OpenLive is a desktop app that adds a voice and vision layer to an AI model or coding agent, with the whole speech loop running on the local machine. It handles voice activity detection, streaming speech-to-text, end-of-turn detection, streaming text-to-speech, and barge-in, and it can pass camera or screen frames with each turn. Audio never leaves the machine: only the final transcript, plus frames when enabled, goes to whichever model or agent answers.
The answering side is either a model reached with your own API key or a coding agent driven locally over the Agent Client Protocol (ACP). Supported providers include Anthropic, OpenAI, Google Gemini, xAI, DeepSeek, Groq, Ollama, MiniMax, OpenRouter, and Mistral, and supported agents include Claude Code, Codex, Cursor, OpenCode, Hermes, Gemini CLI, GitHub Copilot, Kiro, and Pi. With an agent, OpenLive speaks its permission requests, narrates plan steps, and stores calls as the agent's own CLI sessions so they resume from either side.
OpenLive is a cascaded pipeline (speech to text, model, then speech) rather than a speech-to-speech model, which is what lets any model or agent answer. It has three modes: Chat for a voice call in the app window, Flow for a system-wide assistant summoned by a double tap that answers aloud, types, and operates apps through a computer-use helper, and Dictate for voice typing into any text box. The app positions itself as an open alternative to ElevenLabs Agents, Gemini Live, and OpenAI Realtime, and sends anonymous usage counts that can be turned off.
Features
- On-device voice loop: Silero VAD, Smart-Turn end-of-turn detection, and speech-to-text with Whisper on WebGPU or native engines such as Nemotron, Parakeet, Moonshine, and Canary
- Local text-to-speech: Kokoro, Supertonic, Pocket TTS, Kitten TTS, Piper, and Matcha, with streaming that starts before the sentence ends
- Barge-in: interrupting stops the reply mid-word
- Bring your own model: API keys for Anthropic, OpenAI, Gemini, xAI, DeepSeek, Groq, Ollama, MiniMax, OpenRouter, Mistral, and others
- Coding-agent calls: drives Claude Code, Codex, Cursor, OpenCode, Hermes, Gemini CLI, GitHub Copilot, Kiro, or Pi over ACP, with voice answers to permission prompts
- Vision: camera or screen frames ride each turn, a look tool grabs a high-resolution frame, and text-only models can borrow a vision model
- Flow assistant: a system-wide voice assistant that answers aloud, types at the cursor, and opens apps, clicks, scrolls, and runs shell commands
- Dictate: voice typing into any text box with local cleanup rules, an optional AI polish, voice editing, and a custom dictionary
- Voice cloning: optional local zero-shot cloning with ZipVoice from a 5 to 30 second recording
- Ten languages: English, Spanish, French, German, Italian, Portuguese, Hindi, Chinese, Japanese, and Korean across every stage
- Tools and memory: built-in tools, skills, and MCP connectors importable from Claude Code, Codex, Cursor, Gemini CLI, and VS Code, plus editable remembered facts
- Local history: per-mode history with search and configurable retention, and encrypted API key storage
