kestrel-voice-xai 0.1.1

xAI realtime conversation, TTS, and STT providers for Kestrel Sovereign

kestrel-voice-xai

xAI voice providers for Kestrel Sovereign:

• Realtime speech-to-speech over the xAI Voice Agent WebSocket API
• Text to speech, including streaming MP3, WAV, PCM, μ-law, and A-law output
• Batch and streaming speech to text
• Runtime discovery of voice models, built-in voices, and custom voices
• Governed Kestrel function tools with atomic multi-tool continuation

Install the package beside kestrel-feature-voice and set XAI_API_KEY. Providers are discovered through Python entry points; the feature package does not import this distribution directly.

uv pip install git+https://github.com/KestrelSovereignAI/kestrel-voice-xai.git
export XAI_API_KEY=...

All three roles share the operator-facing [voice.xai] configuration. Model and voice IDs are discovered at runtime; selecting no realtime voice lets xAI use its current provider default.

[voice.xai]
language = "auto"
# model = "a-versioned-model-from-/v1/models"  # optional production pin
# keyterms = ["Kestrel", "your product name"]

# Provider-native tools are intentionally opt-in. Kestrel function tools do
# not go here; they remain governed by the Sovereign dispatcher.
native_tools = ["web_search", "x_search"]

Provider-native web_search, x_search, file_search, and remote mcp tools are supported by xAI but disabled by default. Enable only the specific native tool definitions intended for the agent in the provider configuration. Kestrel function tools continue to execute through the Sovereign governed dispatcher.

The voice UI receives the provider's live built-in and team custom-voice catalog, persists voice choices by provider, and uses an xAI ephemeral token for direct browser WebSocket sessions. Long-lived API keys stay server-side.

Upstream contracts: Voice Agent API, ephemeral tokens, text to speech, speech to text, and custom voices.