I Let Claude Click Around Without Me
Ten minutes, one safety toggle, nothing scary, promise.
Quick take
- Claude in Chrome, Anthropic's browser extension, went generally available on every paid Claude plan on August 26, 2026.
- The big change: Claude can now take actions in your browser on its own, instead of asking your permission before every single click.
- A safety classifier checks each action before Claude takes it, and stronger defenses were built in against prompt injection, where a sketchy website tries to hijack what your AI is doing.
- You still get a say. One settings toggle puts Claude back in ask-before-every-action mode, and team admins can restrict it to approved domains only.
Watch me try it and flip the safety toggle
Watch me try it and flip the safety toggle in under two minutes, or read the steps below. Both get you to the same place.
I am, historically, not a hand-the-keys-over kind of person. So when I heard Claude's Chrome extension could now click around my browser without asking me first, my gut reaction was a flat no.
Then I actually read how they built it, and I changed my mind enough to try it, carefully.
What Changed When Claude in Chrome Went Generally Available
On August 26, 2026, Claude in Chrome went generally available on every paid Claude plan. The headline change is that Claude can now take actions on its own instead of stopping to ask permission before every single click. Before you panic, two things ride along with that. A safety classifier checks each action before Claude takes it, and Anthropic built in stronger defenses against prompt injection, the trick where a sketchy website tries to hijack what your AI is doing.
You still get a say. There is a settings toggle that puts Claude back in ask-before-every-action mode, and if you run a team, admins can restrict autonomous actions to an approved list of domains only.
How to Try Autonomous Claude Without Handing Over the Keychain
- Install Claude in Chrome from the Chrome Web Store, if you have not already. It is generally available now on any paid Claude plan, and it opens as a sidebar right inside your browser.
- Open the sidebar on a page you actually need help with. Give it one plain-language task: research a topic across a few tabs, summarize a page, pull information into a doc.
- Check your settings before you trust it with anything real. Autonomous action is now the default, so if you would rather watch Claude ask before every click, switch that off in Settings.
- Start with something small and low stakes the first time. Research and summarizing are a much better first test than anything involving a payment field or a password.
- Running a team? Lock it down first. In Organization Settings, admins can restrict autonomous actions to an approved list of domains before anyone else starts using it.
Try It Yourself
Do Steps 1 and 2 only. Install the extension if you have not already, and give it one small research task, nothing involving passwords or payment info. That's the whole test.
This is Recoded in real time, uncomfortable part included. Handing off real work to AI was never going to feel entirely comfortable on day one. Recoded was never about blind trust, it is about knowing exactly how much rope you are giving something and why.
Here are the step-by-step instructions.
If figuring out how much autonomy to actually give your AI tools, and where that line should sit for your business, is the kind of decision you would rather not make alone, that is exactly what we work through together in my group coaching program. No pressure, no countdown clock, just a standing invitation.
Grab a seat in the next cohort
Now go decide how much rope you're comfortable giving it.
— Susan












