Every service that sees your data.
Every company that sees any part of your data, what they see, and why they see it. If a service is not on this list, Lectavi does not send it anything. What each company does with data on its own account is governed by its terms, linked here, and you should read theirs as well as ours.
Services Lectavi uses on your behalf
| Service | What it receives | Why |
|---|---|---|
| Deepgram | Audio: from version 0.66, unless you turn off Transcribe sessions as they run, a session’s audio, streamed live while the session you started runs, and nothing of it stored by Lectavi; with that switch off, and before version 0.66, the stretch of a session Explain that asks about, up to two minutes, when you press it; a whole recording, streamed live while you record; a saved recording when it is transcribed later, streamed from your disk, which the app does not yet offer a way to ask for; and up to four seconds of room audio after “Hey Lectavi” is heard with that switch on. The audio of a question you ask Lectavi out loud, streamed the same way. The course words you add in Settings, under Audio, with a recording’s audio and the stretch Explain that sends. Before version 0.64, and from version 0.64 when Inworld cannot make a reply or, for someone who updated from an earlier version, until they answer the notice in the app about Inworld, the text of Lectavi’s spoken replies, the lines it speaks over a drawing included, which our server passes on | Turning speech into text, and the text of a reply into speech when Inworld does not. This is the only place your audio goes. The listening for “Hey Lectavi” itself happens on your computer and sends nothing. Deepgram’s terms, read on November 2, 2026, let it store parts of what it is sent and use them to improve its service and train its models unless a request opts out, and Lectavi’s requests do not opt out. |
| Inworld | From version 0.64, the text of Lectavi’s spoken replies, the lines it speaks over a drawing included, which our server passes on with the voice you chose. If you updated from an earlier version, only once you answer the notice in the app that tells you about Inworld. The text can quote a few of the speaker’s words and repeat words of a question you asked out loud. Never audio, and nothing that names you: the request carries our key, not yours | Turning the text of a reply into speech. Without Zero Data Retention, which Lectavi’s Inworld account does not have, Inworld keeps that text 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 (read on November 1, 2026). |
| OpenRouter | Transcript text, the questions you type into the chat with the earlier turns of that conversation, the text of a file you attach to one, and the picture itself when you attach an image, sent with a key issued to your account whose name carries its random identifier. For Explain that and the drawings Live visuals makes on its own, the stretch of transcript they answer, and pictures of your screen only when Live visuals, or from version 0.66 Screen sight, is on: with Live visuals, one with Explain that, up to three recent ones with a drawing made on its own, and one more each time Explain that draws the same stretch again, but none for the one try Lectavi makes again by itself, from version 0.66, when an answer could not be used; and from version 0.66, with Screen sight on, one picture of the screen your pointer is on with an Explain that whose stretch holds too few words, or that asks about something on the screen, when you allowed that look, in place of the one Live visuals would take. Live visuals is off until you turn it on; from version 0.65 the welcome, and the notice an update shows, suggest it and Screen sight, and turn them on only when you press Continue there. The course words you add in Settings, under Audio, with Explain that, a drawing Live visuals makes on its own and the notes. A topic’s notes, its explanations and the speaker’s words it was explained from, when you ask for its quiz, and to mark the quiz, each question you got wrong with its right answer and the option you chose. The words of a question you ask out loud, with the topic you are on and, from version 0.65, the last turns of the conversation it is part of, and, in a conversation about a session or a recording, the notes Lectavi wrote for it and the speaker’s words they cite, as a typed question carries them. From version 0.66, while a session you started is transcribed as it runs, the speaker’s words heard in it in the last two minutes, from the words already transcribed, with a question you ask out loud in it that is not looked up. With Screen sight on, one picture of the screen your pointer is on, each time you ask Lectavi to show you something, with, from version 0.66, the last six turns of the conversation that is open and, during a session or a recording, the topic’s title and Lectavi’s latest explanation of it, never the speaker’s words. Screen sight is off until you turn it on. One picture of the screen your pointer is on with a chat question, when you allowed that question to look. From version 0.64: up to six pictures of the screen your pointer is on for one request to scroll to something, with Let Lectavi use the mouse on, and from version 0.65, with Let Lectavi click for me on, one more picture as it makes a click, with the point it would click marked, only when the place looks different with the pointer on it, as a link or a tab may; the words of a question Lectavi looks up on the web, the question before it, today’s date, and the text of the pages the search returned; and, to check an answer, your question and the answer again, with up to two of your earlier questions in the same conversation and Lectavi’s answers to them and what already went with the question, never a picture a second time, and from version 0.66 never for a plain sum answered by that sum alone, which Lectavi works out again on this computer and finds right, with nothing else sent with the question. | Routing to a language model to draw explanations and write the lines Lectavi speaks over them, to answer a question typed or asked out loud, to write notes, claims, flashcards, quizzes and visuals, to mark a quiz, and to work out where to point. From version 0.64 also to find what to scroll to, to search the web and check each sourced line, and to check an answer before you see it. No audio, and no picture of your screen with Live visuals and Screen sight off and no yes given. 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 send a Claude model’s request to Microsoft Azure or to Amazon Web Services, which run the same models. 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; otherwise it keeps only each request’s counts (its data collection page, read on November 1, 2026). |
| The same text via OpenRouter; with Live visuals on, up to three recent pictures of your screen with a drawing Live visuals makes on its own; with Screen sight on, the one picture of the screen your pointer is on that a chat question you allowed to look carries; and the images you attach to a chat. From version 0.64: the sourced lines of a looked up answer with the page text they cite; the answers Anthropic’s model writes, with your question, up to two of your earlier questions in the same conversation and Lectavi’s answers to them, and what already went with it, the words of a question you ask out loud about what Lectavi is explaining included; each drawing Explain that makes, with the words it was drawn from; and, because Google Cloud runs the Claude models too, everything the Anthropic row lists. When you sign in, which is with your Google account, that you are signing in to Lectavi | Signing you in: Google confirms who you are to Lectavi’s sign in service on Supabase. The Gemini models, which write the notes’ sections, the live summaries, the drawings Live visuals makes on its own, the chat answers, the cards and a topic’s quiz, mark a quiz, and read the pictures of your screen those drawings and a chat question carry, and the images you attach. From version 0.64 they also check each sourced line against its page, check the answers Anthropic’s model writes, spoken replies included, and check each drawing Explain that makes, and read again each drawing Live visuals makes on its own, which a Gemini model drew. OpenRouter is a router, and this is the model it currently routes to. From version 0.64 Google Cloud (Vertex AI) also runs the Claude models: Lectavi asks OpenRouter to send each Claude request to Google Cloud or to Anthropic and to no one else, most to Google Cloud first. The Claude model that checks a chat answer runs on Google Cloud first, so the check of an answer Google’s model wrote reaches Google, which wrote it, and reaches Anthropic only when Google’s is unavailable. A Gemini request goes to Google AI Studio first and to Vertex AI when AI Studio is unavailable. Google’s pages, read on November 2, 2026, say Google Cloud does not use what it is sent to train its models, may keep a request its safety classifiers flag for up to 90 days, where its staff may review it, and may hold a Gemini request in a cache for up to 24 hours; and that Google AI Studio’s paid service does not use requests or answers to improve Google’s products, and logs them for a limited time to detect misuse and for disclosures the law requires. Those pages are Google Cloud’s data governance and abuse monitoring pages and the Gemini API terms. | |
| Anthropic | The same text via OpenRouter, the words of a question you ask out loud or type; with Live visuals on, the picture of your screen Explain that takes each time it draws, a drawing made again included; and with Screen sight on, the one picture of your screen a question asks for each time you ask Lectavi to show you something, or the part you circled, and from version 0.66 the one an Explain that with too few words asks for. From version 0.64: the pages a web search returned; the answers Google’s model writes in the chat, with your question, up to two of your earlier questions in the same conversation and Lectavi’s answers to them, and what already went with it; and with Let Lectavi use the mouse on, up to six pictures of the screen your pointer is on for one request to scroll to something, and from version 0.65 one more as it makes a click, with the point it would click marked, only when the place looks different with the pointer on it, as a link or a tab may | The Claude models, which plan the notes over the whole recording, write each Explain that explanation and what its drawing shows, reading the picture it took with Live visuals on, answer a question and work out where to point. From version 0.64 they also write an answer from the pages a web search returned, check the answers Google’s model writes, and find what to scroll to. The app lays every drawing out and draws every mark itself. From version 0.64 a Claude model runs on Anthropic’s own service or on Google Cloud (Vertex AI), and Lectavi asks OpenRouter for no other: most Claude requests go to Google Cloud first and reach Anthropic when Google’s is unavailable. Anthropic’s pages, read on November 2, 2026, say it does not train its models on what its commercial services are sent, and deletes a request and its answer within 30 days, except one its safety screening flags, which it may keep for up to two years, and one the law requires it to keep. Those pages are its model training and retention articles. |
| Exa | From version 0.64, through OpenRouter, a search query written from the words of a question Lectavi looks up, and when you ask for a check, the words of that answer, which can quote a few of the speaker’s words, or, asked out loud while Lectavi is explaining, the summary of its explanation. Never audio, a transcript, a file or a picture of your screen | Searching the web so an answer can show its sources. 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; Lectavi has no zero data retention agreement with Exa, and OpenRouter’s zero data retention setting does not cover web search (read on November 1, 2026). Turning off Look things up online, in Settings, under Privacy, keeps every question from it. |
| Supabase | Your email address, and the name, profile picture and Google account identifier Google sends when you sign in; synced sessions, notes, claims, flashcards, review history, the explanations you save, without the stretch of transcript they explain, every topic Lectavi explained, saved or not, with its title, its drawings, their steps, examples and questions, and Lectavi’s notes, without the speaker’s words it was explained from, the moments you mark, and your conversations with Lectavi without the speaker’s words they quote and without anything of the files you attached, not even their names, from version 0.64 with each answer’s sources and never with how an answer was checked, and from version 0.65 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, unless it looked the question up online; when each transcription and notes key was handed out and how many minutes of audio were transcribed, how many characters of spoken replies were made, and your account’s OpenRouter key, encrypted. The text of a spoken reply passes through on its way to Inworld or Deepgram and is neither stored nor logged | Sign in, making your notes available on another device, and keeping each plan within its monthly allowance. Audio and transcripts are never sent, though a spoken reply’s text can quote a few of the speaker’s words. With sync switched off no notes or cards are sent, and the app still asks it for transcription, notes and spoken replies, which records only the counts and the key above. |
| Stripe | Your email and payment details | Subscriptions. Lectavi never sees your card number. |
| GitHub | Nothing about you | Hosting the installer and the update feed. The app fetches these anonymously. |
| Wikimedia Foundation | A short search phrase for a topic that can be photographed | Finding a public domain, CC0, CC BY or CC BY-SA picture on Wikipedia or Wikimedia Commons to illustrate a topic in the notes of a recording. The picture is downloaded to your computer with its author, licence and source. |
| Cloudflare | Your browser’s requests for this website, which carry your IP address | Serving this website, fonts included, and on lectavi.com two scripts of Cloudflare’s own: one that checks whether a visitor is automated, and one that turns an email address it has hidden from address harvesters back into a link. This website only; the app does not use it. |
What never leaves
- What a session hears. The last two minutes stay in memory on your computer and are never written to disk. From version 0.66, unless you turn off Transcribe sessions as they run, its audio reaches Deepgram as a stream while the session runs and is never written to disk; with that switch off, and before version 0.66, only the stretch Explain that asks about goes, and only when you press it.
- The recording file. The WAV of a session you recorded is never uploaded anywhere and stays on your disk. Its audio reaches Deepgram as a stream: live while you record, and from the saved file when a recording that has no transcript is transcribed later, which the app does not yet offer a way to ask for.
- Transcript segments. The full transcript is never synced.
- “Hey Lectavi”. The listening for the phrase runs on your computer. What the microphone hears while it waits is never written to disk and never sent.
- What you say to Lectavi. A question’s audio reaches Deepgram as a stream and is never saved or synced. The words of a question you ask out loud stay on this computer, with a drawing it asked for and in its conversation, and never sync. From version 0.65 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 stay on this computer too, unless it looked the question up online.
- Pictures of your screen. Never written to disk and never synced. The last three that Live visuals keeps during a recording can go with more than one of its drawings and are gone when the recording stops; every other picture goes in one request and is gone once the answer arrives.
- How an answer was checked. From version 0.64 each answer keeps how it was checked, on this computer only. It is never synced.
- The files you attach to a chat. Read on your computer; only a document’s text goes with your question and each later question in that conversation, and the text is dropped when you close the Chat panel. A picture you attach goes as itself, the same way, and is dropped with the text. Nothing of a file you attach syncs, not even its name: the conversation keeps each file’s name, kind and length on this computer only.
- Your keys and tokens. They are read only by the part of the application that talks to those services, and are never sent anywhere else.
Services earlier versions could connect
Earlier versions could read your Google Calendar and Google Classroom, your Canvas, a phone used as the microphone, an audio or video file you added, and 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. This page lists them again when one of them returns.
Changes to this list
When a service is added or replaced, this page changes and the date at the top changes with it. Questions go to support@lectavi.com.