Agent-First Browser
VELOX
A Chromium browser where research agents work in visible tabs.
Visible agent tabs
no-key demo
Markdown and JSON export
Problem
Most AI search tools hide the browsing process and only begin once a full query is submitted. Browser automation is also brittle when the user cannot see what the agent opened, read, skipped, or cited.
System
Predictive omnibox to visible agent tabs to research canvas. Cerebras × Gemma 4 hackathon origin for low-latency drafting.
Evidence
Real browser UI as the hero surface; Playwright research runtime; no-key demo via DEMO_MODE_ENABLED; Markdown and JSON export endpoints.
Outcome
The local app ships as a hackathon-ready alpha (v0.3.0) with predictive intent, visible agent tabs, live browsing evidence, cited answer generation, Markdown/JSON export, persisted session snapshots, and a one-command macOS boot path.
System
Visible research flow
Predictive intent opens real Chromium tabs, then streams evidence into a cited answer canvas.
Approach
I built a desktop shell, React browser chrome, Express/WebSocket agent server, Playwright research runtime, source shelf, export path, and no-key demo mode. The research flow streams session state as agents navigate, capture evidence, and hand off to a synthesis agent.
Constraints
Product still evolving; agent actions remain user-initiated and inspectable. Live tab takeover is still listed as future work.
Evidence
Inside the system
