About this project

Jarvis is a personal French-language voice assistant designed to run on a Windows 11 machine. It is activated by the phrase "Hey Jarvis," transcribes voice locally with Whisper, reasons via a configurable LLM, and responds aloud. Three modes are offered: hybrid (cost-effective cloud LLM + local delegation), quality (more powerful cloud model), and local (Ollama + Piper/Kokoro/Windows voice, entirely offline). OpenAI and Claude/Anthropic are integrated; other providers can be added via connector. The project stands out for its broad toolkit: Philips Hue home automation, OBS control, launching applications and PC statistics, Google Calendar, Gmail summaries, Discord, Instagram, web bookings via Playwright, Chrome control via CDP, Twilio phone calls, long-term memory, iPhone bridge via Shortcuts, music recognition, response overlay, and a local web panel. An MCP server exposes the home automation/PC tools to clients such as Claude Desktop. The architecture plans to delegate in-depth reasoning to Hermes, a local deliberative agent in a Docker container, with strict separation: Jarvis executes actions and holds the credentials, Hermes reads the Vault and safe read-only tools. Network audio satellites (Raspberry Pi/Linux available, Android/ESP32 planned) make it possible to offload microphone and speaker into multiple rooms. An old PC can serve as a headless server while a main Windows workstation offloads audio, screen, and applications. Security is graduated by levels N1/N2/N3: safe automatic actions, sensitive actions with revocable confirmation, critical actions with systematic confirmation and never remotely. Secrets and personal data are not versioned. The project is under the MIT license, and adding a tool is done with a simple file in tools/ with a decorator.