Environment Variables

Provider selection

• LLM_PROVIDER = openai | anthropic | google | bedrock | openrouter | zai (alias: glm) | lmstudio | ollama
• TTS_PROVIDER = elevenlabs | openai | piper
• STT_PROVIDER = openai | faster-whisper | elevenlabs-scribe (aliases: elevenlabs, scribe) | openai-realtime (alias: realtime)
• LOCAL_ONLY = true to force local providers

API keys / URLs

• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• GOOGLE_API_KEY
• OPENROUTER_API_KEY
• AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION
• ELEVENLABS_API_KEY
• OLLAMA_BASE_URL, LMSTUDIO_BASE_URL, FASTER_WHISPER_URL, PIPER_URL

Model overrides

• OPENAI_MODEL, ANTHROPIC_MODEL, GOOGLE_MODEL, BEDROCK_MODEL, OPENROUTER_MODEL
• OPENAI_STT_MODEL, OPENAI_TTS_MODEL, OPENAI_TTS_VOICE, OPENAI_TTS_INSTRUCTIONS

Server config

• PORT, HOST
• SYSTEM_PROMPT
• STREAMING_TTS (true/false)
• STREAMING_STT (true/false, default false) - stream mic audio to STT live for interim transcripts; requires a streaming STT provider (elevenlabs-scribe, openai-realtime)

Behavior

• Auto-detection for LLM (when LLM_PROVIDER and LOCAL_ONLY are not set) picks the first provider with a valid key in this order: Anthropic → Google → Bedrock → OpenRouter → OpenAI.
• LOCAL_ONLY=true forces local providers: Ollama for LLM, Faster-Whisper for STT, and Piper for TTS.

Anthropic

Variable	Default	Description
ANTHROPIC_PROMPT_CACHING	false	true places cache breakpoints on the system prompt and rolling history - cache reads cost 0.1x input price

OpenAI voice

Variable	Default	Description
OPENAI_STT_MODEL	whisper-1	Transcription model; gpt-4o-transcribe / gpt-4o-mini-transcribe improve accuracy; with STT_PROVIDER=openai-realtime the default is gpt-realtime-whisper
ELEVENLABS_STT_MODEL	scribe_v2	ElevenLabs Scribe batch model (STT_PROVIDER=elevenlabs-scribe)
OPENAI_TTS_MODEL	tts-1	TTS model; gpt-4o-mini-tts supports delivery instructions
OPENAI_TTS_VOICE	nova	OpenAI TTS voice: alloy, ash, coral, echo, fable, nova, onyx, sage, shimmer (plus ballad, verse on gpt-4o-mini-tts)
OPENAI_TTS_INSTRUCTIONS	-	Natural-language delivery direction (tone, pacing, persona) for instructable TTS models

Z.ai (GLM)

Variable	Default	Description
ZAI_API_KEY	-	Z.ai API key (LLM_PROVIDER=zai)
ZAI_MODEL	glm-5.2	GLM model name

Local vision

Variable	Default	Description
OLLAMA_VISION_MODEL	llava	Vision model used by the CLI's local vision provider. The CLI default stays llava so existing deployments keep working; set qwen3-vl for the current generation (the OllamaVisionProvider class itself defaults to qwen3-vl)