agent-live

Agent Live is a real-time duplex voice advisor and copilot designed for modern AI editors such as Cursor and Antigravity IDE. It is powered by the Gemini 3.8 Live API, full-duplex multimodal streaming, and a tactical HUD. The tool aims to eliminate the back-and-forth between typing descriptions and waiting for model output by acting as a persistent voice strategist and agile execution engineer. It provides end-to-end full-duplex real-time voice with millisecond-level response and supports barge-in interruption. The system includes a global push-to-talk hotkey (Ctrl+Space) for system-level capture, allowing users to speak while browsing code, reading documentation, or debugging. A dynamic tactical HUD embedded in the IDE terminal visualizes microphone energy, conversation status, background actions, and real-time subtitles using ANSI colors. Agent Live concurrently monitors local agent transcripts from both Antigravity and Cursor, intelligently identifying user questions, tool calls, and task completion events to provide voice commentary and battle reports. It also features a single-channel priority queue and dual-IDE conflict arbitration to prevent simultaneous speech from multiple instances. The system intelligently avoids interrupting media playback or meetings by detecting audio activity and queuing reports until media stops. A direct engineering engine allows users to issue verbal commands to execute tasks in the current workspace and receive voice reports. Additional features include hardware echo suppression, adaptive jitter buffer for smooth audio streaming, and a Red Team technical review approach that provides neutral, three-element reports (change substance, impact scope, residual risks) without flattery.