G-Axis

The problem

AI chatbots can answer questions but can’t *do* anything in your browser, and “AI browsers” force you into a separate app — losing your logins, extensions, and habits.

How it works

A Chrome side-panel that lives in your existing browser. It perceives the page with Gemini Vision (screenshots, not brittle DOM selectors), plans, and executes clicks and form-fills through Playwright — with a policy engine scoring every action for risk and gating the risky ones behind approval. A parallel real-time voice companion runs over the Gemini Live API with eight personas and post-session communication scoring.

Key features

• Screenshot-based Gemini Vision perception with confidence scoring
• Policy engine with 3 supervision modes (supervised / semi-auto / autonomous) and approval gates
• 3-tier memory: working, episodic (per-domain), and semantic (cross-site patterns)
• Agent ecosystem: A2A REST + Agent Card, MCP-compatible browser tools, ADK coordinator
• Real-time voice with 8 personas surviving Gemini’s ~10-min session timeout

Architecture

A Chrome MV3 extension (side-panel + service worker + content script + AudioWorklet) connects over WebSocket to a FastAPI + Google ADK backend on Cloud Run — Gemini 2.5 Pro for vision/planning, the policy engine, 3-tier memory, a Firestore audit store, and Playwright. Voice takes a zero-hop direct socket to Gemini Live. Its vendor-neutral sibling, Axis, factors this into an event-sourced control plane with pluggable model adapters.