Environment variables

Copy env.example to .env for a local run, or set these on the container. Only the first group is needed to get a bot talking.

The MeetingBaas API key is not a server setting. Each POST /bots request carries the caller's key in x-meeting-baas-api-key and the bot is created on that account.

Providers (required)

# Which LLM the bots use by default: openai (default), anthropic or zai.
# Callers can override per bot with llm_provider, and bring their own key.
LLM_PROVIDER=openai

# The default models are the fast ones: a voice reply of a sentence or two
# gains nothing from a frontier model and loses seconds to it. Thinking is
# turned off in whatever way the chosen model accepts; models whose thinking
# is always on (Claude Fable, Opus 5.5 and newer) run at low effort instead.
OPENAI_API_KEY=
OPENAI_MODEL=gpt-5.4-mini
# responses (default, newest models) or chat
OPENAI_API_SURFACE=responses
# Optional OpenAI service tier: auto, flex, priority
OPENAI_SERVICE_TIER=

ANTHROPIC_API_KEY=
ANTHROPIC_MODEL=claude-haiku-4-5

ZAI_API_KEY=
ZAI_MODEL=glm-5.2
ZAI_BASE_URL=https://api.z.ai/api/paas/v4/

# Optional global fallback model; the provider-specific variables win
LLM_MODEL=

# Speech-to-text (Deepgram) and text-to-speech (Cartesia)
DEEPGRAM_API_KEY=
CARTESIA_API_KEY=
CARTESIA_VOICE_ID=79a125e8-cd45-4c13-8a67-188112f4dd22
# Default speaking speed when a request sets no speech_speed (0.6 to 1.5)
CARTESIA_TTS_SPEED=1.2

Set the key of whichever LLM_PROVIDER you choose; a request for a provider with no key is refused with a 400 unless the caller supplies one in credentials.

Server

# Port the API listens on
PORT=7014

# Public https URL of this server. MeetingBaas connects back to
# wss://<this host>/ws/<client_id> for audio. Required everywhere except a
# local ngrok run, where it is detected.
BASE_URL=https://speaking.example.com

# MeetingBaas API the bots are created on
MEETING_BAAS_API_URL=https://api.meetingbaas.com

# Writable directory for transcripts, floor files, ready signals, call
# summaries and meeting state. Defaults to the repository root; the
# container image sets /data.
SPEAKING_BOT_STATE_DIR=/data

# Comma-separated CORS origins; unset allows every origin
CORS_ALLOW_ORIGINS=

# When set, MeetingBaas callbacks to /webhook must carry it in x-mb-secret.
# The server passes it to MeetingBaas as the callback secret when creating bots.
WEBHOOK_SECRET=

Conversation defaults

# Voice-activity detection defaults; a request's turn_config overrides them
VAD_CONFIDENCE=0.5
VAD_START_SECS=0.25
VAD_STOP_SECS=0.8
VAD_MIN_VOLUME=0.6

# Seconds of silence after admission before a bot speaks on its own, and how
# many times it may do so
BOT_IDLE_REENGAGE_SECS=15
BOT_MAX_REENGAGE=3

Context and MCP

# Cap on each prompt_data_sources URL fetch, in bytes
PROMPT_DATA_SOURCE_MAX_BYTES=1000000
# Allow prompt_data_sources URLs on localhost / private networks (trusted networks only)
PROMPT_DATA_ALLOW_PRIVATE_URLS=false

# Allow remote MCP servers on localhost / private networks (trusted networks only),
# or list exact private URLs that are allowed
MCP_ALLOW_PRIVATE_URLS=false
MCP_ALLOWED_PRIVATE_URLS=

# Local mcpproxy groups behind the mcp_profile presets (optional)
MCP_PROXY_ALL_URL=http://127.0.0.1:8109/mcp
MCP_PROXY_PERSONAL_URL=http://127.0.0.1:8110/mcp
MCP_PROXY_PROFESSIONAL_URL=http://127.0.0.1:8111/mcp

MCP server credentials travel per request in mcp.servers[].headers; the server holds no global MCP secret.

Persona creation (optional)

Only for POST /personas/generate-image and the persona CLI:

REPLICATE_KEY=
STABILITY_AI_API_KEY=
UTFS_KEY=
APP_ID=

Local development (optional)

# ngrok tunnels for a public hostname during development
NGROK_AUTHTOKEN=

See Local development and Ngrok setup.