This Hermes Agent Project Gives Your AI a Live2D Desktop Pet That Speaks Its Voice

Soundpulse/hermes-live2d is an open-source project that gives your Hermes Agent a physical face: a transparent ATRI Live2D desktop pet that plays the agent's text-to-speech output with lip sync, expressions, and a speech bubble. Built on Tauri 2 and pixi-live2d-display, it routes audio the agent already generates to a pet running on your Mac over a private Tailscale network.

What it does

The repo's pitch is simple: "a transparent Live2D desktop pet (Tauri 2 + pixi-live2d-display) with voice lip sync, expressions, mouse tracking, and a local HTTP API." It currently ships the Sparkle (火花) model. The intended use is that your Hermes Agent runs somewhere remote — say an EC2 box chatting over the Discord gateway — and the pet runs on your Mac. Whenever the agent generates a TTS recording with its text_to_speech tool, a small skill uploads that audio to the pet, which plays it lip-synced with an expression and a speech bubble.

Crucially, normal chat replies are not spoken — only the audio recordings the agent deliberately creates get routed. So there's no risk of the pet blurting out every message; it just animates the voice clips the agent already produces.

How the plumbing works

The flow is a neat example of local-first agent architecture:

hermes agent (remote)                         your Mac
────────────────────                          ────────
text_to_speech ──► hermes-live2d-voice skill ──► POST /speak ──► Sparkle pet
                     (speak.sh, base64)          (over Tailscale)   lip sync + bubble

Requirements straight from the README:

• Tailscale on both the Mac and the agent host (same tailnet) — this is how the remote agent reaches the pet without exposing it publicly.
• A Hermes Agent wired to a TTS engine — any engine works (ElevenLabs, etc.), as long as the agent produces audio files via its text_to_speech tool.
• The hermes-live2d-voice skill — bundled in the repo.

The project also ships a Claude Code skill, which tells you the author is thinking about more than just Hermes — but the headline integration is the Hermes text_to_speech → pet pipeline.

Why it's a fun signal

Hermes Agent has always leaned open-source and self-hosted, and this is the ecosystem doing what it does best: taking an autonomous agent and giving it a presence on your desk. It sits right alongside the deep dive on Hermes Agent as an autonomous coding assistant and the agent's local-first, privacy-minded architecture — the pet only ever sees audio over a private tailnet and the HTTP API listens on 127.0.0.1 with no auth.

The author lists honest limitations: swapping in a different Live2D model requires code changes, lip sync is volume-driven rather than viseme-accurate, and mouse tracking via Tauri can vary across macOS versions. The bundled Sparkle model carries its own license (from booth.pm), separate from the project's MIT code license.

Who it's for

• Hermes Agent users who want a visible, charming indicator that their remote agent is "talking."
• Cron and health-check alerts — the author suggests it's useful for cron reminders and health checks, since those often fire TTS.
• Anyone who likes ricing their setup — a desktop pet that literally speaks your agent's voice is about as personal as it gets.

It's early (the repo was created within the last week and has only a handful of stars), but it's a small, MIT-licensed, self-hosted example of the "agent as a presence" idea rather than "agent as a chatbot in a tab."

FAQ

Q1: Does the pet read out every chat reply?
No. Only the audio recordings your Hermes Agent generates through its text_to_speech tool get routed to the pet. Standard chat text replies are not spoken.

Q2: Is it safe to expose the pet to the internet?
The README advises against it. The local HTTP API listens on 127.0.0.1 with no auth; you're meant to reach it only over a private Tailscale tailnet, never the public internet.

Q3: Do I have to use the Sparkle model?
The pet currently ships with the Sparkle (火花) model, and swapping in a different Live2D model requires code changes. The bundled model also has its own separate license terms from the project's MIT code license.

You already use AI agents for coding. Why limit yourself to one model? aiFiesta ($12/mo) gives you every premium model in one dashboard — Claude, GPT, Gemini, Grok and more.