AI Stays Local is a desktop application that puts a capable AI model on your own computer and lets it read the folders you choose. You get the kind of help you would expect from a chat assistant - drafting, summarising, finding the thing you half-remember - without your documents becoming someone else's training data or someone else's breach.
Prelaunch. The desktop application is in development and the private beta has not opened yet.
You
What notice period do I have to give Kaplan & Ruiz, and does it differ from the Nordwind agreement?
AI Stays Local
Thirty days for Kaplan & Ruiz, sixty for Nordwind. Both are counted from written notice, and the Nordwind clause requires it in writing to the registered address rather than by email.
Neither agreement allows termination for convenience inside the first year.
The pattern is so normal that it stopped registering as a decision. You want help with a contract, a set of case notes, a client’s financial records - so you attach the file, and it goes to infrastructure you do not own, under terms you did not negotiate, into logs you cannot see.
For most people, most of the time, that is a fine trade. For a document you are professionally obliged to protect, it is a trade you may not actually be allowed to make. And the alternative has usually been to do without the help.
The model can come to the document instead.
Four steps, and the fourth is the one most tools leave out.
A normal desktop application. It checks what your computer can handle and suggests a model that will actually run well on it.
Grant the folders you want it to know about. Nothing outside them is read, and you can withdraw a folder at any time.
Ask in plain language. Answers that come from your documents name the file and passage they came from, so you can check them.
A status indicator shows where the work is happening. If something would be better handled by a cloud model, you are asked first - and you can say no.
Only the first of these is what we are building for the private beta. The other two are designed and named here so you can see where the product is going - and so you can see, from the drawing, which one crosses the line.
Everything runs on your device.
The model, the document index and your memory all sit on your computer. In local-only mode the application is designed to work with no network connection at all, so there is nowhere for a document to go.
What this does not doAnswer quality and speed are bounded by your hardware. A laptop will not match the largest cloud models.
Your computer serves your own approved devices.
One machine - a desktop at home, a server in your office - does the work, and the devices you have approved can reach it over your own network. Your documents still never leave equipment you control.
What this does not doYou are responsible for the network it runs on. This mode is designed, not built, and is not part of the first private beta.
Optional, one request at a time, only when you approve it.
When a task genuinely needs a larger model, the request passes the Privacy Gateway first: it is classified, redacted where redaction is possible, and shown to you before anything is sent. If you decline, nothing is sent.
What this does not doThis mode does involve sending data to a third-party provider. What you approve does leave your computer, and that provider’s own terms then apply to it.
Prototype: running in internal development builds. not distributed, not stable, and subject to change or removal. Private beta: what we intend to put in the hands of the first private beta group. scheduled, in active development, not yet released. Planned: designed and specified, not yet built. later than the private beta, and the shape may change.
A familiar chat window answered by a model running on your own computer.
PrototypePoint it at folders you choose and ask questions about what is in them, with answers that cite the file they came from.
Private betaEvery document-based answer names the file and the passage it used, so you can check it.
Private betaNothing is read until you grant a folder. Withdraw a folder and its index is removed.
Private betaPreferences and recurring context stored on your machine, visible and deletable by you.
Private betaThe app looks at your computer and suggests a model it can actually run well, described in plain language rather than parameter counts.
Private betaAn adapter layer so the app is not tied to one inference engine and can adopt better ones as they appear.
PrototypeDrive the same local models and document index from a terminal.
Private betaSpeak to the assistant and hear answers back, processed locally where the hardware allows.
PlannedThe checkpoint that classifies a request before any cloud call, redacts what it can, and asks you to approve what is left.
PlannedSend a single request to a larger cloud model when you decide it is worth it, and only then.
PlannedLet your own approved devices use the model running on your computer, over your own network.
PlannedReach your own machine securely when you are away from it.
PlannedA local HTTP endpoint so your own scripts and tools can use the same models.
PlannedAsk for answers in a fixed shape so they can be consumed by other software.
PlannedMulti-step tasks that pause for your approval at each step that touches a file or the network.
PlannedUpdates that are cryptographically signed, verified before install, and reversible.
PlannedCentral policy, device management and reporting for teams.
PlannedTurning a folder off removes its index from this machine.
You choose folders. AI Stays Local builds an index of what is in them so that a question like “which of these mention the Delacroix contract” can be answered by meaning rather than by exact wording. That index is a file on your disk, like any other file on your disk.
Answers drawn from your documents name the file and the passage behind them, because an assistant that cannot show its source is asking you to take its word for something you are professionally responsible for. Remove a folder and its index goes with it.
Point it at folders you choose and ask questions about what is in them, with answers that cite the file they came from.
Private betaEvery document-based answer names the file and the passage it used, so you can check it.
Private betaNothing is read until you grant a folder. Withdraw a folder and its index is removed.
Private betaPreferences and recurring context stored on your machine, visible and deletable by you.
Private betaThe desktop application comes first, because that is where the processor, the memory and the documents already are. Beyond it, the plan is not to scatter your data across more devices but to let the devices you already trust reach the one machine that holds it.
The application itself. First platform, in development now.
Private betaReach the machine that holds your documents over your own network.
PlannedA private connection back to your own machine, rather than a second copy of your data.
PlannedA client sends a forty-page agreement under an engagement letter that does not contemplate third-party processing.
Ask what the termination and liability provisions actually say, and get an answer that points at the clause it read, without the document leaving the machine it arrived on.
A local model summarising a clause is a reading aid, not advice. The citation exists so you check the source, and you still have to.
Two hundred PDFs from a records request, and the question is which of them mention a particular contract number.
Search across the whole folder by meaning rather than exact string, and follow each answer back to the page it came from.
Recall on a large corpus depends on how the documents are structured. Treat it as a strong first pass, not an exhaustive one.
Notes and recordings that identify someone who spoke on condition that their name stays out of any system you do not control.
Work with the material in local-only mode, where the application is designed to function with no network connection at all.
Local-only mode addresses one risk among several. Device security, backups and physical access remain yours to manage, and this product does not solve them.
More situations, including company knowledge and offline work.
The same models and the same index, reachable from a shell and from your own scripts - on the machine that already holds the data, so there is no key to leak and no request to log.
This request would leave your computer
You asked for a longer draft than the local model handles well. Sending it means the text below goes to the cloud provider you selected, under their terms.
Some tasks are genuinely better on a large cloud model, and pretending otherwise would be a worse kind of dishonesty than the one we are trying to fix. So the plan is not to forbid it. It is to make the moment visible: this request, this text, this provider, your decision.
This mode does involve sending data to a third-party provider. What you approve does leave your computer, and that provider’s own terms then apply to it. That sentence is on this page deliberately. A privacy product that will not say what its exceptions cost you is selling something else.
These editions describe the shape we are building towards, not a price list. Nothing here is a commitment, and the private beta is free for everyone selected.
Anyone who wants to try local AI on their own machine.
No charge
People who want their documents involved.
Anticipated: a low monthly figure
Professionals and power users who live in the product.
Anticipated: a professional monthly figure
Small firms where several people handle the same confidential files.
Per seat, not yet set
Organisations with their own security and procurement requirements.
By agreement
We would rather have fifty testers who tell us the truth about their hardware and their work than five thousand sign-ups we cannot learn from. The application form asks real questions because the answers decide what we build first.