omnilimb-face 0.1.4

Open-LLM-VTuber capabilities (hands-free voice, barge-in, Live2D avatar) as a hermes-agent plugin.

omnilimb-face

English · 中文  |  🌐 face.omnilimb.com

A standalone, installable hermes-agent plugin that gives your agent a face and a voice: hands-free voice interaction (VAD + STT), real-time barge-in, and a Live2D / Live3D avatar with lip-sync and expression driving — all without modifying any hermes core file. Part of the omnilimb family.

Demo

The avatar renders on top and speaks the agent's reply with lip-sync and expressions; you type (or talk) in the dialog box below.

The animation above is a synthetic preview (omnilimb-face); a still frame is in docs/assets/screenshot.png. Avatar: Live2D Cubism sample "Mao" © Live2D Inc. (loaded from CDN, not bundled — see Credits).

📖 Full docs: docs/GUIDE.md — overview / architecture / install / config / commands (CLI · slash · tools) / Live2D ↔ Live3D switching / real-time barge-in / troubleshooting / development.

🎭 Avatar deep-dive: docs/AVATAR_INTEGRATION.md — how Live2D/Live3D expressions, motions and lip-sync are bound, how to import your own model (Cubism / VRM), and the 2D + 3D fusion roadmap.

How it works

The plugin reuses hermes' existing systems instead of carrying its own model config: transcripts are injected via ctx.inject_message, the host agent's normal turn produces the reply (using the user's active model, tools and memory), and the reply text is intercepted through the transform_llm_output / post_llm_call hooks to drive TTS and the avatar. Speech transcription reuses the stt config section; speech synthesis reuses the tts section. The plugin never calls an LLM itself — the avatar always speaks your configured agent's real answer.

Getting started

Requires Python 3.11+. Pick one of the two paths below — each is step by step.

Option A — 1-minute preview (no hermes-agent needed)

# 1) install it (the avatar renders, TALKS via Edge-TTS, and takes typed input)
pip install omnilimb-face
# 2) start it — ONE command serves the web page + gateway and opens your browser
omnilimb-face
# 3) (if the browser didn't open) go to:
#       http://127.0.0.1:12394/      <-- the web page

Type in the page and the avatar replies in Chinese voice with lip-sync + expressions. (Click or type once first — browsers only allow audio after a user gesture.)

⚠️ The page is on port 12394, not 12393. Port 12393 is the WebSocket gateway (ws://…/client-ws); opening http://127.0.0.1:12393/ in a browser will NOT work. The page lives on gateway port + 1 = 12394. (In single-port mode — omnilimb-face --single-port / --https — the page and gateway share 12393.)

Option B — full product (chat in hermes, the avatar speaks the real replies)

# 1) install into the hermes venv (one command, everything included)
<hermes-venv>/python -m pip install omnilimb-face
# 2) enable it: add `omnilimb-face` to plugins.enabled in ~/.hermes/config.yaml
# 3) run hermes — it starts the gateway (12393) + front-end (12394)
hermes
hermes vtuber status                 # check the avatar subsystem came up
# 4) open http://127.0.0.1:12394/ , then chat in the hermes terminal —
#    the avatar speaks each reply with lip-sync + expressions.

Speech uses the host text_to_speech tool when present, else the keyless Edge-TTS fallback (Chinese voice by default; set tts.<provider>.voice to an Edge *Neural voice to change it). The /client-ws gateway works with websockets 12.x and 13–15.x.

Optional extras

pip install omnilimb-face already renders the avatar, talks (Edge-TTS) and takes typed input. The extras below add the full hands-free voice loop and other opt-in pieces:

pip install "omnilimb-face[all]"        # full voice loop: mic capture + STT + wake-word
pip install "omnilimb-face[voice]"      # hands-free microphone capture + VAD
pip install "omnilimb-face[stt]"        # local speech-to-text (faster-whisper) for --stt
pip install "omnilimb-face[wakeword]"   # wake-word activation
pip install "omnilimb-face[dev]"        # test tooling

Enabling in hermes

Discovered via the hermes_agent.plugins pip entry point, or by placing the directory at ~/AppData/Local/hermes/plugins/omnilimb-face/. Enable it by adding omnilimb-face to plugins.enabled in config.yaml.

Mobile (phone) support

The avatar and hands-free voice also run on a phone. Because browsers only allow microphone access from a secure context, the preview can serve over HTTPS with a self-signed certificate on your LAN:

omnilimb-face --lan --https              # HTTPS on your LAN IP (single port 12393)
# or use start.bat options 3 / 4 (LAN HTTPS, optionally with --llm --stt)

Then open https://<your-LAN-IP>:12393/ on the phone (same Wi-Fi) and accept the self-signed certificate warning once — after that the mobile mic works. Cert generation needs the [preview] extra (cryptography).

Troubleshooting

• http://127.0.0.1:12393/ won't open. That's the WebSocket gateway, not a web page. Open http://127.0.0.1:12394/ (gateway port + 1).
• No sound. The Edge-TTS voice ships with pip install omnilimb-face, so this is usually internet/gesture: Edge online voices need internet, and browsers only play audio after you click or type once in the page. The preview prints voice: on (edge-tts …) when the voice engine is active.
• I installed it but nothing happens / no avatar. Installing only adds the code — you still have to START it: omnilimb-face (Option A) or hermes vtuber start (Option B), then open http://127.0.0.1:12394/.
• Avatar doesn't appear. The Live2D model loads from CDN — check your internet. Offline, it falls back to a dependency-free canvas placeholder.

Layout

omnilimb-face/
├── pyproject.toml          # packaging, pinned deps, optional extras, entry point
├── plugin.yaml             # PluginManifest (name/version/hooks/tools)
├── __init__.py             # directory-discovery shim -> re-exports register
├── omnilimb_face/          # plugin package
│   ├── plugin.py           # register(ctx) entry point
│   ├── voice/              # capture, VAD, wake-word
│   └── protocol/           # /client-ws event models + gateway
└── tests/                  # pytest + Hypothesis (unit / property / integration)

Development

pytest                                   # run the test suite
HYPOTHESIS_PROFILE=ci pytest             # heavier property-test run

Releasing (maintainers)

Bump version in pyproject.toml, then publish with the helper script:

py -3.11 -m pip install --upgrade build twine   # one-time
./release.ps1                                   # build + upload to PyPI
./release.ps1 -TestPyPI                         # dry-run against TestPyPI first
./release.ps1 -SkipBuild                        # upload an existing dist/ only

release.ps1 reads the project-scoped PyPI token from a local .env (PYPI_API_TOKEN_FACE) that lives outside the repo and is never committed. The token is used transiently and masked in all output. PyPI rejects re-uploading an existing version, so always bump version first.

License

Licensed under the GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later) — see LICENSE.

In short: you are free to use, study, modify and share this software, including for commercial purposes, but if you distribute it or run a modified version as a network service, you must release your complete corresponding source under the AGPL as well. This keeps every downstream version open.

Commercial / proprietary license available. If you want to use omnilimb-face in a closed-source or proprietary product without the AGPL's source-disclosure obligations, a separate commercial license can be purchased — see COMMERCIAL-LICENSE.md or contact [e-mail].

Copyright © 2025 seanyang1983.

Credits & third-party licensing

This plugin does not bundle or redistribute any avatar models, the Live2D Cubism Core, or any third-party front-end runtime. Everything below is loaded from CDN at runtime only (with a dependency-free canvas fallback when offline). Full details in NOTICE.md and THIRD_PARTY_NOTICES.md.

• Open-LLM-VTuber — the /client-ws protocol here is an independent re-implementation compatible with Open-LLM-VTuber (MIT through v1.2.0 as of 2026-06; see NOTICE.md for the license-transition note). No upstream source is copied. Thanks to the project for the protocol design.

• Live2D Cubism sample model — the default avatar (models/model_dict.json) references Live2D Inc.'s official Cubism sample "Mao" (a commit-pinned CDN URL only). The Cubism Core runtime is proprietary to Live2D Inc. and is loaded from Live2D's CDN, never bundled.

  This content uses sample data owned and copyrighted by Live2D Inc. (the "Terms of Use for Live2D Cubism Sample Data" / Live2D Cubism SDK license — https://www.live2d.com/en/).

  Live2D's official samples are free for general users and small-scale enterprises (latest annual sales under 10,000,000 JPY); larger entities are subject to additional Live2D terms. For commercial / large-scale use, swap in a model you own or are licensed to use by editing models/model_dict.json.

• pixi.js / pixi-live2d-display / three.js / @pixiv/three-vrm — all MIT, loaded from CDN.

• Optional Python deps — edge-tts ([preview], GPL-3.0, never bundled) and openWakeWord ([wakeword], Apache-2.0 library but its pre-trained models are CC-BY-NC-SA-4.0 / NonCommercial) need attention for commercial use.