Why we are building this
Because the most useful software of this decade arrived with a condition attached, and a lot of people quietly decided they could not accept it.
The condition is simple enough that it is easy to miss. To get help with a document, you send the document. For a shopping list that is nothing. For a client file, a case, a set of medical records or a source’s notes, it is a decision with consequences - and for some professions it is a decision somebody else has already made for you, in a code of conduct or an engagement letter.
The result is a strange kind of exclusion. The people with the most document-heavy work, who would gain the most from an assistant that reads carefully and never gets bored, are the people least able to use one. They do without, or they use it on the safe half of their work and feel uneasy about the rest.
Local models have quietly become good enough to change that. Not good enough to beat the largest cloud models - that is not the claim - but good enough to read your folder, answer the question, and show you where the answer came from, on a laptop you already own. AI Stays Local is an attempt to turn that into software an ordinary professional can install, rather than a weekend of configuration.
Who is behind it
An independent developer building a commercial product. Not a venture-funded team, not an open-source project asking for contributions, and not a side experiment of anything else. That is worth stating plainly because it is part of what you are evaluating when you decide whether to put your documents near a new piece of software: a small operation can be more careful with your data than a large one, and it can also disappear. You should weigh both.
What we can offer against that is a habit of saying what is true. Every capability on this site is labelled with whether it exists, and there is a register behind the site recording each public claim, the evidence for it, and the wording we have forbidden ourselves. The feature list is short on purpose, and the privacy page spends as much space on the exception as on the rule.
Where it stands today
Prelaunch. The desktop application is in development and the private beta has not opened yet. Local inference works in an internal prototype on Apple Silicon. Document search, memory and hardware-aware model selection are what we are building towards the private beta. Everything else on the feature list is planned, which is an honest word for not built.
The product is proprietary. It is designed to work with more than one inference engine, including open-source engines used under their own licences and credited in the application’s notices. We do not present other people’s work as ours.