Drop a real-time multimodal AI into any Discord voice channel. Full-duplex audio · vision · function calling · multi-CLI delegation · proactive notifications · post-call transcripts. Built on Google Gemini Multimodal Live, packaged as a self-hostable Hermes Agent plugin. Open source. MIT. Yours.
You've used chatbots. You've used voice assistants that feel like phone trees. This is neither.
Hermes Live puts a genuinely conversational AI into your Discord voice channel — one that hears you, sees your screen, uses your tools, remembers what you talked about last week, and can spin up a Codex session to fix the bug you just described. Sub-second latency. Hour-long sessions. No SaaS, no relay, no vendor lock-in. Just your gateway, your keys, your hardware.
What it feels like:
- You join a voice channel. The agent greets you by name and recalls last week's debugging session.
- You share your screen. It watches, understands, and talks you through the fix in real time.
- You say "ship it" — it delegates to Codex, tracks the PR, and pings you when CI passes.
- You go AFK. It emails you a 3-bucket digest of what happened while you were gone.
- You come back. The conversation picks up exactly where it left off.
This isn't a demo. It's the voice layer your infrastructure has been waiting for.
A proper website is now available in docs-site/ — built from the source markdown in docs/. Open docs-site/index.html in a browser, or serve it with any static host (python3 -m http.server works). 13 pages covering architecture, personality, fallback chain, notifications, email brief, SFX library, webhooks, video feeder, env vars, troubleshooting, and the changelog.
If you prefer raw markdown, every page is also in docs/.
# 1. Install
git clone https://github.com/Capslockb/gemini-live-discord-bridge.git
cd gemini-live-discord-bridge
./install.sh # full install (prompts for env)
./install.sh --from-local # use the current working dir
./install.sh --uninstall # remove
# 2. Restart the gateway
systemctl --user restart hermes-gateway
# 3. From Discord, run:
/voice-live # join your current voice channel
/voice-live-leave # leaveThe installer handles venv, symlinks, env prompts, and SFX directory creation. See install.sh for details.
| 🎙️ Full-duplex voice | Sub-second latency, Discord UDP → Opus → 16 kHz mono → Gemini WSS |
| 👁️ Vision + frame feed | Send images or stream 1 fps screenshare — the model sees what you see |
| 🛠️ Function calling | 30+ voice tools (calendar, mail, Home Assistant, GitHub, Spotify, files, search) |
| 🔁 Multi-CLI delegation | opencode / codex / numasec / gemini / hermes-api with health registry + automatic fallback |
| 📣 Proactive notifications | Voice, DM, channel, webhook, or auto — fires on long-task completion, AFK pings, scheduled alerts |
| 📧 Email brief | Scheduled Gmail digest, 3-bucket importance scoring, AFK delivery |
| 😴 Idle hangup | Two-phase: prompt after N seconds of silence, then auto-leave |
| 📝 JSONL transcripts | Word-level transcripts with tool calls, turns, and idle events |
| 🎵 Bundled sfx library | 4 slots (tool-init / error / notification / transition), env-driven paths |
| 🪶 Self-hostable | No SaaS, no third-party relay. Runs in your existing Hermes gateway's asyncio loop |
| 🩺 Health + control API | Local HTTP on 127.0.0.1:18943 — /health, /frame, /say, /leave |
Discord Voice → Opus Decode → 48kHz PCM → 16kHz Mono → Gemini WSS → Model
↑ │
│ ▼
└──────────── 24kHz PCM ← Gemini WSS ← 48kHz Stereo ← Discord AudioSource
Lies on discord-ext-voice-recv (audio RX) and Gemini Multimodal Live API (WSS). The bridge runs in-process inside the Hermes gateway — no separate services, no queues, no message buses. Full architecture doc: docs/architecture.md.
| Feature | Doc | What it does |
|---|---|---|
| Voice I/O | docs/architecture.md |
Opus in/out, Gemini Live streaming, sidecar HTTP API on 18943 |
| Personality system | docs/personality.md |
14-section system prompt, ping-pong rhythm, boredom switch, vocal expression cap |
| Multi-CLI delegation | docs/fallback-chain.md |
opencode / codex / gemini / numasec / hermes-api with health registry + automatic fallback |
| Proactive notifications | docs/notification.md |
local_notify tool, scheduler, sidecar /notify, AFK DM pings |
| Email brief | docs/email-brief.md |
Scheduled inbox digest, important/fyi/auto buckets, AFK delivery |
| SFX library | docs/sfx-library.md |
4 slots, env-driven paths, local_sfx_test tool |
| Webhooks | docs/webhooks.md |
9 event classes, throttle keys, per-class env-var config |
| Video awareness | docs/architecture.md |
/frame HTTP endpoint, auto-react to video enable/disable |
| Onboarding | — | First-run Q&A for new users, persisted to ~/.hermes/voice-users/<id>.yaml |
| Honcho context | — | Per-user peer memory injected into the system prompt |
| GitHub tools | — | 6 voice tools to manage repos / issues / PRs via the gh CLI |
| Home Assistant | — | Voice-driven HA control |
| Spotify | — | Play/pause/skip/search/volume via voice |
Hermes can now hold a real conversation with you in voice. Not a 30-second demo — sub-second latency, hour-long sessions, remembers what you talked about last time via Honcho memory.
Mid-conversation, it can:
- 🔍 Search the web and read the answer aloud
- 📁 Open your files, review code, suggest fixes
- 📬 Check your email and summarize
- 🎵 Queue Spotify, dim the lights (Home Assistant)
- 🧠 Delegate and track Codex / OpenCode / NumaSec / Hermes (API) sessions
- 👁️ See your screenshare and walk you through a bug
Built in one session. One developer. Shipped.
The minimum required:
DISCORD_BOT_TOKEN=***
GEMINI_API_KEY=***
DISCORD_VOICE_LIVE_USER_ID=1474100257762578597 # your Discord snowflakeFull list of every DISCORD_VOICE_LIVE_* env var: docs/env-vars.md.
Runs on 127.0.0.1:18943:
| Route | Method | Description |
|---|---|---|
/health |
GET | Bridge health JSON |
/frame |
POST | Send a JPEG/PNG frame (?force=true bypasses audio-gate) |
/stop |
GET | Stop the bridge |
/say |
GET | Inject text into Gemini (?text=...) |
/notes |
GET | Recent transcript events (?limit=50) |
/notify |
GET/POST | Proactive notification breakout |
The system prompt is a 14-section behavioral contract, not documentation. Each section addresses a specific regression. Do not add hedging like "be helpful and harmless" — the model interprets that as permission to revert to assistant defaults.
See docs/personality.md for the section index and how to edit.
~$0.03–0.06 / hour of voice on Gemini's Flex tier. Calls cost tokens; a 30-min voice session runs roughly the cost of a long text chat with a generous model.
A proper, designed docs site lives in docs-site/. It's a static site built from the markdown in docs/, so you can host it on GitHub Pages, Vercel, or just python3 -m http.server from the repo.
Individual pages (also browseable as raw markdown):
docs/architecture.md— end-to-end audio path, threading, lifecycledocs/personality.md— system prompt shape and behavioral contractsdocs/fallback-chain.md— multi-CLI delegation with health registrydocs/notification.md— proactive notification breakoutdocs/email-brief.md— scheduled inbox digestdocs/sfx-library.md— slot-based UI sound effectsdocs/webhooks.md— event-class webhook fanoutdocs/video.md— video frame feederdocs/env-vars.md— every env var, defaults, descriptionsdocs/troubleshooting.md— common bridge failures
See CHANGELOG.md for the full release history.
MIT. See top of bridge.py for full text.