Tabbit 浏览器

What Tabbit is

Tabbit is pitched as an AI-native browser for the AI era, with context understanding, an on-call agent, and deep links to Chinese models. It wants to be the browser where work is handed to the assistant rather than the browser where you read pages.

That ambition defines how it should be judged. A conventional browser is evaluated on rendering and speed; this one is evaluated on whether delegating a task to the agent actually produces the result you wanted.

What it does

You browse and use an at sign to pull in text, screenshots, pages or files as context, switch among several widely used Chinese models, and let the agent open pages, fill forms and reply to customers on its own. The loop is reference, ask, run.

The at-sign reference gesture is the part that makes the assistant usable rather than toy-like, because it lets you point precisely at what you mean instead of hoping the model inferred it from the page. Switching between several domestic models means you can match the model to the task, which matters where a general model is weak on a specialised question.

It tidies tabs into a sidebar, translates pages, helps with writing and script generation, and moves bookmarks, history and extensions from another browser in one step, with a command-line tool for power users.

Pricing is unconfirmed: the page asks whether it is free and does not answer, and the privacy and security questions are likewise left blank.

Who it is for

It is aimed at workers and students in the China region who want a browser with local models and an agent for repetitive page work. Anyone already using those models through separate interfaces is the natural adopter.

What to keep in mind

The unanswered questions are the first thing to resolve, and they are not incidental ones. Whether it costs anything, what data the assistant sends and to which model, and how your browsing history is handled are exactly the facts you need before making a browser your daily one. A page that raises these questions and leaves them empty is a page to revisit before installing.

Scope what the agent can touch. An assistant that fills forms and replies to customers acts as you on accounts you hold, and the practical safeguard is to confine it to actions you can undo rather than to hope the model is careful. Customer replies in particular are messages sent in your name, and a fluent answer is not the same as an appropriate one.

No certification was named, so read the policy before using it with sensitive material, and decide deliberately which model receives which content, since routing between providers means the data path changes with the model you pick. It is worth writing down your default choice, because the model selector is easy to change and easy to forget.