Agent Voice Kit: Local talk with your Agents (whisper+supertonic)

Francisco Carlos Serra released Agent Voice Kit, an open-source tool that lets users talk to AI agents like Claude Code, pi, Codex, and Gemini CLI locally, with dictation and optional text-to-speech, running entirely on the user's computer. The kit requires Python 3.11+, ffmpeg, and a GPU for optimal performance, and downloads Whisper large-v3-turbo and Supertonic 3 models from Hugging Face (about 2 GB).

English · Español /franciscocarloserra/agent-voice-kit/blob/master/README.es.md you speak the agent answers you hear it optional press F1, talk, press F1 the text lands in the window the reply is read aloud Press a key, say what you want, press it again. Your words are typed wherever you were writing, in any app. If you want, the agent reads its answer back to you. simplescreenrecorder-2026-09-02 12.44.15 audio v3 1.5x.mp4 It all runs on your own computer. Nothing is sent anywhere, and there is no waiting for a server: the text is there the moment you stop talking. Paste this into your agent Claude Code, pi, Codex, Gemini CLI : Install https://github.com/franciscocarloserra/agent-voice-kit/ following its AGENTS.md What you will have to do by hand. The agent cannot click these for you: - Mac: type your password for brew , assign the two hotkeys in System Settings, accept the Microphone and Accessibility dialogs. - Windows: approve the ffmpeg and AutoHotkey installs when winget asks. - Linux: type your sudo password for the packages, assign the two hotkeys in your desktop's shortcut settings. | key | what it does | |---|---| | F1 | starts recording; pressing F1 again types your words at the cursor, in any app | | Meta+F1 | cancels the recording or silences the agent | /tts | turns read-aloud on or off, in any harness. Persists until you toggle it again | python3 voice.py serve runs the servers in a terminal, install-service runs them at login, doctor reports what is missing, and tts speed 1.5 sets the reading speed. You want a GPU. NVIDIA on Linux or Windows, or any Apple Silicon Mac. On CPU only it works but takes a few seconds per sentence. Models are downloaded, not bundled. Setup pulls Whisper large-v3-turbo and Supertonic 3 from Hugging Face, about 2 GB. A few things get installed on your system. Python 3.11+ and ffmpeg everywhere, AutoHotkey on Windows, xdotool on Linux, current NVIDIA drivers if you have that GPU. Text to speech works out of the box, and can get faster. The default engine needs no build. If you want replies to start instantly, ask the agent to build audio.cpp about 15 minutes, 5 to 8 times faster . macOS will ask for permissions. Microphone and Accessibility, once. Without them dictation records silence or does not type. Read-aloud is off until you turn it on. /tts toggles it for every harness and it stays that way. Dictation needs nothing: it types into whatever window has focus. Everything is customizable. Models, voice, reading speed, hotkeys, ports. Just ask your agent to change it; it all lives in config.json .