Nexus Browser

Nexus is an autonomous, stateful multi-agent research browser: you give it one directive (a sentence or a full brief), and a hierarchy of agents plans the work, drives a real Chromium session page by page, verifies every claim verbatim against the raw page text, and compiles a cited report — while a live dashboard shows exactly what the agents are looking at, thinking, and paying. The problem: real research is tab-juggling. Search engines and chat assistants need constant human prompting, can't traverse multi-step sites (dashboards, publications, paginated docs), and hallucinate numbers with no receipts. What Nexus does differently: • Plan–execute–verify loop: a text-only planner splits the goal into steps; a vision executor drives the browser (navigate/click/read/screenshot); a verifier scores every step and routes done / retry / replan — retries stay local, escalations rewrite the plan (video shows a real verifier-flagged re-plan). • No model grades itself: every claimed quote is re-checked in code against the page's raw text (Unicode-folded matching). Unprovable claims are dropped — a hallucination becomes a retry. • Zero-trust browsing: robots.txt checked before navigation, page content treated as untrusted data, no stealth scripts. • Privacy Ledger UI: every LLM call is metered per endpoint and role — calls, tokens, bytes, images, and billed $ — shown live on the dashboard. Planner/reporter are screenshot-blocked by construction. • One-secret setup, two modes: cloud-only (all Meta) or hybrid (open-weights eyes on Baseten). Built during the hackathon (full build log in the README's commit-range section): LangGraph topology, the verdict router, the deterministic evidence engine, the Privacy Ledger, the mission-control dashboard (Live/Report/Graph), and the demo-recording kit that produced the submitted 3-minute video from one genuine run — real receipts on screen: $0.22 billed, 24 agent calls, 199k tokens, 10+ sites visited, 4 verbatim-verified evidence quotes, 1 verifier→planner escalation.