Tera: Zero Token Browser Use

Today, every agent run starts from zero. Computer-use agents are slow, unreliable, and burn tokens re-reasoning about the same websites every time.

Tera is a native macOS browser (Swift + WKWebView) that turns everyday browsing into reusable workflows. It passively observes your actions, embeds page states as vectors, and compiles an embedding-based Markov Decision Process on your machine. When a new task matches a learned state, Tera replays the stored action with zero LLM tokens (Tier-1). LLM fallback only runs on drift (Tier-2/3).

How it works: DOM snapshots → 1536-d embeddings → cosine match → MDP route → native WebKit execution. Trained on HUD agentic browser trajectories; policies stored locally via NetworkX. No Playwright CDP — the browser is the environment.

Demo: Book a flight on Google Flights. Run 1 learns the workflow (~68s, ~14k tokens, Browser Use–style loop). Run 2 replays via policy (~15s, 0 tokens, HUD LLMJudgeGrader ≥ 0.75).