Booking a flight should take five minutes. Instead you dig out the airline’s password, fail it twice, and end up on hold. The new ChatGPT cloud browser exists to kill that whole loop, and it does it without ever seeing your password.
What is the ChatGPT cloud browser?
OpenAI rolled this out on August 26, and it’s part of ChatGPT Work, the task-focused side of ChatGPT that runs on the $20-a-month Plus plan and up. Free users and the cheaper Go tier don’t get it yet. That’s the first thing to know, because most people hear “ChatGPT can do things for you” and assume it’s free. It isn’t.
Worth a quick note on what ChatGPT Work is, since it keeps confusing people. OpenAI announced it in July, and it’s actually two products wearing one name. There’s Work Cloud, which runs on chatgpt.com and in the mobile apps, and Work Local, which runs on your computer through the desktop app and can touch your files. The cloud browser belongs to Work Cloud. So does a persistent filesystem, access to the GPT-5.6 family of models, and the ability to schedule prompts to run on their own.
So what does the ChatGPT cloud browser actually do? Think of it as a remote browser that ChatGPT controls. OpenAI’s help page describes it as able to “read web pages, click buttons, enter information into forms, and carry out steps on supported public and signed-in websites.” You give it a task, it opens the site, and works through it the way you would, one step at a time, with a running commentary of what it’s doing.
What can ChatGPT actually do for you?
The official examples are the fun part. ChatGPT can now set up utilities for your home, book a passport or a medical appointment, buy or cancel plane tickets, and rearrange a package’s arrival. Checking insurance policies? Also on the list. These aren’t the cute “write me a poem” tricks. These are the chores that quietly eat your Saturday morning.
Here’s a detail that matters: before anything that could be hard to reverse, ChatGPT stops and asks for your explicit confirmation. That includes confirming a booking or making a payment. You’re not giving it a blank check. It’s more like a careful assistant who shows you the paperwork before signing.
It also pairs with the new ChatGPT Work plugins for Gmail and Calendar that landed the same week. So the same system can read an email thread, pull the details, and act on them, instead of just summarizing what’s waiting in your inbox. The Mashable writeup of the launch has the official examples if you want to read them in OpenAI’s own words.
Three chores worth handing over
Concrete prompts help more than abstract promises, so here are three to steal.
Reschedule a delivery. “Reschedule my package from FedEx for Thursday. Ask me before confirming.” ChatGPT opens the tracking page, finds the delivery, and walks through the reschedule flow. You approve the final step.
Cancel a subscription you can’t find. Half the battle with cancellations is locating the account page. “Find my subscription to [service] and cancel it.” The cloud browser digs through the site while you watch.
Book an appointment that only opens at noon. Timing-sensitive bookings are the worst. “Book the earliest passport appointment available in my city, and tell me what you picked before you finish.”
Each of these has the same shape: a task, a deadline, and a confirmation gate before anything irreversible. That pattern is where the ChatGPT cloud browser shines, because the chore is boring but the stakes are real.
How to set up the ChatGPT cloud browser
Setup is surprisingly short. You don’t need to install anything extra.
- Open ChatGPT on the web or the mobile app with a Plus, Pro, or Business account.
- Switch to the Work tab. That’s where the task-based tools live.
- Write your task as a normal prompt. Something like “reschedule my delivery for Thursday” or “cancel my hotel booking.”
- ChatGPT will pause and open a secure form for the website. Log in there with your username and password. This is the key part: you’re typing into the website’s login form, not into the chat.
- Approve the website if asked, and complete any two-factor authentication step.
- ChatGPT resumes, does the task, and reports back with what changed.
The first time feels weird. Typing a password into a browser window you can’t see is a strange muscle memory to build. But it’s exactly what makes the whole thing safe.
Pick your permission level
You control which sites ChatGPT can use from Settings, under Cloud browser. There are three options: Always ask, Auto approve, and Always allow.
Always ask is the sensible default, and honestly the only one I’d recommend for anything involving money or personal data. Auto approve is for sites you trust completely, like a work portal. Always allow is the “I really want this to be hands-free” option, and it’s the one that should make you think twice.
Wait, it doesn’t see my password?
That’s the claim, and the mechanics back it up. OpenAI says the username and password you enter in that secure form are not visible to the model, and ChatGPT does not store those sign-in credentials. You’re authenticating with the website through a remote browser, while the language model just gets to use the session that results.
On top of that, OpenAI runs every sign-in request through an extra model that looks for signs of phishing or deception. So there’s a second set of eyes checking whether the login page you’re being asked to use is actually legit. Simon Willison, who’s been tearing ChatGPT Work apart since July, spelled out the architecture in his explainer, and it checks out.
One warning while we’re here: this is different from typing “my password is hunter2” into a chat. That remains a terrible idea and always will be. The ChatGPT cloud browser keeps your credentials out of the conversation, but only if you use it the way it’s designed. Passwords go in the secure form, never in the prompt.
The limits you should know about
It’s new, so the edges are visible. Some websites restrict access from AI agents entirely, and you won’t know until you try. A few sites will load but behave oddly in an automated browser. Anything requiring unusual steps, captchas with audio, weird banking flows, those will trip it up.
There’s also the subscription wall. Plus starts at $20 a month, and the full Work experience with the cloud browser lives behind it. If you’re on the free plan, none of this applies yet.
And let’s not pretend automation never misfires. If you’ve been following OpenAI’s own agent struggles this month, you know that even their systems have had incidents. That’s exactly why the confirmation prompts exist, and why you shouldn’t set everything to Always allow on day one.
If you’ve read about AI browser agents before, you’ll notice the overlap. The difference here is that everything lives inside ChatGPT, with these permissions and confirmation prompts built in. That makes it far more approachable for non-technical people than running a separate agent tool. Our explainer on AI browser agents covers how the standalone versions work, so you can compare.
Should you turn it on?
Yes, but start small. Pick one repetitive task, a delivery reschedule or a form you hate filling out, and run it through the ChatGPT cloud browser once. See how it feels. Keep the permission level on Always ask until you’ve watched it work a few times.
The pitch here isn’t that ChatGPT replaces your judgment. It’s that you stop being the one who logs into five websites to shuffle a delivery date. That’s the trade worth making, and it’s finally available to normal people, not just developers.
For a broader look at what ChatGPT Work offers beyond the browser, our ChatGPT Work guide covers the whole product. And if you want the developer-flavored version of this idea, the Codex browser automation workflows are worth a read.
