dwr-git/dewiar-app
Голосовая помощница с командой ИИ-персонажей для Windows, Ubuntu и Android — делает дела голосом · Voice AI assistant that actually does things
What it solves
Dewiar is a voice assistant that understands natural speech, not just commands. You can call it by name, interrupt it, and it responds. It solves the problem of having a single assistant with limited capabilities by offering a team of AI characters that can handle different tasks. It also lets you choose where the thinking happens: on your own computer with your own model, or in the cloud using services like OpenAI or OpenRouter, so you keep control over your data.
How it works
The assistant uses speech recognition (Vosk, GigaAM) and text-to-speech (Piper) to listen and talk. It runs a core engine (in Python) with a widget and studio UI on desktop, and an Android app. It connects to any LLM backend: your local vLLM, Ollama, llama.cpp, LM Studio, or cloud APIs. It has a set of characters that can be activated by name; you can create your own with custom name, voice, and photo. It can handle background tasks: you give it a task like "fix the failing tests", and it plans, logs each step, asks questions, and works independently (up to 1200 steps). It has memory with versioning, can answer from your documents (PDF, Word, Markdown, Obsidian) using offline search. It syncs across devices through a dewiar.pro account, and you can also talk to it via Telegram and MAX. Risky actions require your confirmation.
Who it's for
Dewiar is for people who want a hands-free, voice-controlled assistant that respects privacy and can run locally without cloud dependency. It appeals to users who like having multiple personalities and the ability to run their own models. The current interface is in Russian, so it's primarily for Russian-speaking users. It's available for Windows, Ubuntu, Android, and in Safari on iPhone/iPad/Mac (browser version). It's free, so anyone can try it.
Highlights
- Free and ad-free, no subscriptions.
- Runs fully offline with your own model (Ollama, LM Studio) on a computer with a GPU.
- Multi-platform: Windows, Ubuntu, Android, and browser for iPhone/iPad/Mac.
- A team of characters (Lira, Patrick, and more) that can be added or customized.
- Background tasks with planning and logging, up to 1200 steps.
- Memory: remembers what you told it and how it changed over time.
- Can control smart home, set reminders, read documents, and more.
- End-to-end encrypted communication between assistants of different users (X25519 + HMAC-SHA256).
Related
- Project
- Project
- Project
- Project