Chat Without The Internet
Summarize Pages with Q/As
Fetch With Single Command

Project Story

Inspiration

My browser was a disaster. I had like 50 tabs open at any given time—work research, random articles I swore I'd read later, shopping tabs comparing prices. Finding anything took forever. I'd waste time scrolling through tabs, skimming long articles for one sentence, clicking through endless reviews. It was stupid that browsing still worked this way.

Then Chrome announced Gemini Nano—AI that runs right on your machine. No servers, no uploading your data anywhere. I thought: why can't I just tell my browser what I need?

What it does

UniAssist lets you talk to your browser. Say "open YouTube" and boom, it's there. Ask "what's this page about?" and it tells you. "Summarize this article" gives you the highlights. Looking at product reviews? Tell it to analyze them and it'll pull out the good and bad.

Everything happens on your computer using Chrome's built-in AI. Your data never leaves your device.

How I built it

I had two half-finished projects sitting around—one with decent browser automation, another with a nicer UI. I smashed them together and added webpage analysis from scratch.

Built with:

React 19 + TypeScript for the interface
Chrome's Prompt API (Gemini Nano) for the AI conversations
Chrome's Summarizer API for quick summaries
WebKit Speech Recognition for voice
Content scripts to grab webpage text
Session management so you can ask follow-up questions

The hardest part was making webpage analysis work. Every site structures their content differently. I needed something that could pull the actual article or product info while ignoring ads, navbars, and all the other junk. Then I added session management so you could have an actual conversation—like asking "what are the main points?" after getting a summary.

Challenges I ran into

Gemini Nano was barely documented: The API was brand new. I spent forever figuring out how sessions worked, what the token limits were, how to stream responses without breaking things.

Content extraction is a nightmare: News sites, blogs, product pages—they all do their own thing. I had to test on dozens of sites to get the selectors right.

Voice doesn't work everywhere: Chrome blocks voice input on internal pages for security reasons. Had to add text input as backup and make it obvious when voice wouldn't work.

Storage limits: Webpage content gets huge fast. I added chunking, compression, and auto-cleanup so it wouldn't blow up Chrome's storage.

Making streaming look good: Getting AI responses to stream smoothly while keeping markdown formatting working took way more tries than I expected.

Accomplishments that I'm proud of

Built a full webpage analysis system that remembers context between questions
Made the AI responses feel responsive with streaming and proper markdown
UI clearly shows what's happening (execution badges, status updates, all that)
Everything's private—processing happens on your device
Took two okay projects and made something actually useful

What I learned

Chrome's AI APIs are powerful but finicky—you have to babysit sessions and token counts
Voice interfaces need constant visual feedback or people get confused
Extracting content from webpages is way harder than it seems
Streaming makes everything feel faster even when it's not
People actually care about privacy—"runs on your device" is a real feature