Questions and answers

AI Stays Local is a desktop application that gives you a chat assistant and document search powered by an AI model running on your own computer. It is designed so that the documents you ask about stay on your device unless you specifically choose and approve an operation that uses the cloud.

No. The desktop application is in development and the private beta has not opened yet. The application form on this site is how you ask to be in the first group of testers.

Those services run on the provider’s infrastructure: your prompt and any file you attach are sent to them. Here the model runs on your machine, so a local answer is produced without sending your document anywhere. The trade-off is real - a model that fits on a laptop is smaller than one running in a data centre, and it will not match it on the hardest tasks.

You grant specific folders. Their contents are indexed on your computer so that questions can be answered against them. The index stays on your machine, and removing a folder removes its index. Nothing is uploaded as part of normal local use.

More private than a hosted assistant, and not perfectly so - no software should tell you it is. What we can say is specific: in local-only mode the application is designed to run with no network connection, so there is nowhere for a document to go. If you turn on Controlled Cloud and approve a request, the content in that request does leave your computer and the provider you route it to handles it under their own terms. The privacy page sets out each mode and what it means.

The product is being built so that analytics and crash reporting are opt-in rather than on by default, and so that the application works with outbound network access blocked. This is a design commitment during development, not an audited result, and we will publish what the shipped build actually does before the beta opens.

Hardware compatibility varies. Memory is usually the limiting factor: more memory allows a larger, more capable model. A recent laptop with 16 GB of memory should run a useful small model; 32 GB or more opens up better ones. The compatibility page explains the trade-offs without promising a result for a machine we have not tested.

Desktop first. The prototype has been developed on Apple Silicon, and Windows and Linux builds are planned. Mobile access is planned as a private connection to your own computer rather than as a separate assistant on the phone.

You need one to download the application and a model. After that, local-only mode is designed to work without a connection. Controlled Cloud obviously requires one, and only runs when you approve a specific request.

A small first group, chosen for a spread of hardware and for real work we can learn from - people who handle confidential documents as part of their job, and people willing to tell us when something is wrong. We will contact applicants directly; there is no queue position to check.

We use it to select testers and to understand what hardware to support. We do not ask for documents or any confidential material, and the form should never be used to send any. The website privacy notice sets out retention and how to ask for deletion.

There will be a free edition that runs a local model with no per-message charge, and paid editions for document search, larger models and professional use. Figures are not final and the pricing page says so wherever it shows one. The private beta is free for everyone selected.

The product is commercial and proprietary. It is built to work with more than one inference engine, including open-source engines used under their own licences, and those licences and attributions are included in the application’s notices. We do not present other people’s work as ours.

An independent developer building a commercial product, not a venture-funded team and not a side project of anything else. The about page is straightforward about the size of the operation, because that is part of what you are evaluating.