Environment reference (2 of 2): brain, realtime, widget, API, behaviour, webhooks, operations

This page lists every environment variable the gateway reads. Set them in .env (or the container environment).

Default is the value used when the variable is unset. Booleans show as on or off; unset means the build has no default. Where a chapter states a default, this table follows the chapter.
<LANG> rows are families: replace <LANG> with the language code upper-cased and - changed to _, so ml-IN becomes ML_IN (for example GOOGLE_VOICE_ML_IN). A few families use the short code, such as _ML; the row says which.
Chapter links to the section that explains the setting.
The list is split in two: this is part 2 (brain, realtime, widget, API, behaviour, webhooks, operations); part 1 is Environment reference (1 of 2).

AI brain

Variable	Default	What it does	Chapter
`CUSTOM_LLM_AUTH_HEADER`	unset	Header name that carries the token when the type is header.	AI brain
`CUSTOM_LLM_AUTH_TOKEN`	unset	Token for bearer or header authentication.	AI brain
`CUSTOM_LLM_AUTH_TYPE`	`none`	none, bearer or header.	AI brain
`CUSTOM_LLM_RESPONSE_FIELD`	unset	Dot path to the reply text when your JSON is not a recognised shape.	AI brain
`CUSTOM_LLM_SESSION_FIELD`	unset	Dot path to your session id in the reply.	AI brain
`CUSTOM_LLM_TIMEOUT`	`30000`	Request timeout in milliseconds.	AI brain
`CUSTOM_LLM_URL`	unset	Your HTTP endpoint. Required when the brain is custom.	AI brain
`LITEBOT_API_URL`	`http://localhost:3000`	Address of the LiteBot server.	AI brain
`LITEBOT_BOT_ID`	`customer-service`	Bot that answers calls.	AI brain
`LITEBOT_ERROR_TEXT_GUARD`	on	Never speak an error that LiteBot returns as the reply.	AI brain
`LITEBOT_ERROR_TEXT_PATTERN`	unset	Regular expression for what counts as a relayed error.	AI brain
`LITEBOT_HANGUP_TOOLS`	`hangupCall`	Comma-separated tool names that end the call.	AI brain
`LITEBOT_RECONCILE_COST`	on	At hang-up, read LiteBot's session total and correct the per-call cost.	AI brain
`LITEBOT_RECONCILE_TIMEOUT_MS`	`1500`	Hard cap on that lookup.	AI brain
`LITEBOT_SESSION_END`	on	Tell LiteBot the session ended so its end-of-session hooks run at once.	AI brain
`LITEBOT_SESSION_END_TIMEOUT_MS`	`5000`	Time allowed for that notification.	AI brain
`LITEBOT_STREAMING`	on	Speak the answer sentence by sentence as it arrives. Off waits for the whole answer.	AI brain
`LITEBOT_TIMEOUT`	`30000`	Request timeout in milliseconds.	AI brain
`LITEBOT_TRANSFER_TOOLS`	unset	Comma-separated tool names that trigger a transfer. Empty relies on the bot's escalate flag.	AI brain
`LLM_FINAL_TIMEOUT_MESSAGE`	built-in English	Spoken on hard AI failure.	AI brain
`LLM_FINAL_TIMEOUT_MESSAGE_ML`	built-in English	Malayalam (ml-IN) variant; other languages use the base message.	AI brain
`LLM_PROVIDER`	required	Always set `litebot` or `custom`. Unset is not a supported configuration.	AI brain
`LLM_TIMEOUT_MESSAGE`	built-in English	Spoken when the AI is slow, while the gateway keeps waiting.	AI brain
`LLM_TIMEOUT_MESSAGE_ML`	built-in English	Malayalam (ml-IN) variant; other languages use the base message.	AI brain
`LLM_TIMEOUT_RETRY_MESSAGE`	built-in English	Spoken when the gateway retries a slow AI request.	AI brain
`LLM_TIMEOUT_RETRY_MESSAGE_ML`	built-in English	Malayalam (ml-IN) variant; other languages use the base message.	AI brain
`TIMEOUT_MESSAGE_ENABLED`	off	Speak the timeout message while a slow reply is still awaited.	AI brain

Realtime mode

Variable	Default	What it does	Chapter
`BOT_REGISTRY_EXTERNAL_DIR`	unset	Folder holding your bots, prompts and tools (config in the install directory).	Realtime mode
`BOT_REGISTRY_HOT_RELOAD`	on	Reload bot files when they change.	Realtime mode
`DEFAULT_BOT_ID`	`default`	Bot used when a session names none (realtime modes).	Realtime mode
`GATEWAY_COMMIT_ON_SILENCE`	off	Commit the audio buffer when silence is detected.	Realtime mode
`GATEWAY_ENABLE_TRANSCRIPTION`	off	Ask the service to transcribe the caller.	Realtime mode
`GATEWAY_INSTRUCTIONS`	`You are a helpful AI assistant. Be concise and natural.`	System instructions for the session.	Realtime mode
`GATEWAY_LOG_TRANSCRIPTS`	off	Write transcripts to the log.	Realtime mode
`GATEWAY_MAX_ERRORS`	`10`	Errors tolerated before the call is dropped.	Realtime mode
`GATEWAY_MAX_TOKENS`	`4096`	Most tokens per response.	Realtime mode
`GATEWAY_MODALITIES`	`audio,text`	Output types requested.	Realtime mode
`GATEWAY_MODE`	off	Bridge every call to a remote realtime service.	Realtime mode
`GATEWAY_TEMPERATURE`	`0.8`	Sampling temperature.	Realtime mode
`GATEWAY_TRANSCRIPTION_MODEL`	`whisper-1`	Model used for that transcription.	Realtime mode
`GATEWAY_VAD_PREFIX_MS`	`300`	Audio kept from before speech starts.	Realtime mode
`GATEWAY_VAD_SILENCE_MS`	`500`	Silence that ends a turn.	Realtime mode
`GATEWAY_VAD_THRESHOLD`	`0.5`	Speech detection sensitivity.	Realtime mode
`GATEWAY_VAD_TYPE`	`server_vad`	Speech detection type.	Realtime mode
`GATEWAY_VOICE`	`alloy`	Voice requested from the service.	Realtime mode
`GEMINI_LIVE_API_KEY`	unset	Gemini Live key. Falls back to `GEMINI_API_KEY`.	Realtime mode
`GEMINI_LIVE_DRAIN_MS`	`15000`	Ceiling for letting the goodbye finish at hang-up.	Realtime mode
`GEMINI_LIVE_DRAIN_STALL_MS`	`300`	Abort the wait if queued audio stops shrinking for this long.	Realtime mode
`GEMINI_LIVE_INBOUND_GAIN`	`2.0`	Gain applied to caller audio before it reaches the model.	Realtime mode
`GEMINI_LIVE_MODEL`	unset	Model used when the bot configuration names none.	Realtime mode
`GEMINI_LIVE_PRICE_AUDIO_IN_PER_1M`	unset	Audio input price, USD per million tokens, for cost tracking. Unset uses the built-in table.	Realtime mode
`GEMINI_LIVE_PRICE_AUDIO_OUT_PER_1M`	unset	Audio output price, USD per million tokens. Unset uses the built-in table.	Realtime mode
`GEMINI_LIVE_PRICE_TEXT_IN_PER_1M`	unset	Text input price, USD per million tokens. Unset uses the built-in table.	Realtime mode
`GEMINI_LIVE_PRICE_TEXT_OUT_PER_1M`	unset	Text output price, USD per million tokens. Unset uses the built-in table.	Realtime mode
`GEMINI_LIVE_TRANSCRIPTS`	on	Send caller and bot transcripts to logs and webhooks.	Realtime mode
`GEMINI_LIVE_VAD`	unset	Empty uses the model's own speech detection; silero uses local detection.	Realtime mode
`GEMINI_LIVE_VOICE`	unset	Voice used when the bot configuration names none.	Realtime mode
`LITEBOT_LIVE_API_KEY`	unset	LiteBot server key used to authenticate.	Realtime mode
`LITEBOT_LIVE_BOT_ID`	unset	Bot to start on LiteBot. Falls back to `LITEBOT_BOT_ID`.	Realtime mode
`LITEBOT_LIVE_CONTINUOUS_STREAM`	on	Keep a steady audio stream to LiteBot even when the handset sends nothing during silence.	Realtime mode
`LITEBOT_LIVE_DRAIN_MS`	`15000`	Ceiling for letting the goodbye finish when the session ends.	Realtime mode
`LITEBOT_LIVE_DRAIN_STALL_MS`	`300`	Abort the wait if queued audio stops shrinking for this long.	Realtime mode
`LITEBOT_LIVE_FRAME_MS`	`20`	Length in milliseconds of each audio frame sent to LiteBot.	Realtime mode
`LITEBOT_LIVE_INBOUND_GAIN`	unset	Gain on caller audio. Falls back to `GEMINI_LIVE_INBOUND_GAIN`.	Realtime mode
`LITEBOT_LIVE_INBOUND_RATE`	`16000`	Sample rate of caller audio; use 8000 if the bot is set up for 8 kHz.	Realtime mode
`LITEBOT_LIVE_KICKOFF`	on	Ask the bot to greet first once the session is ready.	Realtime mode
`LITEBOT_LIVE_MAX_BACKLOG_FRAMES`	`25`	Frames that may queue toward LiteBot before older ones are dropped.	Realtime mode
`LITEBOT_LIVE_PING_MS`	`30000`	Keep-alive interval; 0 disables it.	Realtime mode
`LITEBOT_LIVE_TOKEN`	unset	A JWT, as an alternative to the server key.	Realtime mode
`LITEBOT_LIVE_TRANSFER_TOOLS`	`escalateToHuman`	Comma-separated tool names whose execution triggers a transfer.	Realtime mode
`LITEBOT_LIVE_URL`	unset	Base address of LiteBot (`ws://` or `wss://`, host and port); the gateway adds `/voice`. Required for LiteBot Live.	Realtime mode
`LLM_API_KEY`	unset	Key sent to the remote realtime service in gateway mode.	Realtime mode
`LLM_WEBSOCKET_URL`	unset	WebSocket address of the realtime service. Required in gateway mode.	Realtime mode
`REALTIME_PROVIDER`	unset	Empty runs the speech-recognition, brain and voice pipeline; gemini-live or litebot selects a realtime mode.	Realtime mode

Variable	Default	What it does	Chapter
`WEBSOCKET_AUDIO_ALLOWED_BOT_IDS`	unset	Comma-separated bot ids a page may choose. Empty allows any.	Browser widget
`WEBSOCKET_AUDIO_ALLOWED_ORIGINS`	unset	Comma-separated allowed origins. Empty allows any.	Browser widget
`WEBSOCKET_AUDIO_API_KEY`	unset	When set, a page must send it as `?apiKey=` or an `x-api-key` header. Unset means no authentication.	Browser widget
`WEBSOCKET_AUDIO_ENABLED`	off	Start the browser widget endpoint.	Browser widget
`WEBSOCKET_AUDIO_GREETING_MAX_CHARS`	`500`	Longest greeting text a page may send; longer text is cut.	Browser widget
`WEBSOCKET_AUDIO_HOST`	`0.0.0.0`	Address to listen on. Use 127.0.0.1 when a proxy on the same host fronts it.	Browser widget
`WEBSOCKET_AUDIO_IDLE_TIMEOUT_MS`	`300000`	Close a session with no audio or control messages for this long. 0 disables.	Browser widget
`WEBSOCKET_AUDIO_MAX_PAYLOAD_BYTES`	`1048576`	Largest single message accepted, in bytes.	Browser widget
`WEBSOCKET_AUDIO_MAX_PER_IP`	`5`	Simultaneous connections allowed from one client address.	Browser widget
`WEBSOCKET_AUDIO_MAX_SESSION_MS`	`3600000`	Hard limit on one session's length. 0 disables.	Browser widget
`WEBSOCKET_AUDIO_PORT`	`8082`	Port the endpoint listens on.	Browser widget
`WS_AUDIO_TTS_STREAMING`	on	Stream Google Chirp3-HD replies sentence by sentence to the browser. Off sends one message per sentence.	Browser widget

HTTP API and outbound

Variable	Default	What it does	Chapter
`HTTP_ALLOWED_IPS`	unset	Comma-separated IPs and IPv4 CIDR blocks allowed to call the API. Unset turns the whitelist off.	HTTP API
`HTTP_HOST`	`127.0.0.1`	Address the API listens on.	HTTP API
`HTTP_PORT`	`8080`	Port the HTTP API listens on. The fallback collides with `AUDIOSOCKET_PORT`, so set it explicitly (this manual uses 8081).	HTTP API
`HTTP_TRUST_PROXY`	off	Read the client address from `X-Forwarded-For`. Set true only behind a reverse proxy.	Browser widget
`OUTBOUND_DEFAULT_GREETING_<LANG>`	built-in text	Greeting text for dynamic outbound greetings in one language, for example `OUTBOUND_DEFAULT_GREETING_ML_IN`.	Conversation behaviour
`OUTBOUND_ENABLED`	off	Open the outbound-call routes; otherwise they return 403.	HTTP API
`OUTBOUND_GREETING_MODE`	`static`	Default greeting mode for outbound calls: static, dynamic or hybrid.	HTTP API
`OUTBOUND_ORIGINATE_ENDPOINT`	unset	Your webhook that places the call.	HTTP API
`OUTBOUND_STATIC_AUDIO_<LANG>`	unset	WAV played in static and hybrid outbound mode for one language, for example `OUTBOUND_STATIC_AUDIO_ML_IN`.	Conversation behaviour
`OUTBOUND_TIMEOUT`	`10000`	Milliseconds to wait for that webhook.	HTTP API

Conversation behaviour

Variable	Default	What it does	Chapter
`BARGE_IN_DELAY`	`200`	Milliseconds at the start of playback during which interruption is ignored.	Conversation behaviour
`BARGE_IN_ECHO_GUARD`	on	Ignore the bot's own voice leaking back from the handset.	Conversation behaviour
`BARGE_IN_ECHO_LOG`	on	Log each ignored echo transcript.	Conversation behaviour
`BARGE_IN_ECHO_MARGIN_DB`	`8`	How many dB louder than the measured echo the caller must be.	Conversation behaviour
`BARGE_IN_ECHO_MAX_THRESHOLD`	`24000`	Upper limit for the echo-adjusted volume threshold.	Conversation behaviour
`BARGE_IN_ECHO_MAX_WORDS`	`3`	Longest transcript treated as echo. Longer ones, doubled words and yes or no words count as the caller.	Conversation behaviour
`BARGE_IN_ECHO_TAIL_MS`	`800`	After playback ends, transcripts that are the bot's own words are dropped for this long.	Conversation behaviour
`BARGE_IN_ECHO_WARMUP_PACKETS`	`10`	Audio packets at the start of playback in which interruption is ignored.	Conversation behaviour
`BARGE_IN_ECHO_WINDOW_MS`	`4000`	Window over which the echo level is measured, sampled only while the bot plays.	Conversation behaviour
`BARGE_IN_FALSE_TIMEOUT_MS`	`1500`	How long to wait for words before resuming. 0 disables holding.	Conversation behaviour
`BARGE_IN_MAX_RESUMES`	`2`	Resumes per turn before the tail is discarded.	Conversation behaviour
`BARGE_IN_MIN_WORDS`	`0`	Words needed to confirm an interruption. 0 means any transcript that passes the gate confirms.	Conversation behaviour
`BARGE_IN_RESET_HYSTERESIS`	`5`	Quiet frames needed before a new interruption can be counted.	Conversation behaviour
`BARGE_IN_RESUME`	on	Hold and resume after a false interruption. Off discards the unplayed tail every time.	Conversation behaviour
`BARGE_IN_THRESHOLD`	`500`	Volume threshold used when the detector is off.	Conversation behaviour
`BARGE_IN_USE_VAD`	on	Decide interruptions with the speech detector, not raw volume.	Conversation behaviour
`BARGE_IN_VAD_THRESHOLD`	`0.5`	Detector confidence needed to count as speech.	Conversation behaviour
`ENABLE_BARGE_IN`	off	Let callers interrupt the bot.	Conversation behaviour
`ENABLE_INCOMPLETE_TURN_DETECTION`	off	Honour the wait-for-more and give-time flags from the brain.	Conversation behaviour
`GOODBYE_MESSAGE`	unset	WAV played at a bot-initiated hang-up when the brain gave no closing words.	Conversation behaviour
`GOODBYE_MESSAGE_TEXT_EN`	`Thank you for calling. Goodbye!`	Spoken goodbye text when no WAV exists.	Conversation behaviour
`GOODBYE_MESSAGE_TEXT_<LANG>`	unset	Spoken goodbye text for one language, short code (`EN`, `ML`, `HI`).	Conversation behaviour
`GOODBYE_MESSAGE_<LANG>`	unset	Goodbye WAV for one language; the short code is used, for example `GOODBYE_MESSAGE_ML`. Beats `GOODBYE_MESSAGE`.	Conversation behaviour
`LLM_HANGUP_DRAIN_MAX_MS`	`15000`	Longest the gateway waits for the goodbye audio to play out.	Conversation behaviour
`MAX_SILENCE_DURATION`	`1000`	Milliseconds of silence before the caller's speech is processed.	Conversation behaviour
`MIN_SPEECH_DURATION`	`200`	Minimum milliseconds of speech before processing.	Conversation behaviour
`MIN_UTTERANCE_LENGTH`	`2`	Shortest utterance, in characters, kept; shorter ones are treated as noise.	Conversation behaviour
`NOISE_FLOOR_CALIBRATION_MS`	`1000`	Listening-only period at the start of the call.	Conversation behaviour
`NOISE_FLOOR_ENABLED`	on	Learn each call's background level and set the interruption threshold from it.	Conversation behaviour
`NOISE_FLOOR_LOG`	on	Log the learned level.	Conversation behaviour
`NOISE_FLOOR_LOG_INTERVAL_MS`	`5000`	Interval between those log lines.	Conversation behaviour
`NOISE_FLOOR_MARGIN_DB`	`6`	How far above the quiet level the threshold sits.	Conversation behaviour
`NOISE_FLOOR_MAX_THRESHOLD`	`2000`	Highest threshold allowed.	Conversation behaviour
`NOISE_FLOOR_MIN_THRESHOLD`	`200`	Lowest threshold allowed.	Conversation behaviour
`NOISE_FLOOR_ROLLING_WINDOW_MS`	`10000`	Window used to track the quiet level afterwards.	Conversation behaviour
`PIPELINE_INBOUND_GAIN`	`2.5`	Gain applied to caller audio before recognition, speech detection and interruption checks.	Conversation behaviour
`PROCESSING_SOUND_DIR`	unset	Folder holding filler files.	Conversation behaviour
`PROCESSING_SOUND_ENABLED`	off	Play a filler sound while the AI works.	Conversation behaviour
`PROCESSING_SOUND_FILE`	unset	Your own filler file, absolute or relative to the install directory.	Conversation behaviour
`PROCESSING_SOUND_HARD_CUT`	on	Cut the filler the instant speech is ready. Off lets the phrase finish.	Conversation behaviour
`PROCESSING_SOUND_LOOP`	off	Repeat the sound while waiting.	Conversation behaviour
`PROCESSING_SOUND_LOOP_DELAY`	`2000`	Milliseconds between repeats.	Conversation behaviour
`PROCESSING_SOUND_START_DELAY`	`2000`	Milliseconds of waiting before the sound starts.	Conversation behaviour
`PROCESSING_SOUND_TRANSITION_DELAY`	unset	Pause in milliseconds after the filler before speech starts.	Conversation behaviour
`PROCESSING_SOUND_TYPE`	`chime`	Built-in filler sound to use.	Conversation behaviour
`PROCESSING_SOUND_VOLUME`	`0.3`	Playback volume.	Conversation behaviour
`REDEMPTION_MS`	`1400`	Wait after the last transcript before the turn is processed.	Conversation behaviour
`TRANSFER_ERROR_MESSAGE`	unset	WAV played when the transfer fails or is disabled.	Conversation behaviour
`TRANSFER_ERROR_MESSAGE_<LANG>`	unset	Transfer-failure WAV for one language, short code (`TRANSFER_ERROR_MESSAGE_ML`).	Conversation behaviour
`TRANSFER_ERROR_TEXT_EN`	`I am sorry, I was unable to transfer you. Please try calling again.`	Spoken text when no error WAV exists.	Conversation behaviour
`TRANSFER_ERROR_TEXT_<LANG>`	unset	Transfer-failure text for one language, short code (`EN`, `ML`).	Conversation behaviour
`TRANSFER_MESSAGE`	unset	WAV played before transferring.	Conversation behaviour
`TRANSFER_MESSAGE_TEXT_EN`	`Please hold while I transfer you to a human agent.`	Spoken text when no WAV exists.	Conversation behaviour
`TRANSFER_MESSAGE_TEXT_<LANG>`	unset	Pre-transfer text for one language, short code (`EN`, `ML`).	Conversation behaviour
`TRANSFER_MESSAGE_<LANG>`	unset	Pre-transfer WAV for one language, short code (`TRANSFER_MESSAGE_ML`).	Conversation behaviour
`UTTERANCE_CHECK_INTERVAL`	`500`	Milliseconds between checks that the system is idle.	Conversation behaviour
`UTTERANCE_COMBINATION_WINDOW`	`10000`	Transcripts arriving within this window are combined.	Conversation behaviour
`UTTERANCE_MAX_WAIT`	`30000`	Longest wait for the system to become idle.	Conversation behaviour
`UTTERANCE_TIMEOUT`	`3000`	Milliseconds to wait for more speech before a turn is closed.	Conversation behaviour
`UTTERANCE_WINDOW_SIZE`	`5`	Most utterances buffered, to bound memory.	Conversation behaviour
`VAD_BUFFER_INTERVAL`	`200`	Milliseconds of audio between detector runs.	Conversation behaviour
`VAD_MAX_SPEECH_MS`	`6000`	Longest single speech segment before it is cut.	Conversation behaviour
`VAD_MIN_SPEECH_FRAMES`	`4`	Speech frames the detector needs to open a segment.	Conversation behaviour
`VAD_MIN_SPEECH_GATE_ADAPTIVE`	on	Discard segments that are both short and quiet.	Conversation behaviour
`VAD_MIN_SPEECH_GATE_DURATION_MS`	`500`	Segments shorter than this are candidates for discarding.	Conversation behaviour
`VAD_MIN_SPEECH_GATE_ENERGY_MULT`	`2.0`	A candidate is discarded if its peak is below the noise floor times this value.	Conversation behaviour
`VAD_MODEL`	`v5`	Speech detector model version.	Conversation behaviour
`VAD_NEGATIVE_THRESHOLD`	`0.35`	Confidence below which speech stops. Must be lower than the start value.	Conversation behaviour
`VAD_POSITIVE_THRESHOLD`	`0.5`	Confidence that starts speech.	Conversation behaviour
`VAD_PRE_SPEECH_FRAMES`	`1`	Frames kept from before speech starts.	Conversation behaviour
`VAD_REDEMPTION_FRAMES`	`8`	Quiet frames tolerated before speech is considered ended.	Conversation behaviour

Webhooks and billing

Variable	Default	What it does	Chapter
`BILLING_CDR_AUTH_HEADER`	`Authorization`	Header that carries the token.	Webhooks and billing
`BILLING_CDR_AUTH_SCHEME`	`Bearer`	Prefix before the token. Empty sends a bare token.	Webhooks and billing
`BILLING_CDR_AUTH_TOKEN`	unset	Token sent with each delivery.	Webhooks and billing
`BILLING_CDR_DIR`	`./cdr`	Store root, relative to the working directory.	Webhooks and billing
`BILLING_CDR_ENABLED`	off	Master switch for billing records.	Webhooks and billing
`BILLING_CDR_KEEP_DELIVERED`	off	Archive delivered records in sent/.	Webhooks and billing
`BILLING_CDR_MAX_ATTEMPTS`	`10`	Attempts before a record is moved to dead letters.	Webhooks and billing
`BILLING_CDR_SECRET`	unset	Signing key, kept separate from `WEBHOOK_SECRET`.	Webhooks and billing
`BILLING_CDR_SWEEP_MS`	`60000`	How often the gateway scans for records due a retry.	Webhooks and billing
`BILLING_CDR_TIMEOUT_MS`	`5000`	Milliseconds to wait for your endpoint.	Webhooks and billing
`BILLING_CDR_URL`	unset	Billing endpoint. Empty means store only.	Webhooks and billing
`COST_RATES_FILE`	unset	Path to your rates file. Empty reads `config/pricing.json`.	Webhooks and billing
`COST_TRACKING_ENABLED`	on	Compute a cost for each call.	Webhooks and billing
`COST_USD_INR`	`87`	Conversion rate from US dollar list prices to rupees.	Webhooks and billing
`STATUS_WEBHOOK_MAX_PENDING`	`500`	Status-update deliveries allowed in flight before new ones are skipped.	Webhooks and billing
`WEBHOOK_AUTH_HEADER`	`Authorization`	Header that carries the token.	Webhooks and billing
`WEBHOOK_AUTH_SCHEME`	`Bearer`	Prefix before the token. Empty sends a bare token.	Webhooks and billing
`WEBHOOK_AUTH_TOKEN`	unset	Token sent with every request. Empty means no auth header.	Webhooks and billing
`WEBHOOK_MAX_PENDING`	`200`	Deliveries allowed in flight; beyond this new events are skipped.	Webhooks and billing
`WEBHOOK_METHOD`	`POST`	HTTP method used for delivery.	Webhooks and billing
`WEBHOOK_SECRET`	unset	Signing key. Empty means requests are unsigned.	Webhooks and billing
`WEBHOOK_SSRF_CHECK`	on	Refuse a URL that resolves to a private or loopback address. Set false for a receiver on your local network.	Webhooks and billing
`WEBHOOK_TIMEOUT`	`5000`	Milliseconds to wait for your endpoint.	Webhooks and billing
`WEBHOOK_TOOL_DETAILS`	off	Include the tool-call list in the call-ended payload.	Webhooks and billing
`WEBHOOK_TOOL_MAX_BYTES`	`2048`	Cap per field; longer values are cut and flagged.	Webhooks and billing
`WEBHOOK_TOOL_MAX_ENTRIES`	`50`	Tool entries kept per call; the rest are counted.	Webhooks and billing
`WEBHOOK_TOOL_PAYLOADS`	off	Also include tool arguments and results, which may hold caller personal data.	Webhooks and billing
`WEBHOOK_URL`	unset	Your lifecycle endpoint. Empty disables webhooks.	Webhooks and billing

Operations and tuning

Variable	Default	What it does	Chapter
`LOG_LATENCY`	off	Log a per-turn response-time line and write logs/performance.csv.	Operations
`LOOP_LAG_WARN_MS`	`40`	Log a stall warning when the process falls this many milliseconds behind.	Operations
`LOOP_PROFILE_DIR`	`/tmp`	Folder where those profiles are written.	Operations
`LOOP_PROFILE_MS`	`0`	Record a CPU profile of this many milliseconds on demand to help attribute stalls. 0 disables.	Operations
`NATIVE_RESAMPLER_AUTO_DISABLE_ON_ERROR_RATE`	`0.01`	Error rate, 0 to 1, at which the built-in resampler switches itself off.	Operations
`RTP_PACER_LEAD_MS`	`350`	Audio cushion for outbound SIP audio.	Operations
`RTP_PACER_WORKER`	on	Run the audio pacer in its own thread. Off paces on the main thread.	Operations
`TX_CATCHUP_SLACK_MS`	`200`	How far behind schedule a sender may fall before it catches up. Leave it.	Operations
`USE_NATIVE_RESAMPLER`	`false`	false uses FFmpeg only; true uses the built-in resampler with fallback; compare runs both.	Operations

Diagnostic and test-only

These switches exist for support and testing. Leave them unset unless Vexyl asks.

Variable	What it does
`GATEWAY_LOG_RATE_LIMITS`	Log the realtime service's rate-limit updates.
`GEMINI_LIVE_AUDIO_DEBUG`	Log Gemini Live audio details for diagnosis.
`RECONCILE_LIVE_SESSION`	Test switch for the end-of-call cost lookup.
`TEST_GROQ_ON_START`	Run a Groq speech-recognition self-test at start-up.
`TEST_ITERATIONS`	Repeat count for the start-up self-tests.
`TEST_LANGUAGE`	Language used by the start-up self-tests.
`TEST_OPENAI_STT_ON_START`	Run an OpenAI speech-recognition self-test at start-up.
`TEST_STT_ON_START`	Run a speech-recognition self-test at start-up.
`TEST_TTS_ON_START`	Run a voice self-test at start-up.

Environment (1 of 2) Cost and quality