BrowseAttune

Inspiration

Most browser AI assistants wait for you to ask a question. By then, you've already lost context—what page you were on, what you clicked, what form you were filling out.

We built BrowseAttune around a different idea: sense first, assist second. The assistant should proactively understand what you're doing in the browser, then step in at the right moment.

What it does

BrowseAttune is a Chrome side-panel agent that:

Proactively perceives your browsing—clicks, navigation, form changes, and page structure—through an event-driven perception layer (no polling, no model required for sensing)
Assists on demand via a native agent chat panel with page reading, browser automation, and domain-specific Skills
Respects privacy with per-domain opt-in, pause/resume controls, and local-only storage

How we built it

We built a Chrome MV3 extension with WXT, React 19, and TypeScript. The agent runtime uses Pi Agent Core and Pi AI for streaming conversations and tool execution—fully client-side, no backend required.

The perception foundation captures L0 raw evidence (user events, navigation) and L1 canonical page snapshots (structured DOM/accessibility representation), designed to work even when the LLM is offline. Skills are stored in IndexedDB and loaded progressively—only metadata at startup, full instructions on activation.

Challenges we ran into

Separating perception from inference: keeping the sensing layer deterministic (L0/L1) while leaving business understanding (L2/L3) to the agent
Chrome extension constraints: side panel lifecycle, content script injection, and permission scoping across tabs
Dynamic pages: element refs invalidate after DOM re-renders—every browser action requires a fresh snapshot

Accomplishments that we're proud of

A perception layer that runs independently of the model—event-driven, debounced, and evidence-chained
End-to-end browser agent with page read, snapshot/action tools, and packaged Skills (e.g. one-click form fill on enterprise workflows)

What we learned

Proactive assistance starts with observation architecture, not a better chat prompt. The hardest part is building a clean boundary between "what happened in the browser" and "what it means for the user."