Codex uses Chrome on desktop

Browser access is the important part of this story — not because Chrome is exciting, but because real work usually happens inside messy, logged-in web software. On May 7, OpenAI shipped a Chrome extension for Codex that connects the desktop app to Chrome on macOS and Windows. That means Codex can now work inside your normal browser session instead of a separate sandbox browser. The practical change is simple: the agent can finally reach the tabs, cookies, and extensions that a lot of actual workflows depend on. ### What actually changed in Codex? Before this, Codex already had an in-app browser and broader “computer use” features in the desktop app. But the in-app browser had a big limitation — it did not support authentication flows, signed-in pages, your regular browser profile, cookies, extensions, or your existing tabs. OpenAI’s own docs basically draw the line: if a task depends on login state, use your regular browser or the new Chrome extension. (developers.openai.com) ### Why is Chrome the big unlock? Because a lot of software is browser software now, and a lot of that software is only useful once you are signed in. Think internal dashboards, admin consoles, SaaS tools, staging environments, bug trackers, CRM pages, and random legacy portals nobody bothered to give a clean API. A browser agent that only sees a fresh, isolated page is like sending a coworker to your office but not giving them your badge. (openai.com) The new extension lets Codex use the browser state you already have. ### Does Codex take over your browser? Not in the old “watch the cursor move and hope” way. OpenAI says the Chrome setup works across tabs in the background without taking over the browser, and the user stays in control of which websites Codex can use. That is a meaningful design choice. It turns the browser from a full remote-control surface into something closer to a supervised work channel. (developers.openai.com) ### What can it do there? OpenAI frames Codex as a coding agent, so the first-party pitch is web-app testing, gathering context across tabs, and using browser tools more efficiently while keeping results organized. Outside coverage also points to signed-in workflows in tools like Gmail, Salesforce, LinkedIn, and internal apps. The pattern is the same either way — Codex is moving from code-only work toward the browser-heavy tasks wrapped around software work. (developers.openai.com) ### Why does that matter beyond coding? Because the browser is where the long tail lives. APIs cover the clean, planned paths. Human workers handle the ugly rest — portals with weird forms, admin pages built a decade ago, systems that expose data but not actions. A browser-capable agent can bridge that gap. That does not magically make every workflow safe or reliable, but it makes supervised automation much more plausible in the places where structured integrations never showed up. (macrumors.com) This is an inference from the product design and the kinds of sites OpenAI highlights. ### What is the catch? The catch is trust. OpenAI’s setup docs explicitly warn users to treat page content as untrusted context and to review the website before allowing Codex to continue. That warning tells you exactly where the risk lives — prompt injection, bad page instructions, accidental actions, and over-broad access once an agent can see what your browser sees. The useful work and the dangerous work are now much closer together. (developers.openai.com) ### Where does this fit in OpenAI’s bigger plan? It fits a clear direction: Codex is becoming a command center for multi-agent work across local tools, cloud tasks, plugins, and now the everyday browser. OpenAI has been adding computer use, plugins, memory, and broader app support over the past few weeks. Chrome is another step toward one agent that can move across the actual surfaces where developers work. (developers.openai.com) ### Bottom line? This is not just “Codex now supports Chrome.” It is OpenAI admitting where the missing context really was. For many tasks, the hard part was never writing code — it was getting the agent into the same logged-in browser reality as the human. (developers.openai.com) (openai.com)