex Computer Use: Background Agents, 90 Plugins, and the Super App Push

OpenAI just shipped the biggest Codex update since the desktop app launched in February 2026. The headline feature is "Computer Use," which lets Codex see, click, and type across every application on your Mac, all while you keep working. Multiple agents run in parallel in the background. There is also a built-in browser, gpt-image-1.5 integration, 90 new plugins, and scheduled task execution.

This is not a coding tool update. This is OpenAI building a super app out in the open, as Codex lead Thibault Sottiaux admitted directly: "We're actually doing the sneaky thing where we're building the super app out in the open and evolving it out of Codex."

Here is what the update actually does, where it falls short, and whether it changes the competitive landscape against Claude Code, Cursor, and the rest.

What Computer Use Actually Does

Computer Use is the defining feature of this release. On macOS, Codex can now operate a separate cursor that clicks, types, and navigates through any application installed on your machine. This is not screen sharing. This is not remote desktop. Codex spawns its own interaction layer that runs concurrently with your normal workflow.

You can be writing an email in one window while Codex triages a Jira ticket in another. You can be debugging in your IDE while Codex tests a frontend change in the browser. The agents run independently, each with their own cursor and context.

Andrew Ambrosino from the Codex team put it plainly during the press briefing: "Codex can actually click on apps, launch apps, and type into apps. This works with any apps on your machine."

How it compares to Anthropic's approach

Anthropic launched Claude Cowork and redesigned the Claude Code desktop app earlier this year. Both Anthropic and OpenAI are chasing the same goal: an AI agent that lives on your desktop and handles work across your entire toolchain. The difference is execution.

Claude Code can interact with your system, but it takes over your screen to do it. When Claude is clicking around, you wait. Codex runs in the background. That is a meaningful UX distinction. If background execution actually works reliably (and early reports suggest it does on macOS), OpenAI has a genuine edge for developers who cannot afford to context-switch every time an agent needs to do something.

Windows users are stuck with the older model for now. Codex can still pull information from Windows apps and surface it in the chat, but the background cursor-level interaction is macOS only at launch. OpenAI has not given a timeline for Windows parity.

The Built-In Browser

Codex now has its own browser. This is more useful than it sounds. Previously, asking an AI agent to iterate on a frontend design meant a clunky loop: agent makes changes, you check them in your browser, you report back what needs fixing. Now Codex renders the page internally and you can leave comments directly on DOM elements with instructions.

The workflow looks like this: you ask Codex to build a landing page, it generates the code and renders it in the in-app browser, you click on the header and type "make this darker," and Codex applies the change. No screenshotting. No copying CSS values back and forth. The feedback loop tightens from minutes to seconds.

OpenAI says the browser currently works with localhost web applications. The plan is to expand it to general web browsing over time. That expansion path is telling. Once Codex can browse the full web, the line between "coding tool" and "general-purpose AI assistant" gets very blurry.

Scheduled Tasks: The Heartbeat Feature

Codex can now schedule work for the future. You can tell it to run your test suite every morning at 9 AM. You can ask it to check for dependency updates on Fridays. You can set it to review a pull request after a teammate pushes changes next week. The agent wakes itself up, does the work, and presents the results.

Caffrey Lynch from OpenAI's developer product communications framed it as a productivity multiplier: agents that work on your schedule rather than demanding your attention in real time.

This is the feature that could make Codex genuinely sticky. Most AI coding tools are reactive. You ask, they respond. Scheduled tasks make Codex proactive. If the execution holds up over days and weeks (not just in demos), this changes the relationship from "tool I use" to "coworker I rely on."

90 New Plugins

The plugin count is where the "super app" ambition becomes obvious. OpenAI added 90 new integrations in this release. The list includes CircleCI, GitLab, Microsoft Suite, Slack, Gmail, Google Calendar, Notion, and dozens more.

The practical impact: you can type "check Slack, Gmail, Calendar, and Notion and tell me what needs my attention" and Codex pulls from all four apps in a single prompt. You can @ mention specific apps to direct the agent, or let Codex figure out which tools to use based on your request.

For developers, the CircleCI and GitLab integrations mean Codex can monitor CI pipelines and act on failures without you switching context. For non-developers (and OpenAI is clearly thinking beyond developers here), the Microsoft Suite integration means Codex can draft documents, update spreadsheets, and manage email.

gpt-image-1.5 Integration

Codex can now generate images using gpt-image-1.5 directly inside projects. This is positioned as a developer convenience: generate mockup assets, iterate on game sprites, create presentation graphics without leaving the environment. In the press demo, OpenAI showed Codex generating hundreds of consistent game assets for a playable PC game.

The image generation is competent. gpt-image-1.5 produces good results for UI mockups and illustrative assets. It is not replacing Midjourney for creative work, but for the "I need a quick icon" or "generate placeholder images for this prototype" use case, having it built in is convenient.

Multiple Terminal Tabs and SSH

Developers get multi-tab terminal support and an alpha SSH feature for connecting to remote devboxes. This is table stakes for a tool that wants to be your development environment, but it is worth noting because it shows OpenAI taking infrastructure seriously. If you are running a cloud dev environment on EC2 or a Codespace, Codex can now SSH in and work directly on the remote machine.

The SSH feature is labeled alpha, so expect rough edges. But the direction is clear: Codex wants to be where your code lives, whether that is local or remote.

3 Million Weekly Developers

OpenAI confirmed that Codex has reached 3 million weekly developers. That number matters for context. When the Codex app launched in February 2026, it was a Mac-only coding assistant with worktree support and a sandbox. In two months, it has grown into a multi-agent desktop environment that controls your entire operating system.

Growth at that scale puts pressure on the competition. Anthropic, Cursor, Windsurf, and Augment Code all need an answer for background computer use. Right now, none of them have one.

What Is Missing

The update is impressive on paper. In practice, there are gaps.

Windows second-class citizenship. Background computer use is macOS only. Windows users can read data from apps but cannot get the parallel agent experience. Given that a large chunk of the enterprise developer market runs Windows, this is a significant limitation.

Reliability questions. Computer use at the OS level is hard. Clicking the right button in an app with a dynamic UI is harder than generating code. If the agent misclicks, it can do real damage to your files or send something you did not intend. OpenAI has not shared details about safety guardrails beyond "it runs in the background."

Pricing and access. OpenAI has not clarified whether computer use requires a specific subscription tier. If it is locked behind a Pro or Team plan, that limits adoption among the indie developers and students who are the earliest adopters of AI coding tools.

The "super app" problem. Building a super app is notoriously difficult. Slack tried it. Microsoft Teams is trying it. The risk for Codex is that it becomes a jack of all trades that masters none. Developers chose Codex because it was good at coding. If OpenAI spreads focus across email triage, calendar management, and image generation, the core coding experience could degrade.

Competitive Positioning

The table tells the story. Codex is pushing ahead on breadth of capability. Claude Code and Cursor are competing on depth of coding experience. Different strategies for different users.

If you need an AI agent that handles your entire workflow across multiple applications, Codex is the only option right now. If you want the best code generation and editing experience, the competition is closer than the feature gap suggests.

Who Should Use This Update

Full-stack developers working on frontend-heavy projects will benefit most. The browser preview plus computer use means you can iterate on UI changes without ever leaving Codex. The feedback loop is genuinely faster.

Team leads and tech leads who spend time in Jira, Slack, GitHub, and email will get value from the cross-app integration. Asking one agent to triage notifications across four platforms saves real time.

Solo developers building SaaS products should pay attention to the scheduled tasks feature. Setting up automated testing, dependency checks, and PR reviews that run on their own schedule is a meaningful productivity gain for someone who is the entire engineering team.

Windows developers should wait. The core Codex experience is solid on Windows, but you are missing the headline feature of this update. Check back when background computer use lands on Windows.

The Bigger Picture

The subtext of this update is OpenAI's strategic direction. Codex started as a coding tool. It is becoming an operating system layer. The blog post title is literally "Codex for Almost Everything." That is not subtle.

The company is betting that developers will prefer a single AI environment that handles coding, communication, project management, and content creation over separate specialized tools for each task. It is a reasonable bet. Context switching is expensive. If Codex can reduce the number of apps a developer needs to touch in a day, that is real value.

The risk is execution. Super apps tend to accumulate features faster than they can polish them. The coding experience that attracted 3 million developers needs to stay sharp even as Codex adds email management and image generation. If OpenAI can pull off that balance, Codex becomes the default developer environment. If they cannot, a focused competitor will eat the coding side of the market.

Bottom Line

The April 2026 Codex update is the most ambitious feature drop from any AI coding tool this year. Background computer use on macOS is a genuine technical achievement. The 90 plugins, scheduled tasks, and built-in browser make Codex feel less like an IDE extension and more like a workstation.

Whether that ambition translates into daily utility depends on your workflow. If you live in macOS and juggle multiple applications throughout the day, this update is worth installing immediately. If your work is mostly in a single IDE on Windows, the value proposition is weaker today but will likely improve.

OpenAI is moving fast. The gap between "AI that writes code" and "AI that runs your computer" is closing. This update is the biggest step toward that gap closing yet.

Sources:

• •Ars Technica: New Codex features include the ability to use your computer in the background (April 16, 2026)
• •VentureBeat: OpenAI drastically updates Codex desktop app to use all other apps on your computer (April 16, 2026)
• •Neowin: OpenAI expands Codex beyond coding with computer use, memory, and plugins (April 18, 2026)
• •9to5Mac: OpenAI's Codex Mac app adds three key features that go beyond agentic coding (April 16, 2026)