ThorOS is an assistant that actually operates your computer — locally, safely, and out loud. Here's what it can do today.
Say the wake word, then your request. Wake-word detection, speech-to-text (Whisper), the language model (via Ollama), and the spoken voice (Piper) run on your own machine by default — nothing leaves unless you choose to connect a cloud model for more power. Prefer to type? A chat window gives you the same assistant.
Your words go to an on-device planner that decides what to do and routes it to specialized agents: files, applications, documents, the terminal, live research, reminders, memory, and security. Each does one job well — and it's doing real work, not just chatting.
Say "I want to build a small game." ThorOS interviews you until it truly understands the project, proposes a complete plan for your approval — language, tools, an Agent team, test stages — then sets up the workspace and its Agents write runnable code, checked and version-controlled, while you watch every step in a live Mission window. Code it yourself, code together, or let the Agents build while you direct — your choice, asked up front.
ThorOS isn't bound to a single LLM. Each job can run its own model: a fast one routes your commands, a specialist coding model writes your programs, another handles writing and reasoning — running simultaneously on capable hardware, swapping in as needed on smaller machines. Manage it all by voice: "what models do I have," "download the qwen coder model," "use it for coding." New and better models come out every month — your OS just adopts them.
If no skill fits your request, a reasoning backbone still helps to the maximum: it answers the question, tells you the simple way to trigger the right skill, or honestly offers the closest thing it can do — never a flat "I can't, try again."
"What's the price of Bitcoin?" "What's the weather?" — it looks things up on the web and reads back a short answer. Ask it to remember: "remind me an hour before my meeting," or "give me the Bitcoin report every weekday at 9 am," and it speaks up at the right time.
You don't have to pronounce a filename perfectly. Say "the reports folder," "my vacation photos," or "the first one," and it matches what's actually on disk. And anything destructive — delete, rename, move — always reads back the resolved name and waits for your yes or no first.
Open a terminal and say "list files" and it runs ls right there. Focus a document and
"make it bold" formats it. The same words do the right thing depending on the window in front of
you.
Beyond clicking menus, ThorOS speaks LibreOffice's own command language: "replace michael with Michael," "insert a page break," "export this to PDF" — driven through the application's real API, not fragile screen automation.
There's no voice app for everything — but Linux already has a command for almost anything. Ask "use the terminal to convert this to a PDF," and ThorOS works out the command, shows you exactly what it'll do, and runs it inside a sandbox locked to a single folder, so a wrong command can't escape.
Browse a catalog and install AI agents by voice — hear what an agent can do, confirm, and it's usable instantly. Untrusted community agents run inside a sandbox (read-only system, their own folder only, no network); reviewed agents run trusted. Anyone can build and publish one.
Ask "why did you do that?" and it tells you the genuine reasoning behind its last action. Tell it your name, your preferences, the way you work, and it remembers across sessions — in a plain file on your disk that you own and can read.
The AI never touches the system directly. Routine actions just happen; anything risky — deleting files, running commands — is gated behind a confirmation or an authorization code, with a hard block that refuses catastrophic commands like wiping a disk.
ThorOS is voice-first by design — so people with severe arthritis, injuries, or no hand use can operate a computer and the web independently, in real time. Most software is built for a keyboard and mouse. This isn't. That's the whole point.
Everything runs on-device by design — private, and yours. The same codebase also supports an optional cloud brain or a thin-client mode for lighter hardware: opt-in, never required, never a fork. Your AI, your choice.
ThorOS is open and growing. Come help shape it — whether
you write code, build an agent, test it on your hardware, or just want to follow along. We chat on
Matrix, self-hosted on our own server — open, decentralized, no tracking. Join from
any Matrix account, or create one on matrix.yggdrasilai.org (registration code:
thoros-6c92e99f).