OpenLive
An open-source, on-device voice and vision layer that gives any AI model or coding agent real-time speech input and output, screen or camera awareness, and the ability to be interrupted mid-reply.
OpenLive provides the pipeline that a hosted voice-assistant product would otherwise rent by the minute: voice activity detection, streaming speech-to-text, a model turn, streaming text-to-speech, and barge-in so a user can interrupt, plus camera and screen input. All of it runs locally in the browser using WebGPU rather than in the cloud.
It works with a model the user already has a key for, including Anthropic, OpenAI, Google, xAI, DeepSeek, Groq or a fully local Ollama model, and can drive a coding agent such as Claude Code, Codex, Cursor, OpenCode or Hermes locally over the Agent Client Protocol. A voice-cloning feature records a short sample of the user's voice so the assistant can speak in it. The architecture is a cascaded speech-to-text-to-model-to-speech pipeline rather than a single full-duplex speech model, which the project describes as a deliberate trade-off that keeps it local and provider-agnostic.