On April 16, OpenAI pushed a major update to Codex. This was not a minor tweak — it transformed Codex from a "chat and write code" tool into an agent that can silently operate everything on your desktop in the background. On the same day, Anthropic released Opus 4.7. Both companies made their moves on April 16, making the direct competition in AI coding tools increasingly evident.
Background Computer Use: What's different from previous "computer control"
The core feature of this update is called "background computer use" — Codex can now run in the background on macOS, using its own mouse cursor to operate other applications on your computer without taking over your entire screen.
Many previous Computer Use demonstrations showed the AI taking over your screen while you could only watch it operate. OpenAI chose a different path:
"It can use apps on your computer in the background, as opposed to taking over your entire computer."
The difference is practical. Codex can help you run front-end interface debugging, test applications, and fix bugs that can only be reproduced through the GUI — all in the background while you continue working normally, without interference. It also supports running multiple agents simultaneously.
This is significant for developers who work with many tools that lack APIs, such as legacy testing software, local simulators, and internal systems with only a GUI.
What it can do: A substantial list of features
Beyond computer use, this update packs in a lot:
- Built-in browser: Can open local or public web pages (no login required), leave comments directly on the page, and have Codex make changes based on feedback
- Image generation: Integrates gpt-image-1.5, allowing direct generation of interface mockups and game assets within the development workflow
- 90+ new plugins: Includes JIRA, CircleCI, GitLab, Microsoft Suite, Neon, Remotion, Render, and more, covering nearly the entire engineering pipeline
- SSH access to remote development machines (alpha): Not just local — can operate remote devboxes
- Monitor Slack/Gmail/Notion: Track task progress, handle PR comments, receive notifications
The most distinctive feature is self-scheduling: Codex can schedule tasks to be executed in the future and then start them automatically, without the user being present, theoretically running continuously for days or even weeks.
Comparison with Claude Code Routines
Interestingly, Anthropic also released Claude Code Routines on April 14 (two days earlier), which is also an automation feature that "can run without the user's laptop being open." Both companies are pushing in the same direction at roughly the same time.
The architectural choices differ:
- Claude Code Routines: Runs on Anthropic's cloud infrastructure, triggered by schedules, GitHub events, or APIs. Suitable for code repository-related automation (nightly bug triage, PR review, code style checks)
- Codex background computer use: Runs locally on a Mac in the background, can directly operate local GUI applications
The former is more pure, the latter is more "wild." The ability to operate local software is irreplaceable in some scenarios, but it also means more complex permission and security boundaries.
Who will use it, who doesn't need it
Scenarios for different developers:
Suitable for computer use:
- Testing mobile app simulators with GUI interfaces
- Projects with legacy internal tools that can only be operated with a mouse
- Front-end development requiring frequent visual comparison between design mockups and implementation
Less suitable scenarios:
- Pure back-end services where all APIs are callable
- In the EEA, UK, and Switzerland — this feature is currently not available in these regions
The exclusion of the EU and UK is not surprising. OpenAI has always been slower to roll out new features in the EU compared to other regions, with AI Act regulatory pressure being the main reason. As for when a Windows version will arrive, OpenAI has not provided a timeline.
What this update means
Looking at Codex's development trajectory, it has completed three stages: "code completion → code agent → cross-application agent." An AI development tool that can control the entire desktop environment in the background is no longer the same kind of thing as the Copilot from a year ago that "helped you complete a line of code."
Automatic scheduling, background operation, cross-application control — OpenAI hasn't explicitly stated where this path ends, but the direction is quite clear.
Sources: OpenAI turns Codex into an always-on coding agent that watches your screen (The Decoder); OpenAI drastically updates Codex desktop app to use all other apps on your computer, generate images, CocoLoop, preview webpages (VentureBeat); OpenAI takes aim at Anthropic with beefed-up Codex that gives it more power over your desktop (TechCrunch); Computer Use – Codex app (OpenAI Developers)