Environment variables

All DISCORD_VOICE_LIVE_* env vars the plugin reads. Defaults shown in bold.

Required

Var	Description
`DISCORD_BOT_TOKEN`	Discord bot token
`GEMINI_API_KEY`	Google Gemini API key
`DISCORD_VOICE_LIVE_USER_ID`	Your Discord snowflake — the bridge listens to this user

Core

Var	Default	Description
`GEMINI_MODEL`	`gemini-3.1-flash-live-preview`	Primary Gemini Live model
`GEMINI_LIVE_MODEL_FALLBACKS`	—	Comma-separated fallback models, tried in order if primary fails
`DISCORD_VOICE_LIVE_VOICE`	`en-US-JennyNeural`	TTS voice (criterion #3 — high-pitched female)
`DISCORD_VOICE_LIVE_PORT`	`18943`	Sidecar HTTP control port
`DISCORD_VOICE_LIVE_ALLOWED_SPEAKERS`	empty	Comma-separated list of user IDs the bridge accepts audio from. Empty = listen to the channel.
`DISCORD_VOICE_LIVE_AUTO_LEAVE_QUIET_SECONDS`	`900`	Idle timeout (15 min) before the bridge auto-leaves
`DISCORD_VOICE_LIVE_AUTO_LEAVE_MIN_UPTIME_SECONDS`	`30`	Don't auto-leave within the first 30s of a session (avoids killing right after start)
`DISCORD_VOICE_LIVE_LEAVE_PHRASES`	(built-in list)	Phrases that trigger `/voice-live-leave` (e.g. "stop", "hang up", "bye", "exit voice")
`DISCORD_VOICE_LIVE_GREETING`	—	Optional greeting text on first turn
`DISCORD_VOICE_LIVE_CLEAR_ON_INTERRUPT`	`true`	When user interrupts the model, clear the audio queue
`DISCORD_VOICE_LIVE_NOTES_DIR`	`~/.hermes/voice-live-notes/`	Where to write per-call notes
`DISCORD_VOICE_LIVE_KEEP_AUTOSTART_FILE`	`false`	If true, the autostart file is not deleted after use
`DISCORD_VOICE_LIVE_AUTOSTART`	`false`	Auto-join the channel in `voice-live-autostart.json` on gateway boot
`DISCORD_VOICE_LIVE_AUTOSTART_FILE`	`~/.hermes/voice-live-autostart.json`	Path to the autostart file
`DISCORD_VOICE_LIVE_GUILD_ID`	—	Guild ID for autostart; required if autostart is enabled
`DISCORD_VOICE_LIVE_CHANNEL_ID`	—	Voice channel ID for autostart; required if autostart is enabled

Voice output

Var	Default	Description
`GEMINI_AUDIO_STREAM_IDLE_END_SECONDS`	`0.25`	Time of audio silence before the model considers the user turn ended
`DISCORD_VOICE_LIVE_OUTPUT_PREROLL_MS`	`200`	Pre-roll audio before first byte lands in Discord
`DISCORD_VOICE_LIVE_OUTPUT_TAIL_PAD_MS`	`120`	Tail padding after last byte (prevents click on natural ends)
`DISCORD_VOICE_LIVE_OUTPUT_FADE_IN_MS`	`60`	Fade-in on each chunk (prevents click on joins)
`DISCORD_VOICE_LIVE_OUTPUT_READ_WAIT_SECONDS`	`0.05`	How long `LiveAudioSource.read()` blocks waiting for the next chunk

Idle prompts

Var	Default	Description
`DISCORD_VOICE_LIVE_IDLE_PROMPT_SECONDS`	`120`	Seconds of inactivity before the model generates a nudge
`DISCORD_VOICE_LIVE_IDLE_PROMPT_GRACE_SECONDS`	`30`	Initial grace period after session start before nudging
`DISCORD_VOICE_LIVE_IDLE_PROMPT_TEXT`	(built-in)	The nudge prompt injected after idle timeout

SFX library

See sfx-library.md for full list. Highlights:

Var	Default	Description
`DISCORD_VOICE_LIVE_SFX_ENABLED`	`true`	Master enable
`DISCORD_VOICE_LIVE_SFX_DIR`	`~/.hermes/voice-users/sfx/`	Default sfx directory
`DISCORD_VOICE_LIVE_SFX_<SLOT>`	per-slot	Per-slot WAV path override
`DISCORD_VOICE_LIVE_SFX_<SLOT>_VOLUME`	per-slot	Per-slot volume (0.0-1.5)

Slots: TOOL_INIT, ERROR, NOTIFICATION, TRANSITION.

Typing sfx (legacy single-slot)

Var	Default	Description
`DISCORD_VOICE_LIVE_TYPING_SOUND`	`true`	Enable the keyboard click sfx on tool calls
`DISCORD_VOICE_LIVE_TYPING_SFX`	`~/.hermes/voice-live-typing.wav`	Path to the WAV
`DISCORD_VOICE_LIVE_TYPING_SFX_VOLUME`	`0.35`	Volume
`DISCORD_VOICE_LIVE_TYPING_SYNTH_FALLBACK`	`false`	If true and the WAV is missing, generate a synthetic click instead of going silent

Video

Var	Default	Description

DISCORD_VOICE_LIVE_VIDEO_MAX_FPS | 1 | Hard cap on feeder frame rate. Gemini rejects faster anyway. |
DISCORD_VOICE_LIVE_VIDEO_MAX_BYTES | 524288 | Per-frame JPEG size cap (512 KB). |
DISCORD_VOICE_LIVE_VIDEO_WHEN_RECENT_AUDIO_SECONDS | 8 | Drop frames if no voice activity in the last N seconds. |
DISCORD_VOICE_LIVE_VIDEO_INITIALIZED_QUIET_THRESHOLD_S | 30 | Webhook announce fires only when a frame is accepted after at least this many seconds of silence. |
DISCORD_VOICE_LIVE_VIDEO_ENABLED | true | Allow video frame input |
DISCORD_VOICE_LIVE_VIDEO_STATE_DETECTION | true | Auto-react to video enable/disable |
DISCORD_VOICE_LIVE_VIDEO_STATE_POLL_INTERVAL_SECONDS | 2.0 | Poll interval for video state changes |

Tool enable/disable

Var	Default	Description
`DISCORD_VOICE_LIVE_LOCAL_TOOLS`	`true`	All local tools (umbrella)
`DISCORD_VOICE_LIVE_WEB_TOOLS`	`true`	Web search / extract
`DISCORD_VOICE_LIVE_SPOTIFY_TOOLS`	`true`	Spotify playback
`DISCORD_VOICE_LIVE_GITHUB_TOOLS`	`true`	GitHub repo / issue / PR tools
`DISCORD_VOICE_LIVE_HA_TOOLS`	`true`	Home Assistant
`DISCORD_VOICE_LIVE_OPENCODE_TOOLS`	`true`	Opencode delegation
`DISCORD_VOICE_LIVE_SYSINSPECT_TOOLS`	`true`	System inspection
`DISCORD_VOICE_LIVE_EMAIL_TOOLS`	`true`	Email read / send / reply / brief

Webhooks

See webhooks.md. One env var per event class, all start with DISCORD_VOICE_LIVE_WEBHOOK_<CLASS>.

Var	Default	Description
`DISCORD_VOICE_LIVE_WEBHOOK_THROTTLE_SECONDS`	`60`	Default throttle window when `throttle_key` is set

Email brief

See email-brief.md.

Var	Default	Description
`DISCORD_VOICE_LIVE_EMAIL_BRIEF_ENABLED`	`true`	Enable the scheduler
`DISCORD_VOICE_LIVE_EMAIL_BRIEF_INTERVAL_SECONDS`	`1800`	30 min default
`DISCORD_VOICE_LIVE_EMAIL_BRIEF_LIMIT`	`8`	Max emails per brief

Per-email reminder loop

Var	Default	Description
`DISCORD_VOICE_LIVE_EMAIL_REMINDER_ENABLED`	`true`	Enable per-email pings
`DISCORD_VOICE_LIVE_EMAIL_REMINDER_POLL_SECONDS`	`120`	2 min poll interval
`DISCORD_VOICE_LIVE_EMAIL_REMINDER_MAX_PER_HOUR`	`3`	Cap pings per hour to avoid spam

Notification system

Var	Default	Description
`DISCORD_VOICE_LIVE_NOTIFY_TIMEOUT`	`10`	HTTP timeout for webhook delivery (s)

Honcho integration

Var	Default	Description
`VOICE_LIVE_HONCHO_CONTEXT`	`true`	Inject Honcho context into the system prompt
`VOICE_LIVE_HONCHO_MAX_CHARS`	`1200`	Cap Honcho context block size
`VOICE_LIVE_HONCHO_PEER`	(user_id)	Override the Honcho peer name

Opencode delegation

Var	Default	Description
`OPENCODE_BIN`	`~/.local/bin/opencode`	Path to opencode binary
`OPENCODE_DEFAULT_MODEL`	(opencode default)	Model passed to opencode
`OPENCODE_TMUX_SESSION`	`voice-opencode`	Tmux session name
`DISCORD_VOICE_LIVE_OPENCODE_WATCHER`	`true`	Watch opencode tmux sessions for status changes
`DISCORD_VOICE_LIVE_OPENCODE_WATCHER_POLL_SECONDS`	`2.0`	Poll interval
`DISCORD_VOICE_LIVE_OPENCODE_WATCHER_INITIAL_DELAY_SECONDS`	`5`	Delay before first poll after session start
`DISCORD_VOICE_LIVE_OPENCODE_WATCHER_MIN_VOICE_GAP_SECONDS`	`10`	Minimum gap between narrations to avoid spam

Misc

Var	Default	Description
`VOICE_USERS_DIR`	`~/.hermes/voice-users/`	Per-user profile directory
`VOICE_OWNER_DISCORD_ID`	(env)	Used for owner-only commands
`HERMES_PYTHON`	`python3`	Python interpreter for subprocess calls
`GOOGLE_API_BIN`	(auto-detected)	Path to `google_api.py` for email + Google Workspace
`HASS_URL`	`http://homeassistant.local:8123`	Home Assistant base URL
`HASS_TOKEN`	—	Home Assistant long-lived access token

Hermes Live v0.3.5 · MIT · github.com/Capslockb/hermes-live-discord-agent-plugin