OpenAI Codex Gets Computer Use, Browser, and PR Reviews — Now the Strongest Claude Code Rival

Sottiaux described the strategy in a media briefing covered by.¹

OpenAI shipped what may be the most significant Codex update since the product launched — transforming it from a coding assistant into a full desktop agent platform. The update, rolled out in April 2026, adds background computer use, an in‑app browser, GitHub pull request reviews, image generation, and more than 90 new plugins. Codex now has 3 million weekly developers, according to Yahoo Tech.

The headline feature — Computer Use — lets Codex see the screen, move its own cursor, click, and type inside any Mac application. It operates in the background rather than taking over the user's screen, meaning multiple agent workflows can run simultaneously while the developer continues working. "Codex can actually click on apps, launch apps, and type into apps. This works with any apps on your machine," Andrew Ambrosino, a Codex team technical staffer, told.²

The New Stack's Jessica Wachtel put the updated Codex through its paces on HTTPie, a popular open‑source Python CLI tool, and came away impressed — with caveats. Codex fixed a real GitHub issue in 3 minutes, reading the bug report directly through the in‑app browser, locating the problem across three files, writing a fix, adding a regression test, and explicitly leaving unrelated changes untouched.

"Codex understands not only the task at hand but also the codebase itself," Wachtel wrote for.³ She noted Codex demonstrated awareness of both the immediate task and the broader project context — it wasn't just blindly editing files.

The in‑app browser, implemented as a plugin, proved particularly useful. Instead of copying a GitHub issue into chat, Wachtel opened the issue directly inside Codex and told it to read and fix the bug. Codex handled the entire workflow in split‑screen, side‑by‑side with the chat interface.

The computer use feature showed real limitations. When Wachtel asked Codex to use Terminal.app through computer use, it refused outright — flagging Terminal.app as blocked from computer use in that session, citing security reasons. It completed the task using its built‑in shell instead.

"A coding agent with unrestricted terminal access is a security risk," Wachtel noted, viewing the limitation as a reasonable safety measure rather than a flaw. But another issue kept recurring: the sandboxed environment blocked the full test suite due to port binding restrictions. "This keeps showing up across every tool I have tested, and it is a real limitation of sandboxed agent environments," she wrote.

Codex head Thibault Sottiaux told:² "It's not just about the growth. It is putting a very capable agent in the hands of builders, and now we're seeing that we're able to expand and do a lot more work entirely across your computer."

OpenAI is positioning the update as "Codex for (almost) everything," The New Stack reported expanding well beyond code editing into a general‑purpose agent workspace. The in‑app browser currently focuses on frontend and game development workflows, with broader browser control planned over time. ¹ described the features as pieces of a future "super app" combining Codex, OpenAI's Atlas browser, and other agentic tools.

Wachtel's verdict, published in:³ "Codex is now the most complete alternative to Claude Code," she wrote. But the two tools embody different philosophies. Claude Code is terminal‑based — an agent that reads codebases, edits files, runs tests, and commits to GitHub from the command line. Codex is becoming a desktop workspace — a unified environment that combines computer control, browsing, image generation, and coding under one roof.

Claude Code added its own computer use feature in March as a research preview for Pro and Max subscribers on macOS, making the feature race between the two companies increasingly symmetrical. The key difference: Claude Code's computer use operates in the terminal context, while Codex's operates as a desktop‑level agent with its own cursor.

The OpenClaw story adds another dimension. Peter Steinberger, the Austrian developer who built the viral open‑source agent framework (60,000 GitHub stars in 72 hours), joined OpenAI in February after being contacted by Sam Altman, Mark Zuckerberg, and Satya Nadella. Anthropic had previously sent Steinberger a trademark complaint over the framework's original name, "Clawdbot" — a move observers believe accelerated his move to OpenAI, Yahoo Tech reported.

The practical takeaway for builders: Codex is now a legitimate daily‑driver coding agent that can handle real production workflows — not just toy examples. The in‑app browser alone changes the workflow: instead of copy‑pasting issues and context between tools, developers can keep everything inside Codex.

But the sandbox limitations are real, and computer use on Windows doesn't yet match the macOS experience. For developers whose workflows depend heavily on terminal access and unrestricted local environments, Claude Code's terminal‑native approach may still feel more natural. For everyone else — especially those who want a single workspace for coding, browsing, testing, and reviewing — Codex's expansion makes it hard to ignore.