SIN

Personal AI OS


A local intelligence layer for your computer.

SIN runs a language model on your own machine, remembers what you tell it, keeps your tasks and reminders, and stops for your approval before it does anything to your computer. No account. No cloud. Nothing leaves the window unless you send it.

Download for Windows What it does 0.8.1 · 2026-08-25

Most assistants are a website with a login. This one is a program on your computer that happens to think.


01
What SIN actually is

A desktop application, a local model, and a hard line about what it may touch.

It runs on your machine

SIN talks to a model served locally by Ollama, on your hardware. SIN can start that runtime for you without opening a terminal. The address it talks to is locked to your own computer and cannot be pointed elsewhere from the interface.

It stops before it acts

Anything that touches your files, your screen or your programs is proposed and waits. Refusals are stated in words, not error codes. Every action is checked after it runs, not just before.

It remembers

Conversations survive moving around the app. Tasks, reminders and notes persist between sessions in a database on your disk. Nothing is uploaded to keep it.

It has a team

Agents you create get a generated name and a personality drawn from twelve bounded traits. Personality shapes how they write. It never grants them permission to do anything — that boundary is structural.


02
How fast it is, measured

We publish the numbers rather than the adjective.

Measured on a desktop with a 14-billion-parameter model, using the runtime's own timing counters. Your hardware will differ; that is the point of saying which hardware.

Quick

2.0 seconds

Short answers with deliberation switched off. Five times faster than the default, and the reply is about as long.

Normal

10.3 seconds

The model's own default: it thinks first, then answers.

Deep

15.3 seconds

Deliberation encouraged and the token budget doubled. Roughly 55% more thinking than Normal.

Cold start

~23 seconds

Once, on a machine that just booted: 5s to start the runtime, 7.5s to load the model, then the first answer. Every turn after that is warm.

What those numbers do not mean: they are how long the whole answer takes. You do not wait that long to see anything — SIN streams the reply as the model writes it, so the first words arrive in well under half a second and you can press Stop at any point. The figures above are completion times, not silence.

03
Where your data goes

Three connections, and you can name all of them.

Your own computer

Every conversation goes to a model on 127.0.0.1. That address is enforced in code and cannot be changed to a remote host from the interface.

This website, for updates

SIN asks whether a newer version exists, at most every six hours. It sends its version number and nothing else — no identifier, no account, no usage data.

Whatever you ask for

Web search and model downloads reach the internet because that is what you asked them to do. Nothing else does.

The full privacy page →


SIN

Version 0.8.1, released 2026-08-25. Windows 10 or 11, 64-bit.

Download · 5.8 MB Release notes