OpenAI launches Codex Chrome extension for background tabs and logged-in sites

• OpenAI's launch post says Codex now works directly in Chrome on macOS and Windows, with background work across multiple tabs instead of taking over the active browser window.
• According to OpenAI's tool-routing demo, Codex can mix plugins with direct browser control, using Chrome for logged-in sites and plugins for steps that do not need the browser.
• OpenAIDevs' developer thread frames the extension as browser-side dev tooling, including web app testing, cross-tab context gathering, and parallel DevTools use.
• Early reactions from embirico's comparison and testingcatalog's pre-release screenshot center on the same shift: Codex already had Computer Use, but the extension changes the speed and tab model.
• Real demos got weird fast, because dkundel's multiplayer test used subagents to drive two browser sessions at once while cedric_chee's ATO simulator run showed the extension completing a tax-return workflow inside a government training site.

You can watch OpenAI's main demo run structured browser work across tabs, see OpenAI's follow-up show Codex switching between plugins and Chrome in one task, and check testingcatalog's settings leak for the per-site control surface that appeared before rollout. The launch also came with developer-specific demos: OpenAIDevs showed parallel DevTools use, dkundel paired Chrome control with subagents for multiplayer testing, and cedric_chee promptly aimed it at Australia's tax simulator.

OpenAI positioned the ship as a Chrome plugin inside the Codex app, available on macOS and Windows now, with EU and UK support coming later.

The pre-release settings screenshot from testingcatalog matters because it exposed two implementation details before launch: Chrome sits under Codex's broader "Computer use" controls, and the UI separates general app control from a browser-specific extension connection.

OpenAI's own rollout copy in OpenAI's thread adds the regional caveat and says setup starts by installing the Chrome plugin from the Codex app. Community posts like testingcatalog's rollout clip and reach_vb's install note confirm the extension was appearing to users that same day.

The biggest product change is not "Codex can use a browser." It could already do that through Computer Use. The change, as embirico's comparison put it, is speed and parallelism: background tabs, multiple agents, and no browser takeover.

Across the launch posts, the new tab model breaks down into three concrete behaviors:

• Background execution across tabs, per OpenAI's launch post
• Multi-tab browser work that leaves the user free to keep browsing, per dkundel's note
• Parallel DevTools and cross-tab context gathering for web workflows, per OpenAIDevs' browser dev demo

That is a cleaner fit for browser-heavy tasks than the older single-surface control loop. jxnlco highlighted the DevTools angle, while gdb's post reduced the whole launch to a simpler phrase: Codex can now drive Chrome tabs in the background.

OpenAI's most useful technical detail is in its tool-routing post: Codex does not blindly stay in the browser. It chooses between plugins and Chrome step by step.

OpenAI describes the routing logic in plain terms:

• Use plugins when they can handle the step, per OpenAI's tool-routing post
• Use Chrome when the task needs a logged-in website, per the same post
• Combine both approaches inside one run, again per OpenAI's tool-routing post

That hybrid model explains why the extension looks broader than ordinary browser automation. In OpenAI's browser-work example, the company says Codex writes and runs code to navigate structured pages and complete complex data-entry flows. Wes Roth's recap describes the same split as backend plugins for some actions and direct interaction with a logged-in Chrome session for others.

The developer framing is narrower, and more concrete, than the broader "do work in Chrome" launch copy. OpenAIDevs' thread says the plugin can test web apps, gather context across tabs, use DevTools in parallel, and keep results organized.

Those examples naturally fall into two buckets:

• Single-app testing: web app testing and browser-flow debugging, per OpenAIDevs and OpenAI's use-case list
• Multi-agent browser work: parallel tasks across tabs, including dkundel's multiplayer game demo where subagents control separate browser sessions

OpenAI's use-case list expands the non-dev side of that list to dashboards, research, and CRM updates. Together, the posts make the extension look less like a one-off browser operator and more like a browser surface for the existing Codex agent stack.

The settings leak in testingcatalog's screenshot included one more detail that OpenAI only described briefly in text: site control. The screenshot shows Chrome as a separately connected control surface, and the quoted product copy says users stay in control of which websites Codex can use.

That permission model is easy to miss in the main demo clips, but it is one of the sharper distinctions from generic browser takeover demos. the changelog repost summarizes the launch as "per-site access controls," while testingcatalog's earlier screenshot showed where those controls would likely live.

The fastest way to understand what people think this unlocks is to look at the first demos people ran, because they escalated immediately.

Two stood out:

• daniel_mac8's hands-on post said Chrome for Codex felt fast enough for flight booking, then attached a Google Flights screenshot from a live search.
• cedric_chee's ATO simulator run showed Codex filing a nil-income return in Australia's myTax simulator, including a receipt number and a follow-up list of due and overdue lodgements.

The ATO example also exposed a bit of failure handling. In

, Codex reports that a declaration checkbox resisted interaction and that the browser-control kernel reset during inspection before it reconnected and continued. That is more revealing than a clean marketing demo, because it shows the extension sitting inside the same messy retry loop as other coding agents, only now with a logged-in browser in the loop.