Somebody on my feed this morning quote-tweeted the “Apple finally fixed Siri” headline with the caption “cool, so is that an AI browser now or not.” The replies went sideways fast. Three people answered confidently and all three described completely different software, which is roughly where the whole category sits right now.

So: what is an AI browser? The term covers at least three products that have almost nothing in common under the hood. The only question that separates them cleanly is what the thing can see. Not what it can do, not how smart the model is. What it can see.

The replacement browser

You download a new Chromium build. Comet from Perplexity, Dia from The Browser Company, Atlas from OpenAI before OpenAI pulled the plug on it. The AI is welded into the chrome of the app itself: a sidebar, an omnibox that answers instead of searching, sometimes an agent mode that clicks around for you.

Because it is a browser, it can see everything a browser sees. Full DOM of every tab, its own cookie jar, your history, your open sessions. That’s real power.

It’s also a migration. Your passwords are in the old browser. Your extensions are in the old browser. Your bank has your old browser fingerprinted and will make you do the SMS code thing again. And the honest reason Atlas died is that asking people to abandon Chrome for an AI feature turned out to be a hell of an ask.

The cloud agent driving a Chrome you’ve never seen

This is the one people mix up most. A cloud browser agent runs Chromium in a data center somewhere, takes your instruction, and drives that remote instance. Browserbase, most of the open-source agent frameworks, most of the “autonomous web agent” demos on X.

That remote Chrome is a stranger. It has no cookies, no logged-in Gmail, no Salesforce session, nothing. It was born ninety seconds ago in a container and it will die when the task ends. So for it to touch anything of yours, you have to hand over access: an OAuth grant, an API key, or, worse, your actual password typed into a remote box you don’t control. We wrote about that gap in more detail in cloud browser agents can’t see your tabs, and nothing since has changed the math.

Great for scraping public pages. Genuinely useless for “reply to the email I’m looking at.”

The side panel in the Chrome you already have

Third category, and the one I think most people actually mean when they search for this. It’s an extension. It opens in the side panel of the Chrome you’re already running, the one where you’re logged into Gmail and Notion and your CRM and that internal admin tool nobody has an API for.

Because it executes inside your browser, it inherits your browser. The active tab’s DOM is right there. Your session cookie for that tab already went out with the request that rendered the page, so the agent never needs its own copy of it, never needs to authenticate separately, never asks Google for a scope grant. Dassi works this way: it reads the page you’re on, and the extent of what it can touch is the extent of what you can touch.

That last part is the interesting constraint. An extension agent has exactly your permissions, no more, which means it cannot quietly acquire a standing credential that keeps working after you close the laptop, and it cannot be revoked-but-still-running the way a stale OAuth token can. The blast radius is a tab.

It also means the tradeoffs run the other way. Your machine has to be on. Nothing runs overnight. There’s no fleet of fifty parallel sessions grinding through a queue while you sleep. If you want unattended scale, category two is the honest answer and I’d say so.

What each one can actually see

Replacement browser: everything, in a browser you had to move into. Cloud agent: a blank Chrome, plus whatever credentials you handed it. Side panel extension: the tab in front of you, already authenticated, nothing more.

Which one you want depends on one question

Do you need the AI to read something you’re already logged into? Then category three is the only one that gets there without a re-login or an OAuth grant. That’s not a marketing claim, it’s just where the session state physically lives.

If the answer is no, if you’re pulling public pricing pages or monitoring competitor sites at 3am, rent a cloud browser and skip the extension entirely.

Dassi is free and sits in the side panel, and you can run it on your existing ChatGPT subscription or your own Claude or Gemini key. Chrome Web Store link here if you want to poke at it.

My guess is the term “AI browser” doesn’t survive 2027 in any useful form. It’s already doing the work that “smart TV” used to do, which is to say: describing four incompatible things to people who just wanted to know if it plays Netflix.