Voice and Companion

frogg has three voice features. All speech runs on the daemon, locally by default.

Feature	What it does	Status
Dictation	Speak into the composer instead of typing	Available
Voice alerts	Agents’ updates are spoken; reply or approve permissions by voice	Available
Companion	A voice conversation that dispatches and tracks coding agents	Preview, off by default

Speech on the daemon

With no configuration, the daemon uses local models (sherpa-onnx) and downloads them in the background into ~/.frogg/models/local-speech:

Role	Default model	Alternatives
Speech-to-text	`parakeet-tdt-0.6b-v2-int8` (English)	`parakeet-tdt-0.6b-v3-int8` (25 European languages)
Text-to-speech	`kitten-nano-en-v0_8-fp32` (voice Rosie)	`kokoro-int8-multi-lang-v1_0`, `piper-ljspeech-medium`, `kokoro-en-v0_19`
Voice activity	Silero VAD

Voice features are on when the local speech runtime is available. Turn them off with FROGG_VOICE=0 or "features": { "voice": { "enabled": false } }.

To use OpenAI for speech instead:

{
  "providers": { "openai": { "apiKey": "sk-..." } },
  "features": {
    "dictation": { "stt": { "provider": "openai" } },
    "voiceMode": { "tts": { "provider": "openai", "voice": "nova" } }
  }
}

All keys and env vars are in the configuration reference.

Dictation

Use the microphone button in the composer. The transcript lands in the composer for you to edit and send.

Voice alerts

Turn on Voice alerts per workspace from the composer. When an agent finishes, fails or needs a permission, the daemon synthesizes a short spoken update. The alert offers Play, Reply by voice and, for permissions, Allow/Deny by saying yes or no.

Settings → General → Auto-play spoken alerts plays them as they arrive while the app is in the foreground.
Confirm voice replies shows your transcript for two seconds before sending.

Voice alerts need a working text-to-speech provider on the daemon.

Companion

Companion is a voice conversation with a model that works through your agents: ask it to start a task, check on a running agent, or answer a permission, and it speaks back results. It is labelled preview and experimental.

Turn it on

Settings → General → Enable Companion (preview). It is off by default on each device, and enabling it does not start the microphone.
In a project workspace, press the Companion button in the composer, then allow the microphone.

Controls: Mute pauses the microphone while tasks continue. Minimize keeps the conversation running with an indicator. End closes it; coding tasks keep running. Type instead of speaking in the text field.

Only one Companion conversation per daemon is allowed. A second device gets Another device is using Companion on this daemon. End that conversation first.

Conversation model

Companion needs a model on the daemon, chosen by features.companion.backend:

Backend	Uses
`subscription` (default)	Claude Code signed in with a Claude subscription, otherwise Codex signed in with ChatGPT
`claude`, `codex`	Only that one
`api`	`ANTHROPIC_API_KEY` or `providers.anthropic.apiKey`, billed as API usage. Used only when selected

Default models: claude-haiku-4-5 (Claude) and gpt-5.6-luna (Codex). Override with features.companion.model.

If the start fails with Sign in to Claude Code or Codex on this daemon., log in to one of those CLIs as the daemon’s user.

Preferences

Settings → General → Companion applies to the next conversation: reply length (Brief or Detailed), spoken task updates, acknowledge tasks, audio mode (Call or Media on mobile), voice speed, pause before replying (0.8, 1.4 or 2.4 seconds), interrupt by speaking, animated graphics and reply text.

Limits

Preview quality: exercised on Linux daemons. Acceptance on physical mobile devices and other daemon platforms is still open, as are microphone and speaker latency tuning.
Codex voice (preview) uses Codex’s native voice transport. It needs a ChatGPT sign-in and features.companion.nativeVoicePreview: true on the daemon, and depends on your account.
Subscription backends consume your plan’s allowance. Check host settings → Usage.