Google quietly released AI Edge Eloquent on iOS this week, and TechCrunch covered it Monday. It’s a free dictation app that transcribes speech using on-device Gemma models, no internet required. You can use it underground, on a plane, in a dead zone. And all the text transformation features run locally too.

But what caught me off guard was not the product. Google’s entire revenue model depends on your data hitting their servers, and they shipped an AI app where offline is the default. They built a Gemini cloud mode for extra polish and made it opt-in. So the cloud is the bonus feature, not the product. I do not know when Google became the poster child for on-device AI, but apparently that happened while everyone was busy arguing about which cloud LLM is cheapest.

What does on-device AI have to do with your browser?

More than you’d think. Eloquent validates something browser agent developers have been arguing for a while: AI works better when it runs where your data already lives, not in some remote environment that starts from zero context.

And the gap is not just theoretical. Eloquent imports your personal vocabulary from Gmail history to get names and jargon right, all without those emails leaving the phone. That’s context engineering at the device level, using data that was already local. The same principle applies to your browser, where the context is even richer.

Your browser right now is logged into Gmail, probably Slack, maybe your company’s CRM or project management tool. Session cookies, form history, open tabs mapping out exactly what you’re working on today. All of that context just sitting there, untouched by most AI tools.

So when a cloud browser agent spins up a disposable Chromium instance in some data center, authenticated to nothing and aware of nothing, it’s throwing away the most valuable thing about your browser: that it already knows who you are and what you’re doing. It’s like hiring someone to organize your desk, but they refuse to enter your building and instead attempt to recreate your entire filing system from memory in a hotel room across town, without your keys or your WiFi password.

Already built

Dassi runs inside the browser you already have open. If your Gmail tab is authenticated, Dassi reads it. Same for Jira, your CRM, whatever tabs you have loaded.

The OAuth handoff nobody wanted

Because Dassi lives in the side panel, it inherits every session and login state you’ve already established just by using your computer normally. And there’s no OAuth handoff involved.

Every third-party AI integration for email or CRM starts with the same ritual: grant tokens, hope the middleware handles your credentials responsibly, give some startup you’ve never heard of read/write access to your entire inbox because you wanted help drafting one reply. I’ve seen people hand over full Gmail scopes to Chrome extensions with 200 reviews and a privacy policy hosted on Google Docs. But a local browser agent skips that entire layer. It uses your existing authenticated session the same way you do when you click around manually, and no third party ever touches your credentials.

Google said the quiet part out loud

The local-first argument is not new. Or novel. But Google building a consumer product around it, shipping it for free, and defaulting to on-device execution is new. Because when the most cloud-dependent company on earth decides that on-device is the better user experience for AI, that is not a blog post from some startup with an agenda. That’s a product bet with real engineering resources behind it.

And it will not stop at dictation. Every browser workflow where you’re already logged in is a candidate for the same treatment: the model runs locally or via your own API key, context comes from your existing session, data never leaves your machine. If you’re already paying for ChatGPT, you can run it inside your browser right now with that exact setup. No new account, no token dance, just the browser you’re already sitting in.

I suspect the whole cloud-agent era is going to look, in retrospect, like that stretch when every mobile app insisted on creating its own login system before “Sign in with Google” made it all irrelevant. Billions spent on credential plumbing that could have been skipped entirely. A hell of a lot of infrastructure that existed only because nobody bothered to use what was already there.