speaker-ai-connector - Let children access AI freely with voice, without a screen.

I built this because I wanted my kid to be able to ask an AI questions without me handing over a phone or sitting them in front of a tablet. The whole interface is a Bluetooth speaker on a shelf. The flow: pair an HFP/HSP Bluetooth speaker (the speakerphone kind, with a working mic) to a Mac, configure it once in the menu-bar app, and from then on — whenever the speaker connects, a Gemini Live voice session opens automatically and bridges the speaker's mic and output to Gemini. Power-cycle the speaker to end the session. No phone, no screen, no parent in the loop. A few things I cared about: - **Cost stays bounded by silence.** A local VAD (Silero by default, WebRTC as a no-model fallback) gates uploads. A left-on speaker in a quiet room sends nothing. - **Everything is recorded locally.** Each input utterance and each Gemini response is saved as a WAV under `~/Library/Application Support/...`, browsable in a Sessions window with click-to-play. Nothing goes anywhere except Google's Gemini Live endpoint. No telemetry. - **Bring your own Gemini API key**, stored in the macOS Keychain.