Does my data ever get sent anywhere?

Not when you use the local model — that runs entirely on the box, from weights on its own drive, with no network call involved. If you choose to send a task to Claude, GPT, Gemini or OpenRouter, it goes to that provider on your key and under their terms, exactly as it would from your laptop. Nothing routes through us either way.

Do you collect any telemetry or analytics from the box?

None. There is no analytics to switch off because there is none to begin with. The box has no reason to report on you and it does not.

Where are my API keys stored?

On the box, on your network. They are not synced to us, not held in a hosted dashboard, and not readable by anyone who has not got the machine.

Private AI, for the question you would not paste into a text box

There is a category of work that never reaches an AI tool at all — not because the tool would handle it badly, but because the document is not yours to upload. Private AI means the model runs where the document already is.

The quiet cost of self-censoring the prompt

Most people who handle confidential material have already worked out the compromise: rewrite the real question into a vague one, get a vague answer back, and do the half that mattered by hand. It does not feel like a loss because nothing visibly breaks. It is still the most expensive thing in the workflow.

Contracts under negotiation. Patient notes. Unreleased code. An acquisition model. A disciplinary file. The work does not get worse because the model is bad — it gets worse because the good model was never shown the real input.

What 'it does not leave' actually means here

The model runs on the box's own silicon, from weights already on its 512 GB drive. Answering a question about a document involves no API call, no key and no packet leaving your network. You can pull the network cable out and ask it anyway.

This is a stronger claim than a policy promise, and deliberately so. A hosted service that says it does not train on your data is asking you to trust a contract; hardware in your own room is asking you to trust a network diagram. One of those you can verify yourself.

The keys, and where they sit

The other half of the privacy question is the one people forget: when you sign into a hosted agent service, you hand it your API keys. They live in their dashboard, on their servers, readable by whatever their access model permits.

Here they sit on the box, on your network. They are not synced to us, not held anywhere we can reach, and not readable by anyone who has not got the machine. Revoking our access means unplugging a cable, because there is nothing else to revoke — there is no telemetry, no analytics and no account.

And the honest limit

The local model is a 7 to 9 billion parameter class model, not a frontier one. For the hardest reasoning you will still want Claude or GPT, and for that you will still be sending something to a provider — on your own key, but off your network all the same.

So the split is yours to make, per task. The point is not that nothing ever leaves. It is that the things which must not leave finally have somewhere to be asked.

Common questions

Does my data ever get sent anywhere?: Not when you use the local model — that runs entirely on the box, from weights on its own drive, with no network call involved. If you choose to send a task to Claude, GPT, Gemini or OpenRouter, it goes to that provider on your key and under their terms, exactly as it would from your laptop. Nothing routes through us either way.
Do you collect any telemetry or analytics from the box?: None. There is no analytics to switch off because there is none to begin with. The box has no reason to report on you and it does not.
Where are my API keys stored?: On the box, on your network. They are not synced to us, not held in a hosted dashboard, and not readable by anyone who has not got the machine.