Beta · 0.8.0Open sourceAndroid

The voice of your Hermes agent.

Hermes Vox is an open-source Android app that lets you talk to the Hermes Agent you already run — hands-free, with on-device speech, and a living presence that shows the agent's real work.

Get it on Obtainium Coming soon toGoogle Play

Free · Apache-2.0 · Download the APK · Requires your own Hermes Agent gateway

Why it exists

A voice that feels present — with a real agent behind it.

The best voice AI answers straight away, lets you interrupt, and sounds like someone. Most of it is also a voice with nobody home, or an agent living in somebody else's cloud. Hermes Agent by Nous Research is a real agent you run yourself — tools, long-term memory, skills, a persistent identity. Vox gives it a voice and a face without adding a second brain.

The entity is Hermes

Not a persona on top. Hermes owns the reasoning, tools, memory and context. The phone owns what has to be instant: hearing, turn-taking, interruption, the voice, the presence.

Sovereign by construction

Speech recognition and synthesis run on the phone. The only network peer is the gateway you configure. No accounts, no vendor cloud, no keys baked into the app.

One agent, many doors

Vox is one more frontend to the same Hermes that already answers you on Telegram, desktop or the CLI. Same memory, same skills — it never forks the agent.

See it

No chrome. Just the being and the words.

Hermes Vox during an open call: a quiet cyan nebula of light on true black, with the call timer below
An open call — hands-free, no push-to-talk.
Hermes Vox speaking: the being becomes a violet soundwave while the reply scrolls beneath it in monospace
Speaking — the words follow the voice.
Hermes Vox settings: voice mode, models, entity and connection, speech and mic, transcription, voice, appearance, visuals
Settings — local, and in plain language.

What it does

A real conversation, on your terms.

Hands-free, interruptible

An open line gated by voice activity. Talk over a reply and it stops and listens — and your interrupting words become the next turn.

Hears you on-device

Silero VAD with Whisper (int8) or NVIDIA Parakeet-TDT via sherpa-onnx. Hermes always gets the whole utterance, checked for hallucinations, and a dim “heard” line shows what was sent.

Speaks on-device

Supertonic: ten natural voices at 44.1 kHz, about ten times faster than real time, chosen by a measured bake-off. Piper as a lighter option.

Enhanced Realtime alpha

An on-device Gemma 4 E2B “soul” on the phone's GPU takes the first beat of every turn in the agent's own voice, hears your tone, and hands over to Hermes for anything real.

Bring your own gateway

Your endpoint, your key, your network. Optional per-device scopes give each person their own memory on a shared gateway.

Built to be trusted

Keystore-encrypted key, TLS everywhere, sha256-verified resumable model downloads, local-only logs, transcript logging off by default. Privacy policy.

The being

It shows the agent's real work.

Every tool call Hermes makes reaches the phone, and the being takes a shape for it. Six thousand points of light, drawn on the GPU with real bloom, morphing between shapes at 60 fps.

  1. An amber radar sweepThinking
  2. A blue hourglassWaiting on the web
  3. An amber terminal window typing itself outRunning a command
  4. An amber constellation of linked starsRecalling memory
  5. A rising orange flameAnswer arriving
  6. A violet soundwaveSpeaking

27 shapes, each one exact

Pick one for rest, listening, thinking and speaking — or let it cycle.

  • The aura shape, drawn in cyan points of lightAura
  • The iris shape, drawn in cyan points of lightIris
  • The vortex shape, drawn in cyan points of lightVortex
  • The jellyfish shape, drawn in cyan points of lightJellyfish
  • The globe scan shape, drawn in cyan points of lightGlobe scan
  • The constellation shape, drawn in cyan points of lightConstellation
  • The terminal shape, drawn in cyan points of lightTerminal
  • The flame shape, drawn in cyan points of lightFlame
  • The ribbon shape, drawn in cyan points of lightRibbon
  • The black hole shape, drawn in cyan points of lightBlack hole
  • The bloom shape, drawn in cyan points of lightBloom
  • The soundwave shape, drawn in cyan points of lightSoundwave
  • The lightning shape, drawn in cyan points of lightLightning
  • The nucleus shape, drawn in cyan points of lightNucleus
  • The eye shape, drawn in cyan points of lightEye
  • The ripples shape, drawn in cyan points of lightRipples
  • The radar shape, drawn in cyan points of lightRadar
  • The octopus shape, drawn in cyan points of lightOctopus
  • The sphere shape, drawn in cyan points of lightSphere
  • The helix shape, drawn in cyan points of lightHelix
  • The knot shape, drawn in cyan points of lightKnot
  • The aurora shape, drawn in cyan points of lightAurora
  • The harmonograph shape, drawn in cyan points of lightHarmonograph
  • The tesseract shape, drawn in cyan points of lightTesseract
  • The mandala shape, drawn in cyan points of lightMandala
  • The butterfly shape, drawn in cyan points of lightButterfly
  • The hourglass shape, drawn in cyan points of lightHourglass

How it works

Hermes decides. The phone makes it feel alive.

  • Gemma speaks; Hermes acts. The on-device soul says “let me look that up”. Hermes is the one that looks it up.
  • No tools at the edge. The phone cannot call a tool, by construction.
  • One identity. The soul borrows Hermes' own SOUL.md through an agent-authored voice file, read-only.

Read the architecture →

Status

Beta. Vox is installable and useful today, and it moves fast: expect rough edges and things that change between releases. Testing on real phones, polish and bug reports are the most valuable help right now.

Shipped — 0.8.0

  • Gemma on the GPU; the soul's beat on every turn
  • Supertonic voice; the soul hears your tone
  • Whole-utterance STT, validator, optional Parakeet
  • The being rebuilt on the GPU, 27 shapes
  • Background, resumable model downloads

Being tested now

  • Field verification on real phones
  • Real-world turn-taking and echo tuning
  • Memory and thermal soak tests
  • Google Play listing

Get it

Install Hermes Vox.

Obtainium recommended

Installs the signed APK from each GitHub release and keeps it up to date.

Get it on Obtainium

Or add github.com/chezgoulet/hermes-vox as a source in Obtainium.

APK

Download hermes-vox-0.8.0.apk from GitHub releases and install it. Android 7.0 or later.

Latest release

Or build from source with one command: bash scripts/gate.sh.

Google Play

A Play Store listing is on the way. It will stay a bring-your-own-gateway client.

Coming soon toGoogle Play
  1. Run Hermes Agent. Vox needs a Hermes Agent gateway you control, ideally reached over a private network such as a tailnet.
  2. Connect. Enter the gateway's HTTPS endpoint and its API_SERVER_KEY.
  3. Download the voice. Three on-device models, about 290 MB, in Settings → Models.
  4. Call. Tap the call button and talk.

Thank you

Built on Hermes Agent, by Nous Research.

Hermes Vox exists because Nous Research builds and open-sources Hermes Agent — the mind behind every word Vox speaks. Thank you for making a real, self-hostable agent that anyone can give a voice to.

Also standing on sherpa-onnx, Supertonic, Whisper, Parakeet-TDT, Silero VAD, Piper, Gemma and LiteRT-LM, and the Rajdhani typeface. Hermes Vox is an independent community project, not affiliated with or endorsed by Nous Research.

Questions

FAQ

What is Hermes Vox?

A free, open-source Android app that lets you talk to your own Hermes Agent by voice. The agent does the thinking, tools and memory; the phone handles listening, speaking and interruption on-device.

Do I need my own Hermes Agent?

Yes. Vox is a client, not a service. You point it at a Hermes Agent gateway you run — locally, on your own network, or on a server you control — and enter its endpoint and API key. There is nothing to sign up for.

Does my voice go to the cloud?

No. Speech recognition and text-to-speech run on the phone. The only network peer is the Hermes gateway you configure, plus the one-time model downloads you start.

Is it on Google Play?

Not yet — it's coming soon. Today, install it with Obtainium or download the signed APK from GitHub releases.

Is it finished?

No. It's a beta and a work in progress: installable and useful today, with rough edges still being worked out on real phones. Contributions are welcome.

How can I help?

Try it on your phone and open an issue with what you find, or read CONTRIBUTING.md and send a pull request.