Maylo Voice Assistant

Offline-first voice assistant stack for macOS (Wake word + VAD recording + local Whisper ASR + OpenClaw agent response + offline TTS via macOS say). Use to i...

Security Scan

The code mostly matches an offline macOS voice assistant, but the package metadata understates real requirements (notably the 'openclaw' CLI, macOS 'say', and model downloads), and the web UI/audio streaming exposes a network surface you should consider before installing.

Purpose & Capability

The skill description (offline macOS assistant) aligns with the code: wake-word (openWakeWord), VAD, local Whisper ASR (faster-whisper), an OpenClaw responder bridge, macOS 'say' TTS, and a web UI. However the registry lists no required binaries or credentials, while the code clearly expects/uses external binaries and services: the 'openclaw' CLI is invoked from bridge/milo_responder_openclaw.py, 'say' is used for TTS, openssl is used to generate certs in run_web_https.sh, and uvicorn is launched for the web UI. The metadata omission (no required binaries) is an incoherence and reduces transparency about prerequisites.

Instruction Scope

SKILL.md instructs copying the app, creating a venv, installing requirements, and running scripts — which is consistent with the code. But there are notable scope issues: (1) Despite 'offline-first' wording, openwakeword.utils.download_models() (in jarvis_minimal_test.py) and model loading (WhisperModel) may download pretrained models or require network access during setup or first run; pip install also downloads packages. (2) The web UI serves audio over WebSocket and the run_web_https.sh defaults to binding 0.0.0.0 with a self-signed cert, exposing the audio upload endpoint to the LAN; that is expected for the UI but is a privacy/network exposure you should explicitly consider. The runtime instructions do not call out these network interactions clearly.

Install Mechanism

There is no centralized install spec in registry (instruction-only), but the included scripts/install.sh will create a venv and pip install requirements.txt. This is normal but requires network access to PyPI and will pull packages (openwakeword, faster-whisper, onnxruntime, etc.). The script also optionally installs switchaudio-osx via Homebrew. Nothing is fetched from untrusted arbitrary URLs in the package itself, but third-party Python packages and model downloads are required.

Credentials

The skill does not request credentials or secrets in metadata. Code reads a few configuration env vars (MAYLO_* such as MAYLO_OPENCLAW_AGENT_ID, MAYLO_INPUT_DEVICE, MAYLO_WEB_WAKE) which are reasonable for runtime configuration. There are no declared or implicit demands for unrelated cloud credentials or secrets. The SKILL.md warns not to commit secrets.

Persistence & Privilege

The skill is not force-included (always:false) and uses standard background processes/threads. It launches background daemons (bridge/responder and main assistant) and a web server that by default binds 0.0.0.0 (network-exposed). While this is coherent with a web UI, it increases the blast radius on a LAN — consider restricting to localhost or firewalling the port if you want to limit exposure.

What to consider before installing

What to check before installing: - Expect to install Python packages and possibly download pretrained models; the 'offline-first' claim is accurate at runtime but not during setup — model and pip downloads may occur. Plan for network access during install. - The code invokes the 'openclaw' CLI locally (bridge/milo_responder_openclaw.py). Ensure you intentionally install and trust the openclaw binary on the same machine; the skill will run 'openclaw agent --local' via subprocess. - This is macOS-focused: it calls the 'say' command and uses macOS audio behavior. It will generate a self-signed cert and default to binding the web UI to 0.0.0.0:8443 (accessible on your LAN). If you do not want LAN access, run the web UI on localhost or use a firewall. - The package metadata omits required binaries (openclaw, openssl/uvicorn/runtime expectations). Treat that as a transparency issue: verify prerequisites before running. - If privacy is a concern: inspect recordings/ directory (logs and possible saved WAVs), and consider running the project in an isolated environment (VM or sandbox) and reviewing the bridge/responder logs to confirm all processing is local. - If you want higher confidence: review/verify openclaw installation/source, audit the exact versions of pip packages installed, and search for any unexpected outbound network connections during install and first run.

Maylo Voice Assistant (macOS)

What this skill ships

• A self-contained app under assets/app/:
  • Wake word listener (openWakeWord)
  • VAD recording (webrtcvad)
  • Local ASR (faster-whisper)
  • Responder bridge via OpenClaw (openclaw agent --local)
  • Offline TTS via say -v Yelda
  • Web UI (HTTPS + WebSocket audio streaming)

Install / Setup (fresh machine)

1. Copy the app to a working directory (recommended):
   • cp -R <skill>/assets/app ~/maylo-voice-assistant
2. Create venv + install deps:
   • Run: scripts/install.sh ~/maylo-voice-assistant
3. Start the assistant (wake word + responder):
   • Run: scripts/run_assistant.sh ~/maylo-voice-assistant
4. Start the HTTPS web UI for phone mic streaming:
   • Run: scripts/run_web_https.sh ~/maylo-voice-assistant --host 0.0.0.0 --port 8443
   • On iPhone/Android (same Wi‑Fi): https://<mac-ip>:8443
   • Accept the self-signed certificate warning.

Normal use

• Wake word: say "Hey Jarvis" near the Mac mini, then speak your query.
• Web UI: hold the mic button to talk; release to send.

Troubleshooting (fast)

• Check logs:
  • recordings/assistant.log
  • recordings/responder.log
  • recordings/web.log
• If wake word never triggers, run the minimal tester:
  • python jarvis_minimal_test.py
• If the assistant responds to itself (feedback loop):
  • Set HDMI output back to Mac speakers / use headphones.
  • Increase MAYLO_POST_SAY_INHIBIT_SEC.

Security / Privacy

• Do NOT package or commit:
  • OAuth client secrets, tokens, refresh tokens
  • Private certs/keys
• This skill ships no secrets. Any tokens/keys must be created on the target machine.