ChatGPT can now read your open Chrome tabs and fill forms in them. Good for one-off tasks, not for recurring processes
ChatGPT can now see the tabs you already have open and act inside them.
What happened
The Verge reports that OpenAI's ChatGPT Atlas browser is gone, and ChatGPT's Chrome extension has picked up its job. With the extension installed, ChatGPT can view your open tabs when you tag Chrome in the ChatGPT web or desktop app. The sidebar now carries your full conversation history. In Work mode, ChatGPT can control your Chrome tabs for tasks like sending messages, filling out forms and downloading files, and it can organize or group tabs.
The Decoder adds that users target tabs with @chrome, Chat mode can pull data from several tabs at once, and Work mode uses Computer Use for the actions. It says the extension works on all ChatGPT plans. OpenAI's examples include putting vendors from several tabs into a comparison sheet with pricing and source links, and filling a registration form from an event brief while stopping before submitting.
My take
This is a real time saver for one-off work. Comparing five vendor pages, pulling details into a sheet, drafting the follow up email. Those are tedious chores it can now handle.
But it is easy to mistake this for automation. A few lines I would draw for clients:
- If a task happens every week, build it as a workflow with APIs, not a browser agent session. It will be faster, cheaper and it will not break when a page layout changes.
- Tabs are logged in as you. Whatever you have open in your CRM, bank or inbox is in reach, so keep sensitive tabs closed when Work mode is on.
- Keep the "stop before submitting" habit from OpenAI's own example. Let it fill, then you click send.
My rule of thumb: use the browser agent to discover the steps of a task, then turn the ones that repeat into a proper automation.
More posts
- MiniMax M3.1 Flash costs $0.15 in and $0.60 out per million tokens, but its hallucination rate is 93%. Give it a way to say unknownOct 20, 2026
- Attackers aimed cheap AI agents at mid-sized online stores. Successful breaches took about 23 hoursOct 20, 2026
- n8n ran a full AI stack on one NVIDIA DGX Spark with no cloud API. Invoice extraction took under a second with thinking offOct 20, 2026
