A French-speaking Jarvis that runs on your own PC — and thinks for itself
A personal voice assistant that keeps the LLM, the mic, and the keys on your machine, with a local deliberative agent doing the deep thinking.

What it does
Say “Hey Jarvis” and a French-speaking assistant wakes up on your Windows PC: local wake word (openWakeWord), local transcription (faster-whisper), an LLM of your choice, and spoken replies. It then reaches into a genuinely large toolbox — Philips Hue lights, OBS streaming control, PC apps and GPU stats, Google Calendar, Gmail, Discord, your real Chrome via CDP, restaurant bookings through Playwright, and even Twilio phone calls. It runs in three modes: hybrid (cheap cloud model + local background agent), quality (the provider’s biggest model), or fully offline with Ollama and local voices.
The interesting bit
The security doctrine is the standout: deep reasoning is delegated to Hermes, a local deliberative agent that only thinks — Jarvis holds the keys and executes every action, and no credentials ever live in Hermes’s environment. Hermes reads safe tools via Jarvis’s MCP server and can only write drafts. That’s a cleaner trust boundary than most assistant architectures bother with.
Key highlights
- Fully offline mode: Ollama + Piper/Kokoro voices, no API, no usage cost; a 6 GB GPU comfortably runs Whisper medium plus a 4B model
- Network audio satellites put mic and speaker in other rooms (Raspberry Pi/Linux today; Android and ESP32 are planned targets)
- MCP server exposes home/PC tools to any MCP client, including Claude Desktop
- Graduated security levels (N1/N2/N3) per tool, revocable “always allow”, and per-provider LLM budgets with automatic fallback to local
- Extras that show the scope: gesture control via webcam (100% local, no frames leave the machine), an A4 morning briefing PDF, music recognition, and an iPhone Shortcuts bridge
Caveats
- Personal project shared as-is, targeting Windows 11 specifically — expect friction elsewhere
- Some integrations lean on unofficial APIs (Alexa/Echo) or are flagged experimental (Google Home/Nest, Pi Zero satellites)
- The README is honest that vision-heavy features (browser, bookings) still work best with a cloud model; a 7B local model handles domotics fine but not everything
Verdict
Worth a look if you want a privacy-first, French-language assistant that actually controls your house and PC — and the Jarvis/Hermes key-separation design is instructive even if you never install it. Skip it if you want a polished cross-platform product; this is one person’s very elaborate rig, documented generously.
Frequently asked
- What is sosoj92/jarvis-assistant-vocal?
- A personal voice assistant that keeps the LLM, the mic, and the keys on your machine, with a local deliberative agent doing the deep thinking.
- Is jarvis-assistant-vocal open source?
- Yes — sosoj92/jarvis-assistant-vocal is open source, released under the MIT license.
- What language is jarvis-assistant-vocal written in?
- sosoj92/jarvis-assistant-vocal is primarily written in Python.
- How popular is jarvis-assistant-vocal?
- sosoj92/jarvis-assistant-vocal has 505 stars on GitHub.
- Where can I find jarvis-assistant-vocal?
- sosoj92/jarvis-assistant-vocal is on GitHub at https://github.com/sosoj92/jarvis-assistant-vocal.