unbrowseI asked my AI agent to post this article for me. It wrote it. It made the images. It opened X.
Then it hit the login screen.
So I'm posting this one myself. That's the whole problem.
Imagine ASI
A superintelligence that plans your week, runs your company, closes your books and negotiates your contracts. Smarter than every expert you know, awake all night, working for you.
Now watch it log into your bank.
"Verify it's you."
It can't.
The smartest mind ever built, stopped by a password field.
Intelligence isn't the bottleneck. Access is.
That's where agents are today. They can read the page, plan the steps and fill in the form.
Then they reach the door. A login screen. A bot check. A captcha on an account you've had for ten years. A 2FA code sent to a phone they will never hold.
Every model release makes the mind smarter. None of them gets it through the door.
The web is checking for a person
The web was built for one kind of user: a human with a screen and a mouse. Every login, every captcha, every "is this you?" email exists to confirm there's a person on the other side.
So we dress agents up as people. A fresh cloud browser that has never met you, on a datacenter IP. Your passwords stored on someone else's server so a bot can type them back in. A new integration for every site.
Sites notice. On 20 September, Amazon cut Meta's Muse shopping agent off from ordering on Amazon.com. Its two complaints: the agent didn't say it was an agent, and it appeared to collect and store customers' credentials. Meta says Muse never sees passwords. Either way, the store's objection was about who was knocking, not about what the agent could understand. (I wrote more about that case in In Defense of Muse.)
You can't disguise your way to ASI.
I should be plain about our own part in this. Unbrowse today does some of that dressing up. When an agent needs a signed-in site, the live path is our cloud browser and a sealed server-side vault for the logins you choose to save; bot checks get handed to you rather than bypassed. It works, and it is still a stranger at the door. The rest of this piece is about how we want to replace it.
One thing on the web is already trusted as you
Your browser.
It has cleared the captchas and 2FA prompts you've passed. It holds your sessions. It sits on your home connection, and the sites you use have known it for years.
Now imagine it reachable from anywhere.
That's the design we're building. Your agent runs in the cloud. When it needs to act as you, it reaches back to your browser and makes the request there, through the session you already have. In the build we're testing, the request is sent by your own browser, on your connection and through your session, so on that path your passwords and cookies stay on your machine, and the site sees the same browser it always has, because it is you.
The guardrails matter as much as the reach. In that build, you approve each site separately; banks, email and health portals can't be approved in one click; and anything other than reading waits for your approval unless you've trusted the site for writes. One gap we know about: when your browser is offline, the agent falls back to Unbrowse's other paths, such as a session you chose to sync or the cloud browser. We're adding an "only through my browser" setting, so that the agent is told your browser is unreachable instead.
Skip the paint
Once it's in, the agent doesn't need to click around like a person. Many modern web apps are an API with a UI painted on top: press a button and the page sends a request.
The agent skips the paint and sends the request.
On our own demo booking site, a browser agent takes a median 36.7 seconds to get the result. Replaying the request the site's frontend makes takes 1.97. The same gap shows up at scale in our April 2026 paper: across 94 live domains, a learned route answered in 950 ms on average against 3,404 ms for Playwright doing the same task.
And don't wait for the official version. We checked the 425 sites agents use through Unbrowse: on 85% of them, no official API covers the task agents actually do there. The official API is often the one that gets switched off.
What we're building
Unbrowse is one MCP server for the web, including the sites with no official API. It learns the requests a site's own frontend makes, verifies them, and replays them. That part is live at unbrowse.ai/mcp. The first request on a site it hasn't seen still runs in our cloud browser while the route is learned, and some sites only work in a browser at all.
As of 12 October 2026, the reach-back through your own browser is built and in testing, but not yet switched on in production. We won't say it works for you until it does.
The idea is simple to state. Your browser supplies the identity. Unbrowse supplies the APIs. Together, an agent could act as you on the sites you already use, from wherever it runs, without your password ever leaving home.
Imagine ASI.
Now imagine it can actually get in.
Your browser already has the keys. We're building the reach. The docs show what you can connect today.