Table of Contents
- What Actually Shipped on April 16
- Background Computer Use
- The In-App Browser
- What This Does to the Codex–Cursor Comparison
- How to Actually Use It
- Where Codex Still Doesn’t Fit
- FAQ
OpenAI ships product updates by the dozen. On April 16, 2026, it shipped one that quietly turned Codex from a coding assistant into something different: an agent workspace that operates your computer. The release was titled “Codex for (almost) everything,” and the framing was honest. Codex still writes code. It now also opens apps, drives a browser, edits documents, and runs multi-step background tasks on the Mac you’re already using.
Roughly 3 million developers were already using Codex weekly before the update. The new release is what those developers have been asking for since the original Cursor versus Codex comparisons of 2024.
What Actually Shipped on April 16
The desktop Codex app became a full agentic workspace with three new capability blocks layered on top of the existing coding interface.
Background computer use. Codex agents can now see, click, and type with their own cursor across every app on your Mac. Multiple agents work on your machine in parallel without interfering with your own work in other apps.
In-app browser. A dedicated browser pane lives inside Codex. You can comment directly on web pages to give the agent precise instructions, useful for frontend development, design review, and game development.
Image generation, persistent preferences, repeatable background tasks. The smaller features that make the agent practical rather than impressive — Codex remembers what you’ve told it to do across sessions, runs the same task on a schedule, and produces images when you ask without leaving the app.
Each capability is incremental in isolation. Together, they make Codex the answer to the question “what does the productivity layer look like when an agent has the same OS-level access a human user has?”
Background Computer Use
The structural change is parallel agents operating your computer without blocking you. You can ask Codex to run an analysis against three different files in three different apps, and three separate agents handle the work concurrently while you do something else. The agents have their own cursors. They don’t fight with your input. The OS-level coordination is the engineering work that makes the parallel pattern actually usable.
Compare this to Cursor 3’s Agents Window. Cursor 3 also runs agents in parallel, but the agents work inside the Cursor environment — local, in worktrees, in the cloud, or over SSH. Codex’s background computer use is the more aggressive version: agents work everywhere on your Mac, including in apps Codex didn’t build. The tradeoff is risk. An agent that can drive every app you have access to is an agent that can break things in places Codex doesn’t directly understand.
The early-user pattern: developers run background computer use against very specific, repetitive workflows — file conversions, data entry into legacy apps, scheduled report generation. The reliability for those bounded tasks is solid. Open-ended exploratory work where the agent has to figure out the environment from scratch is still hit-or-miss.
The In-App Browser
The in-app browser is the smaller-headline feature that solves the largest day-to-day friction. Frontend developers and designers spend most of their day moving between an editor, a browser running the local dev server, and a feedback channel where stakeholders comment on what they see. Codex’s browser pane collapses that workflow.
You view the running app inside Codex. You click directly on the element you want to change. You type the change you want — “make this button blue and bigger” — and Codex acts on the targeted element. The pointing-at-the-thing primitive is the actual feature. Previous coding agents required you to describe the change in language. Codex lets you point.
The same primitive applies to design review (point at the broken element on a deployed page), game development (point at the visual asset that needs adjusting), and any frontend bug report (point at the visual symptom). It works exactly as well as you’d hope and saves more time than the announcement frames it as saving.
What This Does to the Codex–Cursor Comparison
Codex and Cursor 3 are now competing on the same primitive: agents-as-default with the IDE demoted. The difference is the scope of the agent’s environment.
Cursor 3 keeps agents inside Cursor and its connected environments — worktrees, cloud, SSH. Codex pushes agents into the entire OS. For developers who want their agent to operate exclusively inside the coding loop, Cursor’s scoping is the safer default. For developers who want their agent to handle anything they could do on their computer, Codex’s broader scope is the right answer.
Neither product is the answer for every developer. Both products are converging on the same broader idea, which is that the IDE is no longer the primary surface — the agent is. That convergence is the structural shift in 2026 coding tools and Codex’s April 16 release is one of the clearest expressions of it.
How to Actually Use It
Update Codex to the April 16 release or later. The new agent capabilities are on by default for ChatGPT Plus, Pro, Business, and Enterprise users. The in-app browser opens with a keyboard shortcut from any Codex view. Background computer use is enabled per-task — Codex asks before it starts driving your apps.
Do this first: pick one repetitive task you do every week that involves operating an app other than your editor. File conversion. Data entry into a legacy interface. Pulling reports from a web dashboard into a spreadsheet. Tell Codex to do it once with your supervision. If it works, schedule it. The schedule-it-and-walk-away pattern is what makes background computer use worth the trust calibration.
Skip the parallel-agents pattern, at first, on any task where you cannot tolerate the agent making a wrong choice in an app you don’t actively monitor. Single-agent supervised use first. Parallel use after you’ve calibrated trust on your specific workflows.
Where Codex Still Doesn’t Fit
Three real limits.
Codex is Mac-only for the background computer use feature. Windows and Linux developers get the in-app browser and the coding upgrades but not the full agentic OS layer. OpenAI has not announced a timeline for Windows or Linux parity.
The security model for an agent with OS-level access is still being figured out. Codex requests permission per-task and shows what it’s doing, but enterprise security teams reviewing the capability will reasonably ask harder questions about what guardrails exist when an agent can drive any app on the device. The same questions apply to Anthropic’s Orbit phone agent in a different form.
The reliability on unfamiliar apps varies. Codex handles common workflows reliably. Pointed at a niche desktop application it has limited training data for, it fails in unpredictable ways. Watch the agent the first few times you use it on any new surface.
FAQ
What is Codex “for (almost) everything”?
A major Codex update shipped on April 16, 2026, turning the desktop Codex app into a full agentic workspace. New capabilities include background computer use (multiple agents driving apps on your Mac in parallel), an in-app browser for visual feedback, image generation, persistent preferences, and repeatable background tasks.
Does Codex run on Windows or Linux?
The April 16 background computer use feature is Mac-only. Windows and Linux versions of Codex receive the in-app browser and the coding-side upgrades, but not the full OS-level agent capabilities. OpenAI has not announced a timeline for Windows or Linux parity on background computer use.
How does Codex compare to Cursor 3?
Both products are converging on agent-first development surfaces. Cursor 3 scopes agents to its own connected environments (local, worktrees, cloud, SSH). Codex pushes agents into the entire macOS environment. For coding-only workflows, Cursor’s scoping is the safer default. For workflows that span coding plus other apps, Codex’s broader scope wins.
Is background computer use safe?
Codex asks for permission per task and shows what it’s doing. For supervised use on familiar workflows, it works well. For unsupervised parallel use, calibrate trust gradually — pick one task, watch the agent execute, then schedule it once you trust the outcome. Enterprise deployment should pass through a security review of the broader Mac-app-access model first.
What does the in-app browser do?
A dedicated browser pane inside Codex lets you view a running web app and comment on specific page elements. Click an element, type the change you want, and Codex acts on the targeted element directly. Useful for frontend development, design review, game development, and any workflow where pointing at the thing is faster than describing it.
Will OpenAI keep adding to Codex?
Yes. OpenAI has been shipping rapid Codex updates since launch, and the April 16 release expands the surface area of what Codex agents do. Expect continued integration with other OpenAI products (Realtime voice, GPT-5.5, Codex Cloud) and broader OS coverage over the next 6 to 12 months.