Any product that sees your screen and answers you out loud has to send something somewhere. There is no way around it: the reasoning does not fit on a laptop today. The useful question is not whether something leaves, but what exactly leaves, where it goes and what stays.
Almost nobody answers that. We answered it wrong twice, and both times we corrected it the same day we noticed. This article is what there is.
01What leaves when you share your screen
When you turn screen sharing on, those images are streamed to the voice engine as video, at around one frame per second. It is not a single grab taken when you ask a question: it is a stream for as long as sharing is on.
Saelyx also reads the text on your screen on your own machine, so the model doesn't have to decipher a blurry image. That is true and it is a real improvement —the text arrives exact instead of approximate— but it is an improvement on top, not a replacement: the video travels all the same.
For a while we said the opposite. The policy said the screen was not streamed as video and that only the extracted text travelled. That was false, and it was false because there had been a version where the text did go on its own: it was reverted because it broke the audio —sending text while the microphone is open tripped the server's silence detection and the agent came out choppy— and the document was left describing a version that no longer existed.
02The camera, and some small print that matters
The camera works the same way: the frames are streamed. And here there is a detail worth saying out loud rather than hiding it in a table.
There is an on-demand capture setting that cuts the sending down to one frame per turn instead of continuous. In Meeting Mode that setting does not apply: the camera is streamed anyway, even if you chose otherwise. It is deliberate —continuous vision is what lets the agent follow what happens in the room— but it means your preference does not protect you there, and that is something to know beforehand, not afterwards.
03What does stay on your machine
Text recognition happens on your computer. That is not a figure of speech: it is the operating system's vision engine working locally, which is why it is just as fast with a three-hundred-row document as with a label.
What that engine does not do is recognise objects, faces or scenes. Only text. Show it something with no text in it and it returns nothing. When Saelyx describes a scene to you, it is not recognising it on your machine: the model is looking at it, using the image that reached it.
That distinction is exactly what made us write the website wrong: we promised recognising objects and scenes without making clear where it happened. The capability exists; the place was the wrong one.
04The voice
The audio of the conversation is processed by an external provider, with retention switched off at our end: it does not store the audio. It is declared by name in the privacy policy, because European regulation requires identifying whoever processes your data, and because a product that asks you for the microphone should not hide who is listening.
And there is an exception we found while reviewing this, and it does not flatter us: on the Apple Watch the voice does not pass through that provider, nor through our servers. The watch connects directly to a different one. The policy has said so since the day we verified it, with the provider's name and a note that this audio does not enter our European infrastructure.
05The memory, which is the part you do control
What Saelyx remembers between conversations lives in a list that you open, read, edit and delete. It is not a screen buried in a third-level submenu, nor a "forget everything" button that leaves you guessing what it forgot: it is a file with entries, and every one of them can be removed.
We think it is the most honest of the three proofs. A privacy promise you cannot check is a slogan; a list you can empty is a fact.
06How to check this without taking our word for it
Three things you can verify yourself, with no help from us:
- The camera indicator. While the camera is on, the green light on your Mac is on. That LED is driven by the hardware and no application can switch it off.
- The preview. When the camera is active you see exactly the same frame that is being sent. There is no second image.
- The memory. Open it and look. What isn't there doesn't persist.
07Why we tell the uncomfortable part
We could have corrected both sentences without saying they had been wrong. Nobody would have found out: they were two lines inside a legal document that almost nobody opens.
We didn't, for two reasons. The first is that a privacy policy that says something false is not a drafting error, it is a problem, and treating it as a typo would be the second lie. The second is more practical: if in a year's time someone watches the network traffic and finds something that doesn't match what we said, the damage won't be the fact —it will be that we said it.
We would rather the uncomfortable part were written by us.
08Frequently asked questions
Is the video of my screen stored?
No. It is streamed while sharing is on and it is not stored on our systems. The only thing that persists is what you choose to keep in the memory, and that you can see as a list.
Can I use Saelyx without the camera?
Yes. The camera is a separate permission: if you don't grant it, you still get voice and screen exactly the same.
Is the text on my screen recognised locally or on a server?
Locally, with the operating system's vision engine. What travels afterwards is that text and the video it came from.
Where does the reasoning happen?
On servers in Europe, with the Apple Watch exception explained above and detailed in the privacy policy.
How do I get Saelyx?
By invitation. You ask on the website and we approve requests by hand, one at a time, so we can look after whoever comes in.