{ hermes-voice-face-stack }: give the Hermes agent a voice and a face on one Windows box.
Hold Home, talk, and a circuit-board face goes LISTENING → THINKING → SPEAKING while it answers out loud. Whisper does the ears, Kokoro does the mouth, both run locally. This README is also the runbook — an agent can follow it end to end.
Note
Built on the open work of Jared Rhodenizer (@jaredrhod). This repo pulls in backtalk and ai-visualizer rather than redistributing them; both are AGPL-3.0. The only new part is a small shim that swaps the brain from Claude to Hermes.
| Feature | What it does | |
|---|---|---|
| 🎙️ | Push to talk | Hold Home, speak, release. Whisper transcribes, Hermes answers, Kokoro speaks. |
| 🧠 | Hermes brain | Routes the conversation to Hermes' own /v1/chat/completions endpoint on :8642. |
| 🖥️ | Live face | The board animates on every state change, driven by a signal bus on disk. |
| 📦 | One box | Ears, mouth, brain and face all on the same Windows machine. No Linux homelab needed. |
[ you hold Home ]
| Whisper STT
v
backtalk (ears + mouth)
|
v
Hermes brain (:8642)
|
v
.voice_state / .voice_waveform -> ai-visualizer face
backtalk writes its signals straight into the face's bus_dir, so the two never need a network hop. The shim (scripts/hermes_brain.py) is the only piece that knows about Hermes.
Needs: Windows 10/11 · a microphone · Hermes installed · Python 3 · uv · ~1 GB free for the voice models
- Turn on Hermes' API: set
API_SERVER_KEY,API_SERVER_PORT=8642,API_SERVER_HOST=127.0.0.1in its environment, then restart the gateway.GET /v1/modelsshould answer. - Clone
ai-visualizerinto a folder such asC:\my-agent, writeai-visualizer.jsonwithbus_dirpointing at that folder, and start the face:python server.py --no-open. - Clone
backtalk, copyscripts/hermes_brain.pyinto it, changebacktalk/main.pyto importWarmBrainfrombacktalk.hermes_brain, and setsignals_dirinbacktalk.jsonto the same folder as the face'sbus_dir. - Set
HERMES_API_URLandHERMES_API_KEYas persistent user environment variables, then open a new terminal. - Run
uv run python -m backtalk.mainand open the face page on port 8790.
- If nothing happens on Home, check the shim is actually loaded and the API key is set in the same shell.
- The face needs the visualizer's dev server running; the shim needs Hermes' gateway running. Two terminals, or two startup entries.
- Upstream
backtalkandai-visualizerare AGPL-3.0 — keep their terms if you fork or redistribute. The shim in this repo is MIT.






