TaffyGo

A lot of useful work starts in a browser and ends somewhere else. Read a few pages. Compare the details. Copy a table. Go back to the form. TaffyGo starts with that familiar sequence and asks how much of it could stay in one place.

I’m building TaffyGo at Matterward Labs: an AI-native browser for Android, based on Chromium, with one assistant called Taffy. Ordinary browsing remains the foundation. Tabs, bookmarks, history, downloads, private tabs, and ad and tracker blocking come first, and the assistant works alongside them.

A browser first

The new tab is a quiet starting point: original artwork, one address box, and the everyday browser controls. Search the web, ask about a page, or hand Taffy a task. Those are different intentions, and the interface should keep the distinction understandable.

The light and dark surfaces above are rendered from the same new-tab design used on the TaffyGo website. The film shows browsing and research with Taffy.

On a fresh install, TaffyGo is a complete browser before any AI is connected. Blocking is on from the first page, it works in every tab, and you can allow a site you trust. The block lists ship inside the app, so blocking doesn’t depend on a TaffyGo server.

Three ways to use it

People choose how much help they want, page by page. You can browse while Taffy stays out of the way. You can browse together and ask about the page you’re reading. Or you can hand over a piece of research and watch Taffy work.

Taffy’s tabs are ordinary tabs that you can watch, touch and close. Take over whenever you want, and the task stops where it is. Nothing about the assistant hides what the browser is doing.

Research you can check

The useful result of research is something a person can inspect, save and return to. Bring information from several pages together, compare products and reviews, and keep the sources beside the answer. In TaffyGo, results collect in a workspace where every fact shows its source, and you keep what you want in your Library.

Taffy can be wrong, and pages can fail to load. When that happens, it reports what it found and what it couldn’t, instead of rounding partial work up to success. When it needs an answer, it asks you.

Nothing consequential without you

Before Taffy fills in a form, it stops and asks. You can approve it, change it or do it yourself. Taffy asks before opening more tabs, and it doesn’t buy things, send messages or post on your behalf.

Passwords, one-time codes, payment details and passkeys are never sent to an AI provider. That’s part of how TaffyGo is built, not a setting you have to find.

Your provider, your phone

There’s no TaffyGo account and no TaffyGo server for the app to sign in to. People connect the AI they prefer: their own provider key, an existing plan at a supported provider, or a model server they run themselves. Keys stay on the phone, and requests go directly from the phone to the provider the person chose. To answer a question about a page, Taffy sends that page’s text to that provider.

History and Taffy’s memory stay on the phone. Without a provider, TaffyGo is still a complete browser with blocking, and Taffy waits until one is connected. TaffyGo isn’t a VPN: the blocker works on the pages TaffyGo loads.

The hard parts

The first is staying close to Chromium. A browser built on Chromium has to keep up with it, because security fixes land there first. TaffyGo keeps its own work apart from Chromium’s code and treats every change to Chromium itself as a cost that has to justify itself.

The second is keeping the assistant inside the rules. The part that plans Taffy’s next step can only suggest it. Whether the step is allowed is decided separately, and the browser carries it out. A planning mistake becomes a suggestion that can be refused.

The third is making approval feel natural. Asking too often trains people to tap yes without reading, and asking too rarely breaks trust. Taffy asks at the moments that change something, such as opening more tabs or filling in a form. I wrote more about these decisions in Building an AI browser for Android that asks first.

Where it stands

Version 1.0 is the first release. It needs a 64-bit ARM phone with Android 10 or later, and so far it has been tested on one physical phone. The source is on GitHub, and the README asks people to open an issue with their phone model if anything misbehaves.

Build the browser with us

TaffyGo is available on Google Play. Its first-party source is licensed under the Mozilla Public License 2.0. Developers, designers, translators and testers are welcome to help shape it.

Visit taffygo.com for the product story and contribution information. The launch film above brings the direction together in under two minutes.