Lectavi

Last updated November 3, 2026

Privacy policy.

Lectavi listens to what you are listening to, so it handles recordings of people speaking. This page says exactly where that audio goes, what is kept, what leaves your computer, and how to get your data out or deleted. Every sentence here describes what the app does today.

Lectavi never sends audio by itself: audio leaves this computer only while a session you started is running, on a press of Explain that, a question you ask, or a recording you chose to make, and the orb on your screen shows which is happening.

The short version

Who we are

Lectavi is a desktop application published at lectavi.com and operated by its developer. For anything in this policy, including requests about your data, write to support@lectavi.com.

What Lectavi stores on your computer

Everything below lives in Lectavi’s application data folder, on your disk, under your user account. None of it is encrypted at rest beyond whatever your operating system does, and none of it is read by anyone but you and the application.

What leaves your computer, and when

While a session listens

A session keeps the last two minutes of what it hears in memory on your computer, in a ring that overwrites itself, and never writes it to disk. The orb on your screen shows a session is on.

From version 0.66 a session is also transcribed as it runs, unless you turn off Transcribe sessions as they run, in Settings, under Privacy, so Explain that can answer at once. The welcome shows that switch on, as suggested, and Leave them off, or turning it off there, keeps it off. Nothing is streamed before you answer that step, or, if you start a session before then, until you do. If you updated from an earlier version, nothing is streamed until you answer the notice the update shows by the orb, whose box, read when you press its Continue, keeps it off when you untick it. With it on, the audio of a session you started is streamed to Deepgram for as long as the session runs, to be turned into words as they are said, with the course words you add in Settings, under Audio. A quiet room is still sent; only the complete silence of a microphone that sends nothing, or of the seconds held back while Lectavi speaks, stops being sent after a second and a half, until there is sound again. None of that audio is written to disk, and nor are the words: they stay in memory on your computer while the session runs, for Explain that to read, and only the words each explanation was drawn from are kept, in History, as below. Deepgram’s terms let it store parts of the audio it is sent and train its models on them, since Lectavi’s requests do not opt out, as What we do not do says, so with this on that can be any part of a session, the speaker’s voice and anyone else’s nearby included. The seconds sent count toward your plan’s monthly hours. When the connection drops, Lectavi connects again by itself, at most twice in a stretch of listening, holding what is said meanwhile in memory, never on disk, and sending it first once the connection is back, except a stretch a press of Explain that already sent. When it cannot run, because you are signed out or offline, your hours are used up or the connection fails and does not come back, Lectavi says so once by the orb and the session goes back to sending a stretch only when you press Explain that. A recording has its own transcript, below.

With that switch off, and before version 0.66, nothing is sent while a session listens until you press Explain that, ask a question, or turn on Record this session.

When you press Explain that

Explain that sends the stretch it explains, the last 30 seconds or the last 2 minutes, to be understood. In a session that only listens, that stretch of audio is sent to Deepgram to be turned into words, and the words then go to OpenRouter, which routes them to a language model to draw the explanation. From version 0.66, while the session is transcribed as it runs, above, the words are already there, so the press sends no audio and only the words go. While it is still connecting, or connecting again after a drop, a press sends its stretch of audio as before, counted as before, and that stretch is not sent again once the connection is up. A silent stretch is not sent at all. While a recording runs the words already exist, so only they are sent. A drawing Live visuals makes on its own sends the last minute and a half the same way. The course words you add in Settings, under Audio, go with the audio to Deepgram and with the words to OpenRouter, so names and terms are heard and spelled right. From version 0.64, with Double check answers on, a second language model reads each drawing against the same words while its first line is said, and a drawing it finds wrong is drawn once more, or taken down. A drawing Live visuals makes on its own is read the same way, by the model that drew it, and one it finds wrong is taken down.

Lectavi talks you through the drawing, a line at a time. Each line is Lectavi’s own writing, and its text passes through our server to the speech service to be spoken, as a spoken reply does below. With spoken replies off, each line is shown on the drawing’s card for as long as it takes to read, and nothing is sent to be spoken.

A picture of your screen goes only when Live visuals is on, or, from version 0.66, when Screen sight is on and the words are too few. Live visuals is a setting, and it is off until you turn it on. From version 0.65 the welcome, and the notice an update shows, call it Live diagrams and suggest it, and it is turned on only when you press Continue there. Its switch there is shown on, unless you had turned it off before; if you updated from an earlier version, it is shown as you had it until you answer that notice. With it on, while Lectavi is hearing your computer (Online or Hybrid), Explain that sends one picture of your screen at the moment you pressed it, and each time it draws that stretch again, because you asked for it again or another way or the check found a drawing wrong, it takes and sends one more. From version 0.66, a press whose card for a look at the screen, below, you answered Not now, or did not answer, takes none, and nor does drawing that stretch again. While it records with your computer’s sound, Lectavi also looks at your screen every 20 seconds and keeps, in memory only, the last three pictures that changed, until the recording stops; a drawing Live visuals makes on its own sends those, or one taken then when there are none yet, and the same picture can go with more than one drawing. With it off, no picture of your screen is taken or sent for Explain that, apart from the one look Screen sight can ask for, below. Every picture is held in memory only and is never saved, on your computer or anywhere else.

From version 0.66, with Screen sight on, a press of Explain that whose stretch holds fewer than twelve words, or that you asked with words pointing at something, such as this or that, also takes one picture of the screen your pointer is on, with Lectavi’s own pill and marks hidden from it, and sends it with the words to the model that draws the explanation, through OpenRouter. It asks you first on the pill, with the card for a question about what is on the screen, which says why Explain that is asking, unless you chose a standing answer for that kind of look, or the welcome or the notice kept Always allow for it, as above, which Settings, under Privacy, can change or reset; a Not now, or a card nobody answers, sends the words alone, without the picture Live visuals would take either. If you updated from an earlier version, no such picture is taken, and nothing is asked, until you answer the notice the update shows by the orb. That picture is taken in place of the one Live visuals would take for the press, so no second one goes, and it is held in memory only and never saved. With Screen sight off, nothing is asked and no picture is taken.

From version 0.66, when the stretch you asked about holds only a few words, Lectavi explains the last two minutes of words it already has instead, and sends no audio for that. When the model’s answer cannot be used, Lectavi asks it once more by itself, at most twelve times an hour, with the same words: nothing is heard again, no audio is sent again, and no picture of your screen is taken or sent again. The drawing’s card shows that it is drawing while the answer is made and says so when it could not be drawn; its Try again explains the same words again, and only a stretch that could not be heard is sent to be heard again. You can drag the card anywhere on your screen.

While you record a session

Record this session is a switch in the pill, and the orb on your screen turns red while it is on. Audio is streamed to Deepgram to be turned into text, continuously, for as long as the recording is running. Without this there is no transcript, so it is not optional while transcription is on. The course words you add in Settings, under Audio, go with it, to help Deepgram hear names and terms.

When a recording cannot be transcribed as you make it, because you are signed out or offline, your plan’s hours for the month are used up, or the connection is too slow, it is still recorded and kept on your computer without a transcript, and none of its audio is sent. The app does not yet offer a way to have it transcribed later; what would be sent then is described below.

When you talk to Lectavi

Pressing Talk, finishing a circle around part of the screen, or saying “Hey Lectavi” with that switch on, opens the microphone for one question. It closes when you stop talking, when you press Talk again, or after twenty seconds, and Lectavi never opens it on its own: there is no conversation mode, and nothing Lectavi says opens it again. The question’s audio is streamed to Deepgram to be transcribed. Lectavi does not keep it: it is never written to disk, and it is not recorded into a session you are recording. The words of your question are sent to OpenRouter, with the topic you are on, its title, the explanations and notes Lectavi wrote for it and the speaker’s words it was explained from, so a language model can work out what you asked and answer it. From version 0.65 a question about the topic you are on goes with the conversation it is part of too: its last six turns, your questions, typed or asked out loud, and Lectavi’s answers, each cut to 600 characters. A question with no topic on is answered the way the chat answers one, below, and from version 0.65 in the conversation that is open: when that conversation is about a session or a recording, the notes Lectavi wrote for it and the speaker’s words they cite go with your question, as they go with a typed one. From version 0.66, while a session you started is transcribed as it runs, as above, a question you ask out loud in it also goes with the speaker’s words heard in the last two minutes, taken from the words already transcribed in memory, so no more audio is sent for it: with a topic on, the words heard since Lectavi’s latest explanation of it, and with nothing explained yet, the last two minutes, unless the conversation that is open is about a session or a recording, whose notes and cited words go instead. They go as the speaker’s words, never with a question that is looked up on the web, never while you record the session, never while its transcript is connecting or after it has stopped, and, if you updated from an earlier version, not until you answer the notice the update shows by the orb. They are not kept: only what Lectavi’s answer quotes of them is kept with that answer in History, on this computer, and never synced. From version 0.64, a question about the world can first be looked up on the web, as described under When Lectavi looks something up; with a topic on, only a question about something that changes over time, or one you ask Lectavi to look up, and from version 0.65, in a conversation about a session or a recording, only one you ask Lectavi to look up.

From version 0.64, a spoken answer is checked too, unless Double check answers is off in Settings, under General. Before an answer is shown, a second model checks it: your question and the answer go to OpenRouter again, with up to two of your earlier questions in the same conversation and Lectavi’s answers to them, and the recording’s words, the notes, the file text and the sources that already went with the question. A picture of your screen or an attached picture is never sent twice. A question about what Lectavi is explaining is checked without your earlier questions, though from version 0.65 its answer is written with them. From version 0.66, as in the chat, a plain sum answered by that sum alone, which Lectavi works out again on this computer and finds right, is not sent again when nothing else went with the question; a question about what Lectavi is explaining goes with the speaker’s words it was explained from, so even a sum asked about it is sent again. A reply with no number, maths, date, unit, moment in the recording, fact that changes over time, grade of your answer or line it says the speaker said, and no drawing, is said at once and checked as it is said, and a correction is said only if the check finds it wrong.

A spoken reply is made from its text. The text of Lectavi’s reply, which can quote a few of the speaker’s words, passes through our server to the speech service, which turns it into speech: from version 0.64 that is Inworld, and Deepgram when Inworld cannot make the reply, and before 0.64 it was Deepgram. If you updated from an earlier version, Deepgram speaks every reply until you answer the notice in the app that tells you about Inworld. From version 0.66 Lectavi’s own lines no longer repeat your question, and the AI model that writes a chat or a spoken answer is told to answer without repeating your question, but an answer can still carry some of your words, so the words of a question you asked out loud can be in the text of the reply that is turned into speech. Also from version 0.66, a quick command is never answered out loud: opening something, scrolling, moving the pointer or clicking, stop, quiet, thanks, opening Settings, Chat or History, asking only where something is on your screen when the answer is one short sentence that only says where it is, and a request Lectavi refuses or cannot do yet. Its line, when it has one, is shown in the small bubble by the orb, or in the captions when there is no room there, and a line that says what Lectavi opened, found or clicked names that thing, often in your own words, but nothing of it is sent to be spoken; an open that worked shows nothing at all, since what opened says so. Our server does not store or log that text, only how many characters it was, to keep your plan within its monthly allowance of spoken replies, and which service spoke it. Inworld receives the text and the voice, never audio and nothing that names you: the request carries our key, not yours. Without Zero Data Retention, which Lectavi’s Inworld account does not have, Inworld keeps the text it is sent in its logs and does not publish for how long. Its terms say it does not train its general models on material that is not public, and let it use what it is sent to run and improve its service (Inworld’s pages, read on November 1, 2026). While a reply plays the room is not heard, and while you ask a question the room’s microphone is not: a recording holds silence for those seconds and notes which seconds they were, so nothing Lectavi says is recorded as the speaker, and nor is your question. The short sounds the app plays are held back from a recording the same way. Spoken replies can be turned off in Settings, and replies are then shown in words instead: as captions or, from version 0.65, in a small bubble by the orb or in the Chat panel.

A question you ask out loud with nothing explained yet to answer from, or one Lectavi looks up online, is kept on this computer as a Chat conversation with your words and the answer, like a typed one, and syncs without your words, its conversation titled “Asked out loud” when a question asked out loud began it. Before version 0.65, any other question you asked out loud, such as one about the topic you were on, was kept with its reply in memory only, and dropped when the topic changed or the app closed. From version 0.65, the questions you ask out loud that Lectavi answers are all kept on this computer, with the answers, in the conversation that is open, which History lists, whether an answer was shown in the Chat panel, in the small bubble by the orb or with a drawing; the bubble itself keeps nothing. From version 0.66 a quick command, as above, is not kept at all: it is not in History, starts no conversation and joins none. Your spoken words never sync, and Lectavi’s other answers to what you ask out loud, about the topic you are on, a drawing of it, a place on your screen or the mouse, stay on this computer too and never sync, since they can quote the speaker or describe your screen. The words of a question that asked for a drawing are kept on this device with that drawing, and are never synced.

When you say “Hey Lectavi”

“Hey Lectavi” is a setting, and it is off until you turn it on. Settings describes it in a few words; in full, it does this: Lectavi keeps the microphone open and listens, on this computer, for those two words. Nothing is sent anywhere until it hears them. Then it listens for your question exactly as a press of Talk does, and the question goes to the speech service. The listener is a small speech model on this computer, not a service, and it listens for those two words and nothing else: what it hears is never written to disk. A recording that runs on the same microphone takes the microphone over, and the listening stops while Lectavi speaks or a question is open. When it hears the phrase, the microphone opens for one question exactly as a press of Talk opens it, above, and what you say next is streamed to Deepgram. If you say nothing in the next four seconds, that much room audio has been sent and the microphone closes. Before version 0.65 Lectavi then said “Heard Hey Lectavi, then nothing” by the orb; from version 0.65 it says nothing and asks no language model anything, and the orb goes back to rest. The Test button in Settings listens for five seconds and sends nothing. From version 0.64 “Hey Lectavi” has three levels, and the most sensitive also wakes on “Hi Lectavi”. While “Hey Lectavi” is on, Lectavi notes numbers in its own log on this computer: how loud the room’s speech is, how often it woke, how loud a phrase it almost heard was, why it let a wake go, when the microphone sent only silence, and how soon after speech you pressed Talk, never the sound and never the words.

When you ask Lectavi to show you something

Screen sight is a setting, and it is off until you turn it on. When you ask Lectavi to show you something, it takes one picture of the screen your pointer is on, at that moment and only then. The picture goes to the AI model in one request, through OpenRouter, to work out where to point. Lectavi does not keep it: it is never written to disk, and it is gone once the answer arrives. With more than one screen, the system pictures the others too, and Lectavi throws those away at once without looking. From version 0.66 your question goes with the conversation that is open, so the model can tell what “this” or “that one” means: its last six turns, your questions and Lectavi’s answers, each cut to 600 characters, though never a line Lectavi said out loud about the topic you were on, a drawing, a place on your screen or the mouse, since such a line can quote the speaker, and, while a session or a recording runs, the title of the topic you are on and Lectavi’s latest explanation of it, each cut to 300 characters, never the speaker’s words. If you updated from an earlier version, none of that goes until you press Continue on the notice an update shows by the orb, which says so. From version 0.65 your question and Lectavi’s answer are kept in words, in the conversation that is open on this computer, and never sync; the picture never is. From version 0.66 they are not kept when you asked only where something is and the answer is one short sentence that only says where it is, which is shown by the orb and not said aloud; from version 0.66 an answer that explains, such as what a mistake is or what a button does, is kept and said aloud.

Circling part of the screen counts as asking: the picture is taken when you start, held in memory for up to a minute until you ask your question about it, and then sent the same way. The model answers with the places to mark, and Lectavi draws the marks itself. Text in the picture is never taken as an instruction, and links, email addresses and phone numbers are taken out of what Lectavi says aloud. With Screen sight off, asking Lectavi to point takes no picture, and it says so, and Point is not offered. With it on, pointing and circling each ask you first on the pill, with Allow once, Always allow and Not now, unless you chose a standing answer for that kind of look, which Settings, under Privacy, can change or reset. From version 0.65, pressing Continue in the welcome, or in the notice an update shows, when that turns Screen sight on, keeps Always allow for both, and for a question about your screen in the chat, unless you had chosen Never; when Screen sight was on already and is left on, Continue leaves those answers as they were, and turning Screen sight on in Settings does not keep Always allow either. From version 0.65, pressing Stop, or saying stop, while the card for a circle is still waiting takes that card down, and the part you circled is sent nowhere.

When you let Lectavi use the mouse

From version 0.64, on Windows only, Let Lectavi use the mouse is a setting in Settings, under Privacy, and it is off until you turn it on. It needs Screen sight. From version 0.65 the welcome, and the notice an update shows, offer the same switch, shown off unless it is on already; turning it on there sets no Always allow for it, so it still asks first. Settings describes it in a few words; in full, it does this: When you ask Lectavi to find something on the screen, it can move your pointer to it and scroll the window under it. It asks first, unless you asked it to click with Let Lectavi click for me on, and stops the moment you move the mouse or press Esc. It never types, drags or right clicks, and it clicks only when Let Lectavi click for me is on. To find what you asked for it may look at the screen your pointer is on up to six times for one question, each picture sent to the AI model through OpenRouter and never stored. It needs Screen sight.

Lectavi asks first on the pill, with Allow once, Always allow and Not now, and when you ask through “Hey Lectavi” it asks every time, even after Always allow. In versions 0.65 and 0.66 it also asked every time you asked it to click something, even after Always allow; from version 0.67 a click you ask for with Let Lectavi click for me on asks no card, as When you let Lectavi click for you says, and a card nobody answers counts as Not now. Every card’s buttons answer nothing for a moment after the card appears, so a click meant for something else is not taken as an answer, and a card whose question was overtaken, by a newer question, Stop or a lock, is taken down at once and counts as Not now. The first picture is taken after you say yes, or, for a click, after you ask, and each one after it only right after Lectavi’s own scroll or its own move of the pointer: for a click, back where you left it once it has scrolled to the thing, and, after the ring has been shown, onto the thing, never while you are using the mouse. Each picture that goes to the AI model goes in one request and is not kept: a small copy, in grey, stays in memory only until the next picture, to tell whether the page moved, and nothing is written to disk. With this switch alone Lectavi makes only two kinds of input, a move of the pointer and a turn of the wheel, and nothing is ever typed or clicked by it; from version 0.65 Let Lectavi click for me adds a third, one press of the left button, below. Nothing on a page can make Lectavi type, and text in a picture is never taken as an instruction, but a page can be written to mislead the AI model about what is where. At worst Lectavi then marks the wrong thing, moves the pointer to it or scrolls the wrong way, within what you asked for; with clicking on, the ring is on that thing, where you can see it, and the strip says what Lectavi is about to click, for a moment before it clicks. Lectavi never moves your mouse by itself: it moves the pointer, scrolls or clicks only when you ask, after a yes on the pill or, for a click, with Let Lectavi click for me on, shows a strip on the screen while it does, and stops the moment you touch the mouse or press Esc.

When you let Lectavi click for you

From version 0.65, on Windows only, Let Lectavi click for me is a setting in Settings, under Privacy, below Let Lectavi use the mouse. It is off until you turn it on, and it works only while the mouse switch and Screen sight are on; turning either off turns it off. Settings describes it in a few words; in full, it does this: When you ask Lectavi to click something, it finds it, draws a ring on it and clicks it: turning this on is your yes, so it does not ask again, and it clicks at most ten times an hour. Before it clicks, the ring and a strip saying what it is about to click stay up for a moment, so you can press Esc or move the mouse to stop it. It then puts the pointer on the thing and checks, on this computer, that the place looks as it did when it was found; only if it looks different with the pointer on it, as a link or a tab may, does the AI model read one more picture of the screen, with the point it would click marked, and it clicks only if the thing at that point has the same words and does the same as when it was found. It clicks once, with the left button, where the ring is, and never types, double clicks, drags or right clicks. It checks your words, the words on the thing and the window’s title against lists of words about buying or paying, sending or posting, deleting or installing, and passwords and sign ins, and does not click on a match or on an ad; the lists are not a guarantee, and in a browser, a PDF reader, Word, Excel, PowerPoint, OneNote, Notepad or VS Code the window’s title, and the words of a link, a video or a result it would open, are checked only for the names and phrases on them, such as checkout, sign in or buy now. It stops the moment you move the mouse or press Esc.

Before version 0.67 a click needed two yeses on the pill, each asked for that click: the mouse card above, asked every time, even after Always allow, and a second card asked with the ring on the thing. From version 0.67 there is no card: with Let Lectavi click for me, Let Lectavi use the mouse and Screen sight on, a click you ask for goes ahead, and that is so whoever asks, a press of Talk or “Hey Lectavi” alike, so with the switch on, someone near your computer who says “Hey Lectavi” can have it click; turn the switch off to stop that. A scroll or a move you ask for that is not a click still asks the mouse card. If you turned this switch on before version 0.67, when Lectavi still asked twice, it is off after you update, and Lectavi clicks nothing until you turn it on again. Lectavi finds the thing as above and draws the ring on it; while it finds it, it moves the pointer only to scroll, never onto the thing, though the page it scrolls can bring the thing to rest under the pointer. When it scrolled to find the thing, it then puts the pointer back where you left it, waits a moment and takes one more picture, one of the six, which stays on this computer and is what the click is later compared with, so nothing the pointer brought up while it scrolled is part of it. Then, with the ring on the thing and the pointer where you left it, the strip says what Lectavi is about to click and that Esc cancels, for a little over a second, before the pointer moves at all; Esc, a touch of the mouse or a key, Stop, a newer question or a lock then ends it with no click.

You can ask for a kind of thing rather than its words, such as “click on a random video”, “click the first result” or “click any link”: from version 0.67 the AI model is then told to pick one item of that kind that the page shows as its content, never an ad, a sponsored or promoted item, a Subscribe, Sign in, Buy or Download button, or a menu or a sidebar, and the words on it are that item’s own title. And when you ask Lectavi to open something and then click, such as “open Khan Academy and click on a random video”, from version 0.67 it opens it, waits a few seconds for it to load, and then takes the click as if you had asked for it on its own, with every check here; with clicking off it opens it and says what it did not do.

Before the ring is drawn, Lectavi refuses the click, on this computer, when what you said, the words on the thing or what the AI model calls it name buying or paying, sending or posting, deleting or installing, passwords or sign ins, the words you quote from the thing included, as in “click the button that says Submit”, and anything you ask for after “then”; when you ask for a right, double or middle click; from version 0.67, when the words on the thing mark it as an ad, sponsored or promoted, which Lectavi says as an ad; when the window is an app’s own, not a browser, a PDF reader, Word, Excel, PowerPoint, OneNote, Notepad or VS Code, and its title names money, a sign in or a password, installing or updating, writing or sending a message or a post, deleting an account or a file, or the computer’s own settings and tools; when the window is one of those and its title holds a name or a phrase on Lectavi’s list, such as checkout, sign in, PayPal or online banking, or is the browser’s own Settings, Downloads, Extensions, Passwords or Payment methods page; when the AI model reads no words on the thing, or says it does anything but take you somewhere, play, open, select or expand, and from version 0.67 Lectavi says so in its own words when the model could not tell what it does; when none of the words you called it by are on it, so say the words it shows, such as “click Next”, and not only its colour or its place, unless you asked for a kind of thing, as above; when what it found is not one small thing, is one of Lectavi’s own windows, or is in a window with no title; and when the ring would sit where the orb, or what the pill shows above or below it, could cover it, in which case Lectavi asks you to drag the orb somewhere else. A click that names nothing, such as “click here” or “click it”, is not a click: Lectavi asks you to say what to click. It offers the click only when Lectavi’s check of its own overlay shows the ring drawn exactly where the click would go. These are checks of words against lists, and the words on the thing are the ones the AI model copies off the screen, so they are not a guarantee: a page can label a button anything, and a title is matched word by word. The browsers and PDF readers Lectavi knows are Chrome, Chromium, Edge, Firefox, Brave, Opera, Vivaldi, Arc, Safari, Adobe Acrobat, Foxit Reader and SumatraPDF. In those, and in Word, Excel, PowerPoint, OneNote, Notepad and VS Code, the title is the page’s or the file’s own words, so there its topic words are not checked, unless one is the whole title: a page or a file whose title names a bank, an install, an update or a message, such as a bank statement open in Excel, but no name or phrase on the list, is not caught. Any other app’s window is checked for the topic words anywhere in its title. From version 0.67, in those windows the words on a thing that takes you somewhere, plays or opens, when they are three words or more, are read as its title too: they are checked only for the names and phrases on the list and phrases such as buy now, add to cart or install now, so a video called “How to install Python” can be clicked, while a button that says Install, Buy now or Sign in cannot. In an app’s own window, a bank that does not say bank in its title is not caught either. That is why the ring, and the moment the strip says what Lectavi is about to click, come first.

With the ring up Lectavi waits for you to let go of the mouse, holds Esc and shows the strip, which says what it is about to click for a little over a second, then puts the pointer on the thing and waits a moment. It checks on this computer that the screen is the same one, that the window under the pointer is the same window with the same title, which still names nothing above, that no key or mouse button is held and that you have not touched the mouse. It then takes one more picture and compares a small area around the point it would click with the same area of the picture it found the thing in, or, when it scrolled to find it, of the picture it took with the pointer back where you left it: almost every point must be close to what it was, and a change of shade over the whole area, such as a highlight the pointer brings up, does not count. The ring is drawn just outside the thing, and Lectavi’s own ring and spark are left out of the comparison, never what is inside the thing; when too little of the area is left to compare, Lectavi says it could not check that place, does not click and sends nothing. When the area looks the same, that picture stays on this computer, is sent to no one and is not kept. Only when it looks different with the pointer on it does the AI model read that one picture, once, with the point it would click marked on the copy that is sent, never on your screen; Lectavi first moves the pointer back where you left it, off the thing while the model reads. A link or a tab often changes when the pointer is on it, an underline or a colour, so a click on one may send that picture. Lectavi goes on only if the model names one small thing whose box holds that point, with the same words on it and doing the same as when it was found; it then puts the pointer on the thing again, waits a moment, takes one more picture, which stays on this computer, and clicks only if the area still looks as it did in the picture the model read. That is the only time a picture taken after the ring is shown leaves your computer, and it is not kept either. The grey copies go when the click is made or refused. If you use a tool that clicks wherever the pointer rests, it can click where Lectavi’s pointer rests while it scrolls to find the thing, or on the thing before Lectavi’s own click; Lectavi takes that as your own input, stops, and does not click again. After Lectavi’s own click it puts the pointer back where you had it and stops watching the mouse, so such a tool can click again there; Lectavi cannot see or stop that click. A click the mouse driver says it made is told as made, even when Stop, Esc or a newer question arrives while it is on its way. It does not click when Windows has the left and right mouse buttons swapped, and it clicks at most ten times an hour. The AI model that finds the thing is told it may only point to it: it never decides to click, and what you said, not what it saw, decides whether a click is asked for; what it reads off the screen can stop a click, and can never ask for one or move it somewhere else. Lectavi still never types, never double clicks, drags or right clicks, and never presses a key. Lectavi clicks only when you ask, with its switch on, once, where you can see the ring, and never types.

When you let Lectavi open apps, sites and folders

From version 0.65, Let Lectavi open apps, sites and folders is a setting in Settings, under Privacy, and it is off until you turn it on. It needs neither Screen sight nor the mouse. Settings describes it in a few words; in full, from version 0.66, it does this: When you ask Lectavi to open a website, a folder or an app, it opens it for you straight away: turning this on is your yes, so it does not ask again, and it opens at most twenty times an hour. It opens a site it knows at that site’s own address, a folder on its list, an app on its list or one installed on this computer with a Start menu shortcut, or the front page of a website whose address you say, never a file. It skips any shortcut whose name says it installs, removes, updates or repairs something, or is a command prompt, a shell or a system tool. An address its rules read as a copy of a site it knows is never opened as you said it, though its rules can miss a copy, and it refuses a name or an address its lists tie to paying, shopping, signing in, sending or installing. A site it opens can still offer a download, a purchase or a sign in, as any site can. Opening never types, clicks or moves your mouse. On a Mac, where Lectavi opens no app, the setting is called Let Lectavi open sites and folders, and its words name only a website or a folder and say it never opens a file or a program.

Before version 0.66 Lectavi asked first on the pill, with Allow once and Not now, every time and whoever asked, and showed on that card what it would open. From version 0.66 there is no card: with the switch on, an open you ask for happens at once, and what opens is how you see what Lectavi opened, since Lectavi then says and shows nothing more. That is so whoever asks, a press of Talk or “Hey Lectavi” alike, so with the switch on, someone near your computer who says “Hey Lectavi” can have it open a site, a folder or an app; turn the switch off to stop that. It opens at most twenty times an hour; an ask Lectavi refuses, or one a newer question, Stop or a lock overtakes before it opens, does not count toward that. With your screen locked it opens nothing. If you turned this switch on before version 0.66, when Lectavi still asked every time, it is off after you update, and Lectavi opens nothing until you turn it on again.

What it opens comes from its own lists on your computer, never from an AI model or from your screen. The sites it knows are a few hundred well known ones, such as Khan Academy, Quizlet, Wikipedia, Google’s, Reddit, Netflix and Spotify, each opened at the address on Lectavi’s list, whatever you called it, so a name it heard a little wrong still opens the real site. The folders are your Downloads, Documents, Desktop, Pictures, Music and Videos. The apps, on Windows only, are Notepad, Calculator, Paint, File Explorer, Word, Excel, PowerPoint and OneNote, and from version 0.66 any app installed on this computer that has a shortcut in your own or every user’s Start menu Programs folder, found by the shortcut’s name on this computer and opened by handing that shortcut to Windows’ File Explorer. Lectavi reads only the names of those shortcuts and their folders, never what is inside them, never the Startup or Windows’ administrative tools folders, and it skips any shortcut whose name or folder names an installer, an uninstaller, an updater or a repair tool, a command prompt, a terminal or a shell, the registry, settings, security or administration tools, a store, a wallet, a bank or a password manager, or a site it refuses. That is a list of words, and it can miss a program whose name says none of them. From version 0.66 those words include every name on Lectavi’s own lists of password managers, wallets, banks, brokers, app stores and game launchers, such as Bitwarden, MetaMask, Steam, Epic Games and Battle.net, in a shortcut’s name or its folder’s. From version 0.66 it also skips a script host or a language’s interpreter, such as Python, and a remote access tool, such as TeamViewer, AnyDesk, Remote Desktop or Quick Assist, by name; a few of those names that are ordinary words too, such as Ruby, Node or Remote, only alone or beside a version number or a word that marks the tool, such as Desktop in Remote Desktop. A request that also asks for something Lectavi refuses, such as “open Discord and send a message”, is refused whole, and nothing is opened. An app on the list, or a folder, that is not on this computer is said to be missing, and nothing is opened. From version 0.66, with this switch on, the AI model that works out what you asked can name what you meant to open, and it names only that, such as Wikipedia or your Downloads folder, as a plain name that Lectavi looks up on the same lists; no AI model ever writes an address that Lectavi opens. Nothing new leaves your computer for that: the words of your question already go to that model.

For a web address you say out loud, Lectavi opens only the site’s front page, https and the site’s name with nothing you said after the name. It opens no address with a user name or a port in it; none whose name is a bare IP address, a name whose ending Lectavi’s lists say only your own computer or network uses, such as .local or .home, a name ending in .zip or .mov, or a lookalike written in another alphabet; none whose name holds a word on Lectavi’s lists about paying, signing in, sending or installing, such as paypal.com; and none where what you said after the name names a program, an installer, an archive or a script by an ending on Lectavi’s list, such as .exe, .msi, .zip or .ps1. These are lists, and a list can miss a name or an ending, but what you said after the name is never opened: only the site’s front page is. From version 0.66 an address you say is compared with the sites Lectavi knows, by rules that catch some kinds of copy. A slip, such as a letter dropped, doubled or swapped in a name of five letters or more on the site’s own ending, or .con for .com, opens the real site at the address on Lectavi’s list, and so does the name of a site that is not on .com said with .com, such as archive.com for archive.org, though an address on Lectavi’s list of other real sites, such as interest.com, is opened as you said it. What the rules read as a copy is not opened as you said it, and Lectavi says which site it thinks you meant: a digit or a lookalike letter in place of a letter, one changed letter in a name of six letters or more, a slip on another real ending, such as netflx.net, a site’s name and ending inside another address, such as youtube.com.evil.io, a site’s name of four letters or more joined to another word by a hyphen, such as youtube-login.com or hulu-free.com, a site’s name of five letters or more with a word from Lectavi’s list glued on, such as youtubepremium.com, or two words glued after it, the first from that list, such as robloxfreerobux.com, a site’s name with a second ending after its own, such as netflix.com.co, or a site’s name on a cheap ending such as .xyz or .info. A near miss of a site for paying, shopping, banking or sending, such as paypa1.com, is refused, and so is such a site named by its name or its address alone, such as PayPal, Amazon or WhatsApp. From version 0.66 a name of three letters or more of a bank, a payment site or a messaging site on Lectavi’s lists, such as Chase’s, is refused on a country’s ending too, such as chase.co.uk, and, when it is longer than three letters and not an ordinary word, on any other ending, such as telegram.dog. These are rules and lists, and they can miss a copy: a site’s name with a word glued on that is not on Lectavi’s list, such as youtubevanced.com, a name of three letters or fewer, or one that is an ordinary word, such as Weather, with a word added, a lookalike of a site Lectavi does not know, or a site’s name on another real ending alone, such as .org or a country’s, may not be caught, and nor is a public name that leads back to your own computer or network, such as one ending in nip.io, so look at what opened.

Like any question you ask out loud, the request is first transcribed by Deepgram. Before version 0.66, like any spoken reply, Lectavi’s answer, which names what it opened, went as text to the speech service, Inworld or Deepgram, to be said. From version 0.66 an open is a quick command: one that works says and shows nothing, one that does not is shown by the orb, and nothing about it is sent to be spoken or kept in a conversation. The page you open is fetched by your browser, as if you had opened it yourself, and it can offer a download, a purchase or a sign in as any page can. The log on your computer says what kind of thing was opened and, for a site, a folder or an app on Lectavi’s own lists, which one; never an address you said, and never the name or the place of an app you installed. Lectavi opens an app, a website or a folder only when you ask, only with this switch on, and only a site it knows, an app on this computer, a folder on its list or the front page of a web address you say.

When you chat with Lectavi

The questions you type into the chat are sent to OpenRouter, with the earlier turns of that conversation and, when the conversation is about a session or a recording, the notes Lectavi wrote for it and the speaker’s words they cite. When you attach a file, a PDF, a Word document, a slide deck or a text file, it is read on your computer and its text is sent with your question; the file itself is never uploaded. The text stays with that conversation, in memory only, and goes with each later question you ask in it, until you close the Chat panel, which drops it. When you attach a picture, a PNG, JPEG, WebP or GIF, the picture itself is sent with your question, and is kept and dropped the same way. Asking a question that needs a look at your screen, such as “what is this”, needs Screen sight on, and asks you first, with Allow once, Always allow and Not now, unless you chose a standing answer for it, Always or Never, or the welcome or the notice kept Always for it, as above; a standing answer is kept in Settings, under Privacy, and can be changed or reset there. With that yes, one picture of the screen your pointer is on goes with the question, the way Screen sight sends one, and is not kept.

From version 0.65 a conversation holds what you type and what you ask out loud alike. It ends when you close the Chat panel, press New conversation, or ten minutes after its last question or answer, and the next question starts another. History keeps each one, and opening one there continues it. The earlier turns that go with a question are that conversation’s, so they can include what you asked out loud and Lectavi’s answers about the topic you were on or your screen.

From version 0.64, two things are added. A reply that has no source says it is from memory. And unless Double check answers is off in Settings, under General, every answer is checked. Before an answer is shown, a second model checks it: your question and the answer go to OpenRouter again, with up to two of your earlier questions in the same conversation and Lectavi’s answers to them, and the recording’s words, the notes, the file text and the sources that already went with the question. A picture of your screen or an attached picture is never sent twice. From version 0.66 one kind of answer is not sent again: a plain sum, such as “what is 15% of 80”, answered by that sum alone, which Lectavi works out again on this computer and finds right, with nothing else in the answer to read and nothing else sent with the question, such as the recording’s words, a file, a source or a picture of your screen. With the switch off, an answer is not sent again to be checked: its arithmetic is still redone on this computer, and a looked up answer’s lines are still checked against their pages, below.

When Lectavi looks something up

From version 0.64, when a question is about the world rather than your recording, and Look things up online is on in Settings, under Privacy, Lectavi searches the web before it answers. The search call goes to OpenRouter with your key: the words of your question, the question before it in the same conversation, and today’s date. OpenRouter sends a search query written from those words to Exa, and a model reads the pages Exa returns to write the answer. A second model checks each line against its page before you see it. Your audio, your screen, the files you attach and a recording’s words are never part of a search. When you ask Lectavi to check one of its answers online, that answer’s words go too, and they can quote a few of the speaker’s words. When you ask Lectavi out loud, while it is explaining a session, to check its explanation online, the summary of that explanation goes too, as the answer to check; otherwise a question you ask then goes to the search as any other does, without Lectavi’s explanation. A question you ask out loud is sent to the search as text when it is looked up. Lectavi never visits the pages itself: a source opens in your browser only when you press it. The answer keeps each source’s address, title, publisher and the words it quoted, at most 300 characters of the page, and these sync with the conversation; where an answer and a recording disagree, only the source’s line syncs. From version 0.65 Look things up online is one switch, on until you turn it off, and the welcome and the notice tell you so rather than ask. Before 0.65 the setting had three choices, Automatically, When I ask and Off, and a When I ask chosen then is kept. From version 0.66 fewer questions are looked up by themselves: maths, from a sum to calculus, and questions about a definition, settled science, grammar or code are answered without a search, unless you ask Lectavi to look one up; a question about people, places, events, products, prices, laws, a medicine’s dose, side effects or safety, a condition’s symptoms, or anything that changes over time still is. Lectavi tells these apart on this computer by lists of words, such as the course acronyms it reads as settled science or code, like TCP or DNA, so a list can miss one and send a settled question to the search, or keep a question about the world off it. A question that asks for a price or a rate it does not give, a schedule, or what is current, latest or happening today is looked up even when it holds numbers or a percentage, and converting a currency is looked up rather than worked out as a sum, since its rate changes, unless the question gives the rate itself, such as “if 1 dollar is 0.9 euros”, when it is worked out as maths.

What the two services keep, as their own pages said when we read them on November 1, 2026. OpenRouter keeps no prompt or answer unless the account turns its logging on, which Lectavi’s has not, or its safety screening flags it as possible child sexual abuse material, which it may keep for review and report where the law requires, and it keeps each request’s counts, such as its tokens and how long it took. OpenRouter’s zero data retention setting does not reach web search, which follows the search service’s own terms. Exa’s privacy policy says it uses search queries to improve its products, including to train its models, and does not say how long it keeps them; Exa offers zero data retention to its enterprise customers only, and Lectavi has no such agreement with it. So take the words of a question that is looked up as words Exa may keep, and turn Look things up online off to keep every question off the web.

Lectavi can look up a question about the world on the web, and shows every source it used with a link. Only the words of your question and the one before it go to the search, through OpenRouter and Exa: never your audio, your screen, a file or what a recording says, unless you ask it to check an answer online, which sends that answer’s words.

When a saved recording is transcribed

A recording kept while you were offline, signed out or over the month’s hours has no transcript yet. The app does not yet offer a way to have it transcribed later, so for now none of its audio is sent. When that is offered and you ask for it, once the month’s hours are back, Lectavi sends that saved recording to Deepgram, streamed from your disk faster than real time, and only then: never on its own, and never for a recording that already has a transcript. Deepgram receives the audio as a stream, the same way it does while you record. The file itself is not uploaded and stays on your disk. A recording that was heard whole and had nothing in it is marked, and is never sent again.

After a recording

The transcript text, with the course words you add, is sent to OpenRouter, which routes it to a language model, to write the notes, the claims, the flashcards and the visuals. OpenRouter currently routes to models from Google and Anthropic, depending on the task. From version 0.64 Lectavi tells OpenRouter which services may run each model, and no other: a Claude model runs on Anthropic’s own service or on Google Cloud (Vertex AI), and a Gemini model on Google, in AI Studio or Vertex AI. Before version 0.64 OpenRouter chose, and could also run a Claude model on Microsoft Azure or Amazon Web Services. Writing the notes sends text, never audio.

When a topic is something that can be photographed, such as an organ, an apparatus or a place, a short search phrase for it is sent to the public APIs of Wikipedia and Wikimedia Commons. Only pictures in the public domain or under a CC0, CC BY or CC BY-SA licence are used, and each is downloaded to your computer with its author, its licence and a link to its source. Topics about ideas, such as a formula, are drawn by the app instead and send nothing.

When you ask for a quiz

Quiz, on a topic in History, sends that topic’s notes, its explanations and the speaker’s words it was explained from to OpenRouter, which routes them to a language model to write the questions. When you mark the quiz, each question you got wrong goes again with its right answer and the option you chose, so the model can mark it and say why. The quiz and your answers are kept on this computer and never sync.

If you sign in

You sign in with your Google account, in your browser. Google confirms who you are to Lectavi’s sign in service, which runs on Supabase, and tells it the basic profile it shares at sign in: your email address, your name, your profile picture and your Google account’s identifier. Supabase keeps them with your account, and of these Lectavi itself uses only the email address. Google learns that you signed in to Lectavi, as it does for any app you sign in to with Google. Lectavi’s use and transfer of information received from Google APIs, at sign in and through a connection an earlier version made, adheres to the Google API Services User Data Policy, including the Limited Use requirements. You can remove Lectavi’s access to your Google account at myaccount.google.com, under Third-party apps with account access.

Your email address, your sessions, notes, claims, flashcards and review history, every topic Lectavi explained in a session or a recording, saved or not, with its title, every drawing in it with its steps, example and question, and Lectavi’s notes on it, the explanations you save from Explain that, the moments you mark, and your conversations with Lectavi are synced to Supabase so they are available on another device and survive the loss of this computer. A saved explanation syncs without the stretch of transcript it explains. A topic syncs without the speaker’s words it was explained from. A drawing you asked for out loud syncs without the words of your question. A conversation syncs without the speaker’s words it quotes and without anything of the files you attached, not even their names. From version 0.64, an answer’s sources sync with it, each page’s address, title, publisher and the words it quoted, and how an answer was checked never syncs. From version 0.65 a conversation also syncs without Lectavi’s answers to what you asked out loud about the topic you were on, a drawing of it, a place on your screen or the mouse, which stay on this computer; one it looked up online syncs like a typed one. Audio and transcript segments are never synced, which is a deliberate design decision rather than a policy one: a fifty minute recording is around a hundred megabytes, and transcripts are the largest table in the database. Sync can be switched off under Account in Settings, and your notes and cards then stay on this computer.

Your plan pays for transcription, notes and spoken replies through our allowance server, a function that runs on Supabase, whether or not sync is on. For each recording, each press of Explain that in a session, each saved recording once that is offered, and each question you ask out loud to be transcribed it hands the app a Deepgram token that lasts thirty seconds, and once a month it issues an OpenRouter key of your account’s own, whose spending limit is your plan’s. To keep each plan within its monthly allowance it records when each was handed out, how many minutes of audio were transcribed and how many characters of spoken replies were made. It never receives your audio, your transcripts or your notes. With sync on, your notes go to our database on Supabase as the paragraph above says, never through it. The one thing of yours that passes through it is the text of a spoken reply, on its way to the speech service, Inworld from version 0.64 or Deepgram, which it neither stores nor logs. The OpenRouter key’s name carries your account’s random identifier, so OpenRouter can count what it spends against your plan.

If you subscribe

Payment is handled by Stripe. Lectavi never sees or stores your card details. We receive your email address and which plan is active, which is what decides what the app lets you do.

Updates

The app checks GitHub for a newer version, anonymously. On Windows it downloads the update and installs it when you quit or when you ask it to. On a Mac it only says one is out and, when you ask, opens that version’s disk image on GitHub in your browser. It never installs an update, or starts downloading one, while you are recording. Nothing about you is sent with the check.

What earlier versions could do

Earlier versions could connect Google Calendar, Google Classroom and Canvas, use a phone as the microphone, import an audio or video file, and fetch the captions of a video posted online. None of these can be reached in this version: nothing from them is read now, and a stored connection is not used. If a Google connection was made in an earlier version, Lectavi only reads with it and never writes, and the scopes it asked for do not allow it to; it is held to the Google API Services User Data Policy as If you sign in says, and the same page at myaccount.google.com removes it. This section changes when one of these returns.

The full list of companies, with what each receives and why, is on the subprocessors page. If a service is not on that list, Lectavi does not send it anything.

What we do not do

Why we process your data

Where the law asks for a legal basis, ours are these. Providing the service you asked for, which is the explanations, the transcription, the notes and the sync you signed up for. Your consent, for anything optional: recording a session, turning on Live visuals, Screen sight, “Hey Lectavi”, Let Lectavi use the mouse, Let Lectavi click for me or Let Lectavi open apps, sites and folders (called Let Lectavi open sites and folders on a Mac), letting a question look at your screen, attaching a file, subscribing. Our legitimate interest in keeping the service secure and up to date, which covers the update check and the protection of your account.

How long we keep it

Your rights

Wherever you are, you can do all of the following. Depending on where you live, some of them are also rights under law, including the GDPR in Europe and the CCPA in California, and we honour them the same way everywhere.

Security

Everything that leaves your computer travels over encrypted connections. Files on your computer are protected by your operating system’s user account, and stored tokens are written in a file only your account can read and never shown to the interface. No system is perfectly secure, and if we learn of a breach affecting your data we will tell you.

International transfers

The providers above operate in the United States and may process data there. Where your law requires safeguards for that transfer, we rely on the providers’ standard contractual terms and their published data protection commitments, linked from the subprocessors page.

Recording other people

A recording contains the speaker’s voice, and often other people’s voices. You are responsible for having permission to record. Many places require the consent of everyone recorded, and most universities separately require the lecturer’s permission regardless of the law. Lectavi asks you to confirm you have permission the first time you turn Record this session on, and cannot verify it. A session that only listens keeps no audio, only the words of each stretch Lectavi explained. From version 0.66, unless you turn off Transcribe sessions as they run, it streams what it hears to Deepgram to be transcribed while it runs, the voices of the people around you included, so where you may not record, or may not send a class’s audio to a transcription service, turn that switch off. See the Terms of Service.

Children

Lectavi is not intended for anybody under 13, and you must be at least 16 in parts of the European Economic Area where that is the minimum age for consenting to data processing. We do not knowingly collect data from anyone younger, and if we learn that we have, we delete it.

This website

Lectavi sets no cookies of its own on lectavi.com and runs no analytics there. The fonts are served from lectavi.com itself, so a page makes no request to Google or to any other font service. The pages are served by Cloudflare, which handles each request the way any web host does. On lectavi.com it can add two small scripts of its own: one that checks whether a visitor is automated, which may keep its answer in a cookie of Cloudflare’s, and one that turns an email address it has hidden from address harvesters back into a link. The installers are downloaded from GitHub.

Changes

When this policy changes, the date at the top changes with it, and a change that matters to what you have already agreed to is announced in the app before it takes effect.

Contact

support@lectavi.com, for questions, requests to see or delete your data, or anything this page did not answer.