Hermes Vox

What Hermes Vox does

Hermes Vox is a thin client. Hermes owns the reasoning, tools, memory and context, while the phone captures your voice, shows Hermes's work as a visual presence and speaks the answer. There is no hosted service: you enter the endpoint and API key of a Hermes gateway you run yourself, whether local, on your own network or on a remote machine you own. Speech processing runs on the device, using Whisper for speech-to-text, Silero for voice activity detection and Piper for text-to-speech.

Two voice modes can be chosen in Settings. Realtime is a hands-free line gated by voice activity detection, with barge-in, and Enhanced Realtime (alpha) adds an on-device Gemma 4 E2B expression layer. The README's rule is that Hermes decides and Gemma expresses: Gemma may generate conversation but never runs the agentic workflow or calls tools. The interface shows a generative particle being that reacts to the agent's real tool calls, and replies appear as a crawl synced with speech.

Key features

• Native Android client for a self-run Hermes gateway
• On-device Whisper speech-to-text, Silero VAD and Piper text-to-speech
• Hands-free Realtime mode with barge-in
• Enhanced Realtime (alpha) with an on-device Gemma 4 E2B expression layer
• Particle-being visuals that react to real tool calls
• Endpoint and API key entered in the app, with nothing bundled

When to use it

• Talking hands-free to a Hermes agent running on a home server
• Keeping speech processing on the phone instead of a cloud service
• Watching what Hermes is doing through the visual presence during a task

Who it is for: Hermes Agent users who want a voice interface on an Android phone for an agent they host themselves.

How it fits with Hermes Agent

It is a client for Hermes Agent, in which the voice entity is Hermes itself rather than a separate persona.

Requirements: An Android device and a Hermes gateway you run yourself, with its endpoint and API key

Note: The README labels the project beta, an early MVP that is changing quickly, and the Enhanced Realtime mode is alpha and experimental.

FAQ