OpenAssistant

A fully private, offline voice assistant that runs entirely on-device — no cloud, no telemetry, no network required.

Features

Offline voice recognition — Speech-to-text powered by Vosk; all audio processing happens locally and is never transmitted.
On-device LLM — Responses are generated by LiteRT-LM running fully on the device, with no external API calls.
Vision and camera — Assistant interprets the environment through the device camera.
MCP tools — Model Context Protocol support for extending assistant capabilities with local tool integrations.
Routines — Schedule and automate multi-step actions that trigger on voice commands or time-based events.
Skills — A modular skill system that lets the assistant handle specialised tasks.
Default system assistant — Can be set as the default Android digital assistant, replacing Google Assistant or Gemini, and invoked via the system assist gesture or button.
Privacy-first design — No data ever leaves the device. There are no accounts, no analytics, and no remote services.

You can contribute to the project with pull requests, ideas and bug reports.

Licensed under the Apache License, Version 2.0. See LICENSE for the full text.