OpenAI ships Codex Chrome extension

OpenAI just pushed Codex one step closer to being a real browser coworker, not just a coding agent that lives in a sandbox. The new Chrome extension, released on May 7, lets Codex work directly inside Chrome on macOS and Windows, including on sites where you’re already signed in. That matters because a lot of real software work now happens in the browser — testing web apps, checking dashboards, reproducing bugs, updating records, and poking around DevTools. The gap before this was obvious: Codex could write code and run tasks in controlled environments, but it was still clumsy around the messy, logged-in web. (developers.openai.com) ### What actually shipped? A Chrome extension called Codex showed up in the Chrome Web Store with OpenAI as the publisher and version 1.1.4 dated May 7, 2026. OpenAI’s own changelog calls the release “Codex for Chrome” and says the point is to make Codex better at working with apps and websites in your browser. This is not a separate browser — it is a bridge between the Codex app and your existing Chrome session. (chromewebstore.google.com) ### Why does Chrome matter so much? Because the hard part of modern software work is often not writing the code — it is dealing with the environment around the code. A bug might only show up after login. A workflow might depend on account state, cookies, browser history, or a weird sequence across several tabs. OpenAI’s doc(chromewebstore.google.com)e in-app browser for localhost or public pages that do not need login. Basically, Chrome is for the real web; the in-app browser is for safer, cleaner testing. (developers.openai.com) ### What can Codex do in the browser? The short version is — more than just inspect a page. OpenAI says Codex can read or act on sites like LinkedIn, Salesforce, Gmail, and internal tools. The Chrome Web Store listing adds research, dashboard review, form filling, multi-step workflows, website testing, and tab cleanup. Other coverage notes that it ca(developers.openai.com)will notice first. That turns the browser into something closer to an active test bench. (developers.openai.com) ### Does it take over your whole browser? Not exactly — and OpenAI is making that distinction on purpose. The extension works in task-specific tab groups and runs in parallel “without taking over your browser,” which means Codex can do background work while you keep using Chrome normally. Think of it less like remote control of your whole desktop and(developers.openai.com)tive, it can stop and hand control back to you. (developers.openai.com) ### So what’s the catch? The catch is permissions. OpenAI’s docs warn that page content should be treated as untrusted context. Codex asks before interacting with new websites, before using browser history, and before downloading or uploading files. The docs also flag “always allow browser content” and browser-history access as elevated-risk settings, because h(developers.openai.com)es. That is the real tradeoff here — convenience versus exposure. (developers.openai.com) ### How does this fit into the bigger Codex push? It fits cleanly. Codex started as a cloud-based software engineering agent in 2025. In April 2026, OpenAI broadened it into a more general workspace with computer use, an in-app browser, memory, and a pile of plugins. Chrome is the next logical step. First Codex could work on your repo. Then it could (developers.openai.com)d — inside the browser. (openai.com) ### Why should developers care? Because browser work is where AI agents usually get awkward. They can write a Playwright test, sure, but the real headache is reproducing the bug on the actual page, under the actual account, with the actual state. This extension narrows that gap. It will not remove the need for review — if anything, it makes review more important — but it does make Codex more useful on the last mile of web work. (developers.openai.com) ### Bottom line? This is a small release with big implications. OpenAI is turning Codex from a coding agent into a browser-native work agent, one permission prompt at a time. That is powerful — but the closer AI gets to your real tabs, real accounts, and real history, the more the product question stops being “can it do this?” and becomes “when should you let it?” (developers.openai.com)