Drutak — Privacy Policy

Last updated: 2026-10-12

Drutak is a voice assistant app that runs on your phone. "Drutak" is the app you installed, the thing that listens and answers, and the word you say to wake it — one name for all three. This policy explains exactly what leaves your device, who receives it, what is kept, and how to erase it.

If you installed an earlier build, the assistant introduced itself as "Loki". That was the same software under a different name, and saying "Loki" still wakes it — nothing about what is collected, sent or stored changed with the name.

It is written to be read, not to be survived. Where something is uncomfortable — and one thing genuinely is — it is stated plainly rather than buried.

The short version

• There is no account. No sign-up, no login, no email, no password. The one identifier is a random number your phone makes up on its own, so the free plan can count how many questions were asked today (§1.7). It is not linked to your name, number or anything else about you, and we have no profile of you on our servers.
• Your voice commands are recorded and uploaded to be transcribed. This is the one disclosure that matters most, and it is unavoidable in a product that answers when you speak. It is explained in full below.
• Almost everything Drutak knows about you stays on your phone, in the app's own storage, and is deleted when you uninstall.
• We do not sell your data. We do not share it with anyone for advertising. There is no analytics or tracking software in this app — not Firebase Analytics, not Crashlytics, not an advertising SDK, not any third-party tracker.
• Nothing you say is used to train AI models by our AI provider, Groq, which is contractually forbidden from doing so. When Groq cannot answer, or is slow to, your message may go to a backup provider, Hugging Face, whose terms we describe separately in §4.
• One exception while Drutak is being built: a request of several steps. Each step after the first goes to Google's Gemini with a picture of your screen, and on the free tier we use for now Google may use it to improve its products (§4). This moves to Google's paid tier before Drutak is offered to anyone else.

1. What leaves your phone

Everything Drutak sends goes first to a server we operate (a Cloudflare Worker), which passes it to the providers that do the work: the AI providers, a weather service when you ask about the weather, and the reference sources in §1.6 when you ask something they answer. There are no other destinations.

1.1 Audio of your commands

When you speak a command, Drutak records that utterance and uploads it to be converted into text. Recording begins only at the wake word, when you tap Drutak's microphone, press the assistant button on a headset or hold your phone's power button with Drutak as your digital assistant (below), when you talk over Drutak's reply, or for a few seconds right after Drutak answers you.

The moment around the wake word. So you can say "Drutak, open the camera" in one breath, the phone keeps the last second and a half of sound in memory while it listens. When the wake word is heard, the part of that sound that belongs to what you just said — the wake word itself and anything you began saying with it — is sent with your command. Sound from before a pause preceding the wake word, such as a TV or another conversation, is cut off on your phone and never sent. Speech that runs straight into the wake word with no pause cannot be separated from it and is sent with the command.

Interrupting Drutak. While wake-word listening is on, the microphone also stays on while Drutak is speaking a reply, so it can notice you talking over it and stop. During a reply the phone keeps only the last two seconds of sound in memory, each moment overwritten by the next, and erases them when the reply ends unless you talk over it. When you do, Drutak stops and keeps listening: what the microphone hears from about half a second before it stopped, for up to about fifteen seconds, is held in memory. When Drutak listens on after stopping — on your screen or, since 2026-10-03, not — that sound is sent to Groq, via our server, to be turned into text, and the text is treated like words heard in follow-up listening (below); the copy on your phone is erased once it has been sent. A lone "stop", and words Drutak does not listen on for — for example when you say its name again, a call begins, or it has already listened on as many times as one wake allows — are erased on your phone without being sent, within about a minute. This sound is never written to a file. When wake-word listening is off, the microphone is not used while Drutak speaks.

Drutak listens continuously for the wake word using a speech model that runs entirely on your phone. That continuous listening never leaves your device and is never uploaded. Audio is captured and sent only for the command that follows. If you taught Drutak your voice (§3), each time it thinks it heard its name it also compares that sound with what it learned from you, on your phone, to tell your call from a video or a TV that only sounds like its name.

Follow-up listening. After Drutak answers a spoken command, it listens once more for a few seconds without the wake word, so you can keep talking. Whatever the microphone hears in that window is converted to text and sent to the AI model with your conversation — including speech that was not meant for Drutak, such as a TV or other people nearby. The model decides whether it was said to Drutak; if not, Drutak stays silent, the sentence is not added to your chat, and nothing is done with it. When Drutak answered while it was not on your screen, the words in that window are turned into text by Android's own speech recogniser (§2), which is Google's and may work over the network, not by Groq; the text then goes to the AI model the same way. What you said over its answer after it stopped is the one exception: that is turned into text by Groq, as above.

Holding the power button. Since 2026-10-05, on Android 12 and newer, you can make Drutak your phone's digital assistant (§6). Holding the power button — or your phone's own way of calling its assistant, such as a swipe from a bottom corner — then calls Drutak from any app to listen for one request, as its name does, also while wake-word listening is off. With Drutak's own screen in front, it listens there as it does after its name, or, with wake-word listening off, as when you tap its microphone; a press never switches the microphone off, and if Drutak is already listening for your words, what you say next, within 45 seconds, is taken as a request, with its name or without. Otherwise the microphone opens as it does after its name while Drutak is not on your screen, and your words are turned into text the same way, by Android's own speech recogniser (§2). If Drutak is saying an answer or a reminder, it stops and listens. If it is at work on a request, it keeps your press for up to thirty seconds and listens once that request is over, and with its screen not in front it may say it is still on it; a press never stops a request. If a call has your phone's audio, a press does not listen. If Drutak is not set up yet, or you switched its microphone off, a press opens Drutak's screen and nothing more. If its microphone is not allowed, a press opens Drutak's screen and asks you for it with Android's own dialog, and listens once you allow it; if Android will no longer ask, it opens Drutak's page in Android's settings instead, where you can allow it. Only your own press counts: when Android says an app asked for the assistant, Drutak does nothing. Drutak takes nothing from your screen when it is called this way — no screenshot, and nothing of what the screen shows.

The recording is a temporary file, deleted after transcription. Transcription is performed by Groq, via our server.

Please understand what this means: if you say something private in a command, that audio is transmitted. Drutak cannot know in advance which sentences you would rather keep to yourself.

1.2 The text of your conversation

To answer you, Drutak sends the AI model the transcribed command together with context about you. That context includes:

• Your name, which you typed during setup.
• Facts, goals and traits Drutak has inferred about you from past conversations — including short quotes of things you said, which is how Drutak explains why it believes something about you. Since 2026-10-03, a fact can be sent with the value it replaced and the day it was noted, and with the last day it holds when you named one, where its line has room for them; a fact past that day is not sent; and up to two more facts can come with the ones chosen for your request: what Drutak knows about someone they name, or the fact that names whom they are about — asked about your sister, what it knows about her.
• A one-line summary of each of your last few days of conversation, so Drutak can pick up where you left off — for example planning a trip to Goa. You can see and delete each one on the Memory screen.
• The names and relations of contacts you saved in Settings — for example Amma — mother. Phone numbers are never sent. When Drutak places a call, the number is looked up on your phone and passed straight to the dialer.
• The name of the app currently in the foreground and how long you have used it today — for example WhatsApp — 42 min — so Drutak can respond usefully when you ask about your own focus.
• The text on the screen in front of you, but only on a turn where you asked about it, asked Drutak to press, type or open something on it, or named something on it that can be pressed — "what does this say", "read this", "press 7", "type lo-fi in the search box", "open Display & brightness" with Settings open. Drutak reads that window's text at that moment and sends it with your request, because nothing can answer that question, or find the button or box you meant, without it. Since 2026-09-27, when your words name something on that screen that can be pressed — the 7 of "press 7" with the Calculator open — it goes with your first request, whatever you asked: "play some music" with a Play button on the screen sends it too (§6). It requires Assist mode to be on (§6), it never runs in the background, on a request that is about neither your screen nor a press or typing on it it runs only in the cases §6 lists, it never reads a password field, and no copy is kept once you have your answer.
• A picture of the screen in front of you, but only for a step of a request of several steps (since 2026-09-29) — each step after the first of "open YouTube and play some lo-fi" and the like. Your phone takes one picture of the screen for that step, made small, and sends it with that screen's text, so the AI can tell what the text alone reads badly: a video's card, an advertisement, a dialog drawn as a picture. It goes to Google's Gemini and nowhere else (§4), one picture per step, never on a timer; it is never written to your phone's storage and never appears in your gallery, and neither your phone nor our server keeps a copy. Android leaves a window marked secure — banking apps and password screens usually are — out of the picture: it comes out black. No picture is taken while your phone is locked or its screen is off, during a call, or while Drutak itself, a payment app, an app store, or a permission or install screen is in front (since 2026-10-02, §6); that step goes with the screen's text alone. It requires Assist mode, on Android 11 or later (§6).
• One photograph from your camera, but only on a turn where you held the phone up at something and asked about it — "what is this", "read this sign". Drutak takes a single frame at that moment, sends it to be described in words (§1.6, §4), and answers your question from that description. The photograph is never written to your phone's storage and never appears in your gallery. It requires the camera permission (§6), it never happens while the app is closed or the screen is off, and it never happens unless you asked for something that cannot be answered without it.
• More than one photograph, but only if you asked Drutak to keep looking — "tell me when the kettle boils", "tell me when the light turns green". Drutak then takes one frame every couple of minutes, has it described in the same way as above (§1.6, §4), and asks itself whether the thing you described has happened yet. What is sent with each description is your own words for what to watch for, and nothing else about you — no memories, no contacts, no other part of the conversation. It stops the moment it can tell you, and it stops on its own after at most twenty looks or half an hour, whichever comes first, and says that it has stopped. It only ever looks while Drutak is open on your screen: never in the background, never with the screen off, and closing the app ends it, so a watch never survives being closed and reopened. There is only ever one at a time, and saying "stop" ends it.
• What your phone is playing, but only on a turn where you asked about it — "what song is this?", "who sings it?", "which film is this from?", in any language. Drutak reads, at that moment, what the app playing it has told Android about it: the title, the artist, the album, the app's name, whether it is playing, paused, loading or stopped, and how far into it you are and how long it is. It sends those with your question, so the answer can name it in your language and say what the app does not — which film a song is from. It never sends the artwork, the rest of the queue, or which app package it was. It requires Notification access to be on (§6), it never runs in the background, and it is never read as one step of a longer request (since 2026-09-28).
• How many photos and videos you have, and from when, but only on a turn where you asked about them — "how many photos did I take today?", "when did I last take a video?", "what's in my Goa album?", in any language (since 2026-10-04). Drutak reads, at that moment, the list your phone keeps of your photos and videos, and from it only how many there are, the names of the albums (the folders) they are in, the days they were taken on, and, for up to fifteen of them (the newest, the oldest, and the newest in each of the largest albums), when each was taken and which album it is in. It sends those with your question, so the answer can count them in your language. Drutak itself never opens a photo or a video: no picture, no file's name or path, no caption, and nothing of what is in one is read or sent, so it cannot find a photo by what it shows. When you ask it to open one of those it listed ("open the second one", since 2026-10-04), your phone's own gallery app shows it: Drutak hands the gallery that one item, with leave to read that item only. Which items it listed is kept in your phone's memory for at most ten minutes for that, never written down and never sent. It requires photo access (§6), it is read only at the moment you ask, and never while your phone is locked or its screen is off.
• What is on your calendar, but only on a turn where you asked about it — "what's on tomorrow?", "when is my dentist appointment?", "am I free on Friday?", in any language (since 2026-10-05). Drutak reads, at that moment, the events your phone's calendar lists for the days you asked about — or, when you named no day, those still to come in the next thirty days, or in the next year when you named a word in an event's title — and from each only its title, when it starts and ends, whether it fills whole days, where it is, and whether you said no to it. It sends those for up to twenty events, with how many there were, with your question, so the answer can tell you in your language. When where an event is is a link, such as a video call's, what follows a "?" or "#" in it, where a meeting's password usually is, is left out. Who is invited to an event or organised it, its notes, its reminders, and which account or calendar it is in are never read or sent. Drutak only reads your calendar: it never adds, changes or deletes an event itself (§2). It requires calendar access (§6), it is read only at the moment you ask, and never while your phone is locked or its screen is off.
• The current date and time, and your language setting.
• The language your voice was heard in, and when speech recognition was unsure, up to four other ways it heard the same sentence, so a mishearing can be undone.
• Your battery level and whether it is charging, whether you are on Wi-Fi, mobile data or offline, and whether the phone is running hot — so Drutak can say when something will not work before it starts. The kind of connection only: never the network's name, its address, or anything that passes over it.
• When your next alarm rings, whether your phone is set to ring, vibrate or silent and whether Do Not Disturb is on, and how much storage is free — so Drutak can answer "when is my alarm?", "am I on silent?" or "do I have room for this?". The alarm's time only: never its label or which app set it (since 2026-09-28).
• What Drutak is watching your phone for, and until when, while it is running and watching for something you asked it to tell you about (§3) — in a few fixed words, such as "the battery at 80 percent or more", and the minute it stops watching; never the sentence it will say then (since 2026-10-03). With a setting you asked it to change then (§3), that setting and its value too, in the same few fixed words, such as "dnd on".
• The weather Drutak already shows you, when it has it — the reading, not your location (§1.5).
• The items on your plan (reminders and tasks you asked Drutak to keep), so it can answer "what's next?".
• The names of the routines you saved, when you have saved any — for example good night — so Drutak knows one when you say it, even inside a longer sentence. The names only: the steps each one stands for stay on your phone (§3). Since 2026-09-28.
• The last messages of this conversation, so "and tomorrow?" means something: up to three before what you said, or up to nineteen before what you typed — never an answer that read out what came in on your phone (§6). Since 2026-10-02 they include what you said to Drutak while it was not on your screen and what it said back (below), which Drutak's screen shows as part of the one conversation (§3), and since 2026-10-03 what it said out loud then with nothing asked of it (§3).

When you talk to Drutak while it is not on your screen — the app closed, another app in front of you, or, with the wake word on, the screen off — it sends less than the list above. Only: your name, your language setting, the current date and time, your battery level, whether you are on Wi-Fi, mobile data or offline, whether the phone is running hot, when your next alarm rings, whether the phone is set to ring, vibrate or silent and whether Do Not Disturb is on, how much storage is free (since 2026-09-28), the names of the routines you saved — their names only, never their steps (since 2026-09-28), the name of the app in front of you — its name only, and only while the phone is unlocked with its screen on and Assist mode is on (§6) — a copy of what it remembers about you (up to five facts, since 2026-10-03 the five that share the most words with what you asked, chosen on the phone from a copy of your twenty newest kept in the app's private storage; three goals, two day summaries and two traits, without the quotes), since 2026-10-03 up to five of the things you asked it to remember while it was not on your screen and its screen has not taken in yet, each with when you said it (those that share the most words with what you asked; never one whose answer had read your screen or looked something up, and none while What I keep is off, below), since 2026-10-03 the request of several steps you made before this one, while it is still being carried out or waits for your answer on your phone's own screen — your words for it, when you made it, where it stands, how many steps it has taken, how many of them failed and how many of its actions Drutak's own action journal could not say happened, and when it must be done by — or, with none, how the last one ended, if it ended in the last thirty minutes on the same start of your phone: when and how it ended, and those counts, never your words for it, which its record no longer keeps (§3); never the ids either goes by, since 2026-10-03, while Drutak is watching your phone for something you asked it to (§3), what it watches for and any setting it will change then, in a few fixed words, and the minute it stops watching — never the sentence it will say then, up to three earlier messages from the same conversation (since 2026-10-02 the one conversation you have with Drutak, on its screen or not, since 2026-10-03 with what Drutak said with nothing asked of it among them, and none from before a pause of more than five minutes, §3), up to four other ways your phone's speech recogniser heard the same sentence, so the AI can tell from your screen or the conversation which you meant (since 2026-09-28), and, when a question needs something current, what Drutak looked up for it (§1.5, §1.6) — including what your phone is playing, or how many photos and videos you have and from when (since 2026-10-04), or what is on your calendar (since 2026-10-05), when that is what you asked (§6), and, if your question pointed at your screen instead of naming what it was about, or asked Drutak to press or open something on it, the text on your screen at that moment (§6). And when your words, or one of those other ways of hearing them, name something on your screen that can be pressed, the text of that screen goes with your first request too, whatever you asked: your phone first looks, on the phone only and sending nothing, at what on that screen can be pressed, and reads and sends the screen only when your words or one of those name one of those things (§6). When you ask for more than one step at once, each step after the first is asked with the screen as the step before left it and, for each step so far, your phone's own words for what it did — "went back", "scrolled down", "opened Settings", "pressed one line of the screen they were on", "typed into one box of the screen they were on, and pressed nothing" or, since 2026-09-30, "…and pressed its Enter", "…and could not press its Enter" or "…then pressed its Enter if it had one" — and never the words on what it pressed or what it typed (§6). Since 2026-10-01 each of those steps also carries your phone's fixed sentences for what it found after the step before: that the screen did not move, what Android shows playing — whether something new plays, never what — that the result was already there before, or that the screen already shows what you asked; never what it looked for (§6). Since 2026-09-29 each of those steps — and a step picked up again after Android closed Drutak (below) — also sends a small picture of that screen, to Google's Gemini and nowhere else (§4, §6). When the AI's first answer only opens an app, that app's name goes once more, with your words and nothing else, in a one-word question: whether you asked for more than opening it (§6). When you ask for more than one step, your words alone go once more, inside fixed words, in a question about what your screen will show once it is all done, so your phone can look for it there (§6). Since 2026-10-01 that question also asks which of four kinds of thing your request does — sends or calls, deletes, pays or grants — used only to ask you before a press the screen chose and you did not (§6). The same question is asked beside your first request when your screen went with it, and beside the request that carries your screen when the AI asked to read it first — even when the request turns out to be a single press, or none — so that a first press for something to play is followed by a look at what plays, and a request that goes on has its answer already (§6). Since 2026-10-01 a first press is no longer checked against it, but for one case since 2026-10-02: when its answer names only switches or tabs, and the screen that press was chosen from already shows them so, the press is not made, and the AI is asked once more, with your screen, your same words and your phone's fixed sentence that the screen already showed what you asked, never what it looked for (§6). Beside your first request, it is asked from your words as the recogniser first wrote them or, when those name nothing on your screen and one of its other ways of hearing them does, from that one (since 2026-09-28). When its answer names no app to look in, it is asked once more, your same words in the same question (§6). If Android closes Drutak in the middle of such a request and Drutak picks it up again when it restarts (§6), your words go again: in the question about the next step, with your screen as it is then and your phone's words for the steps already taken — "opened an app", "pressed one line of the screen they were on", "went back, scrolled or opened a panel once" — written from a record that keeps only what kind each step was (§3); and, if the AI had not yet said what your screen will show once it is all done, in that question too. It does not send your contacts, how long you have used any app, your location or your plan. Until 2026-09-23 it did not send the app's name either; it does now because without it, "press 7" with the Calculator open reached an AI that did not know a screen was there.

Since 2026-09-30, a request you make on Drutak's own screen can go on this way too. When the AI's answer to it opens an app, searches the web or YouTube, or plays music, and Voice replies is on, your phone hands your words, and nothing else, to the part of Drutak that listens while it is not on your screen. That part asks the AI again with your words, sending only what the paragraph above lists — its earlier messages are those of the one conversation (§3), never the answer that read out what came in on your phone (§6), and with no app's name the first time, since Drutak itself is in front — then opens the app and goes on step by step, as above, until the request is done, speaking for itself. So your words go to the AI twice: once from Drutak's screen, with the list above, and once more this way. "Open Chrome and search for Bahubali songs" opened Chrome and stopped there, because Drutak's screen carries out one action and ends. With Voice replies off, or with that part busy with a request of its own, nothing is handed over, and Drutak's screen carries out the one action itself, as before — except, since 2026-10-02, while that part is still carrying out a request of several steps: then Drutak's screen opens nothing over it and says the last request is still going. One only waiting for your answer on the phone's own screen, or one left on file when Android closed Drutak (§6), gives way to your newer request instead, and ends without a word. Until 2026-10-02 your words went with up to three earlier messages from Drutak's screen, which took the place of that part's own, and the wake word had to be on as well. Since then, with the wake word off, or with Drutak not allowed the microphone, that part is started for your request alone: it opens no microphone, listens neither for its name nor for anything you say next, and stops once the request is over (since 2026-10-03, or, with a watch over your phone you asked for on file, once that watch ends, §3), while Android shows it at work on a request (§6). If it cannot start within three seconds, Drutak's screen carries out the one action itself, as before. Since 2026-10-03, a watch over your phone you ask for on Drutak's screen (§3) is handed to that part too, because only that part looks at your phone again later: what it watches for and the sentence the AI wrote for Drutak to say then, never your words. With the wake word off, or with Drutak not allowed the microphone, that part is started for the watch alone, the same way, and opens no microphone; if it cannot start within three seconds, nothing is kept, and Drutak says so.

Things you ask Drutak to remember while it is not on your screen. Since 2026-10-03, when you ask Drutak to remember something for later with the app closed — "remember that the spare key is with Sita" — the AI's answer names what to keep, in your words, and Drutak writes those words to one small file in the app's private storage (§3) before it answers. It holds at most twenty: when it is full, Drutak keeps nothing more and tells you so. Until its screen takes them in, up to five of them, with when you said each, go with every question you ask Drutak while it is not on your screen (above), so the AI can answer from them; never words from an answer that had read your screen or looked something up. The next time Drutak's screen is open, they go to the AI once more, on their own, with your name, the day you said them and the facts and goals Drutak already keeps about you, the same as after a conversation on its screen, and what the AI finds in them is kept in Memory. If it finds nothing to keep, or not all of it could be kept, your words are kept as well, word for word, as a note (§3). Words from an answer that had read your screen or looked something up may be the screen's or the search's and not yours: they are kept only as a note and never go to the AI again. When no AI can be reached, the words wait in the file for a later try, and after a day of waiting they are kept as a note instead. Each one is deleted from the file once it is kept. While What I keep on the Control screen is off (Stop keeping new things), Drutak writes nothing new to that file and tells you it is not keeping new things, and anything already in it waits there, sent nowhere, until you turn it back on; the switch is copied to the part of Drutak that listens in the background for this.

Checks you set for a time. Since 2026-10-05, when you ask Drutak to tell you something at a set time, every day, on some days or once — "every morning at 7, tell me my day", "every evening at 9, remind me to take my pills" — on its screen or not, the AI's answer names the time, the days and your request in your words, and Drutak keeps them in one small file in the app's private storage (§3) before it answers. At that time an exact alarm (§6) wakes the part of Drutak that runs while its screen is closed, started for it when it is not running, with no microphone opened. Once your phone is unlocked and its screen is on, that part sends your request, in your words, to the AI as a question of its own, with your name, the time and date, your language and the facts Drutak keeps about you. The AI may ask for up to two live lookups — the weather, the time somewhere, an exchange rate, a web search, Wikipedia or your own calendar (§1.6, §6) — and what they found goes back to the AI for its answer. It is never allowed to do anything on your phone: no call, no message, no press, nothing opened or changed. Drutak says the answer out loud once. When the AI cannot be reached or gives no answer, it says your own request back to you instead. While your phone is locked or its screen is off, nothing is sent: Drutak shows "Unlock to hear your update" and waits, for at most three hours, then lets that time go. It never says a check during a call, while you are talking to it, or over something else it is saying. While checks are on file, each one's time, days, next time and your request go with what you ask Drutak, on its screen or not, so you can ask about them or stop one (§1.2).

Drutak also talks to the AI model when you have not asked it anything. It is allowed to speak first in three fixed windows a day — morning (6am–12pm), evening (7pm–10pm) and late night (11pm–5am) — and it asks the model to word what it would say. Most of those asks come back "say nothing", which is the intended answer. The limits are fixed in the app and the model cannot widen them: at most three unprompted messages a day, and at most eighteen such asks a day. Since 2026-10-03 those limits are one count whether Drutak is on your screen or not: its screen and the part of Drutak that listens in the background keep them in the same small file on your phone (§3), and an ask is written there before it is made. Drutak never speaks first while your phone is on silent, on vibrate or on Do Not Disturb. If it speaks first from its own screen, the message may also arrive as a notification. Since 2026-10-03 the sentence may end by asking whether Drutak should do one thing on your phone about something you told it — set a reminder, an alarm or a timer, or change a setting. Nothing is done unless you answer, and your answer is a request like any other.

What it sends in that case is less than the list above, not more. Only: your name, your language setting, the current date and time, the name of the app in the foreground, up to six facts you told Drutak, and up to three things you told it were coming up. It does not send the traits Drutak has inferred about you, your day summaries, your contacts, your battery level, your weather or your plan — an unprompted message is built only from the clock and from things you actually said.

With Drutak not on your screen, since 2026-10-03, the part of Drutak that listens in the background asks the AI model in those same windows, under those same limits, while it runs with the wake word on. It sends less again: your name, your language setting, the current date and time and up to six facts you told Drutak — not the app in the foreground and nothing coming up. It asks only when a sentence would cut across nothing: never while your screen is off or locked, while something is playing, while the microphone is switched off, during a call, or while Drutak is answering you, saying something else or waiting for your answer — and it looks again once the answer is back. It says the sentence out loud once and posts no notification; of it, only which window it spoke in is written to the file above (§3), and since 2026-10-03 the sentence itself is held in memory only, with the one conversation (§3), so your answer to it follows on from it.

Drutak can also say one thing out loud without asking the model anything, and because it does not use the internet at all it is not counted in the three above. There are exactly three of these sentences and they are written into the app, word for word, in each language: your battery has gone under 10% with nothing charging, your phone is running hot, or your phone is dangerously hot and may shut itself down. Nothing is sent anywhere — no model, no server, no network request of any kind. The phone reads its own battery and its own temperature, decides on the spot, and speaks one of three fixed lines. It says each one once per occasion and never repeats a line within 15 minutes, it stays quiet while the app is open in front of you, while the microphone is switched off, and while it is in the middle of answering you. Since 2026-10-03 the line it said is held in memory with the one conversation (§3), as its answers are, so it is among the earlier messages that go to the AI with what you ask after it.

These are said with the app closed too, and no one having spoken — which is the point of them: the phone that is about to shut itself down is exactly the phone whose app is not open.

You can switch both off in Settings, under "Speaking first". With it off, no request is made at all, with the app open or closed — Drutak does not ask the model and then stay quiet; nothing about you leaves the phone on this path, and the three spoken device warnings above are silent too.

When Groq cannot answer, or is slow. If Groq is unreachable, has hit its rate limit, fails on its side, or runs out of time, our server may send the same request — the same text, nothing added — to a backup provider, Hugging Face, so you still get an answer. Hugging Face passes it to one of the companies that run models for it (§4). Since 2026-10-01 it is also sent there when Groq has not answered within four seconds: then both have it, the first answer is the one you get, and the other request is cancelled, though by then it may already have been received. It is never sent on to Hugging Face after Groq refused it because something was wrong with the request itself; if Groq took more than four seconds to refuse it that way, Hugging Face was already asked, and that request is cancelled.

1.3 Your IP address

Making any internet request reveals your IP address to the server receiving it. Our server sees yours. It is used for one purpose — limiting how many requests a single address can make, to stop abuse — and it is stored for 120 seconds and then discarded automatically. It is not logged, not retained, and not linked to anything about you.

Because requests reach the AI providers through our server, the providers see our server's address, not yours.

1.4 The voice Drutak speaks in

Drutak can speak in one of two voices, and which one you hear decides whether anything leaves the phone.

The phone's own voice. The text-to-speech engine built into Android. Drutak sends nothing anywhere for it — not to us, not to anyone. This is what you hear whenever the other voice is unavailable, including with no network, and it is the only voice used for a language the better one cannot read, and for what came in on your phone (§6). One thing to know about the engine itself: some engines also offer voices that are made online, and Android marks those as needing the network. If the voice your engine uses for a language is one of them, it is the engine, under its maker's terms, that sends the sentence, as it would for any app on the phone that speaks.

The better voice, through our server. When it is available, the sentence Drutak is about to say is sent to our server, which asks a speech provider to record it and sends the recording back to your phone to play. What travels is Drutak's own reply, together with the language it is in. Your recorded voice is never sent for this, and neither is your question.

Be aware of what a reply can contain: Drutak says things like "Calling Amma" or "I couldn't reach Amma", so a name Drutak speaks aloud is a name the speech provider receives. If that matters to you, Drutak sends nothing for the phone's own voice, and you can stay on it by leaving the premium voice off.

The recording is played and discarded. We do not keep it, and we do not keep the sentence. The provider is Google Cloud Text-to-Speech, or Groq's PlayAI voice when Google is not configured — both listed below.

Settings also contains a "Premium voice" preview button, in the Developer section and only once you have switched developer tools on there. Pressing it sends one fixed demonstration sentence, which is not your words and says nothing about you.

1.5 Weather

When you ask about the weather, Drutak looks it up live through our server, which asks Open-Meteo, a weather service. What is sent depends on how you ask:

• You name a place ("weather in Delhi"): only that place name is sent.
• You mean where you are ("is it going to rain?"): the first time, Android asks whether Drutak may use your approximate location. If you allow it, Android gives Drutak a position blurred to about 3 km; Drutak rounds it further, to about 11 km, and sends only that rounded position. It is not stored, on your phone or on our server, and it is used for the weather and nothing else. If you say no, Drutak asks which city you mean instead.

Drutak also looks the weather up without being asked, so the home screen can show you the sky when you open the app. This is the same lookup, through the same server, and it is subject to two limits. It never raises the location dialog: if you have not already allowed approximate location, it simply asks for no position and the home screen shows no weather. And it does not run on every glance — one reading is kept in memory for 30 minutes, so opening the app repeatedly costs one lookup. Nothing about it is written to disk.

Drutak never asks for your precise (GPS) location, and never reads your location in the background — including for this: the home screen looks the weather up while you are using the app, never while it is closed. Ask about the weather where you are while the app is closed, and Drutak asks which city you mean. Open-Meteo sees our server's address, not yours, and receives no information about you — only a place name or a rounded position.

1.6 Looking things up

Some questions need something current: the time somewhere else, an exchange rate, or a fact that may have changed. Drutak looks these up live through our server, and each source receives only what is needed to answer — never your location, your name or your conversation:

• The time in another place ("what time is it in Tokyo?"): only the place name is sent, to Open-Meteo's place search, to find that place's time zone. The clock is read on our server.
• Exchange rates ("100 dollars in rupees"): only the two currency codes and the amount are sent, to Frankfurter, a free service publishing the European Central Bank's daily reference rates.
• Facts ("who is the president of Brazil?"): only a short topic is sent, to Wikipedia (the Wikimedia Foundation), in your language first and in English if nothing is found.
• Something current ("who won the match today?", news, prices) (since 2026-09-28): only a few search words, which the AI writes from your question (at most 200 characters), are sent to Groq, whose built-in web search passes them to its search partner, Exa, and reads the pages it finds. Groq sums up what they say, with the names of the sites, and Drutak answers you from that. Nothing else goes with the words — not your location, not your conversation. If the search cannot be done, Drutak may ask Wikipedia with the same words instead, as above.
• A hard problem (a plan with many parts, weighing options, maths past a quick sum) (since 2026-09-28): the problem, as the AI writes it out from your question with the details you gave (at most 1,500 characters), is sent to a second Groq model that works it through, and Drutak answers you from its answer. The AI writing it has your conversation (§1.2), so the problem can carry anything from it that the problem needs; it goes to the same provider, under the same terms (§4), and nowhere else.
• What your camera sees ("what is this?", asked with the phone held up at something): this is the one lookup that is not words. One photograph (§1.2) is sent to a vision model — Groq, or Hugging Face if Groq cannot answer or is slow to (§4) — with a fixed English question that asks only for a description of the image. Your own question is not sent with it, and neither is anything else about you; the description comes back and Drutak answers you from it, in your language.

Each source sees our server's address, not yours. Nothing about these lookups is stored, and answers built from them are never saved as recent answers, because they change.

1.7 Counting free questions

The free plan includes a set number of AI answers a day. To count them, the first time Drutak runs your phone makes up a random install number. It is not derived from your phone, your name, your number or any account, and it is not shared with any app or advertiser. Each question you ask gets a second random number, so a question that needs several steps is counted once. A request Drutak's screen hands on to the part that listens in the background (§1.2, since 2026-09-30) gets a new one there, so it is counted twice.

When you ask Drutak something that needs the AI, these two numbers go to our server with the question. Our server keeps only the install number and the question numbers for that day — never what you asked — and deletes them automatically after two days. They are not passed to Groq or anyone else. Things Drutak does on the phone itself, like calling, opening apps or setting alarms, are not counted, and neither is anything Drutak does that you did not ask for.

Delete all my data (§5) throws the install number away, and a new one is made.

1.8 Reports you send

Only if you choose Report this reply on something Drutak said. The report — that reply and the message of yours that produced it (§10 lists exactly what it holds) — goes to our server, which keeps it for 90 days and then deletes it automatically. It is not passed to Groq or anyone else.

1.9 Checking for a new version

Since 2026-10-05. Drutak comes from our website, not from an app store, so no store tells your phone when a newer Drutak is out. Instead, about once a day by your phone's clock, while Drutak's screen is open, it asks our server which version is the newest on the website; a question that gets no answer is asked again the next time Drutak's screen comes to the front. The question carries no words, no install number, no question number and nothing Drutak keeps about you. Like any internet request it shows our server your IP address (§1.3), which our server does not store for it, and our server keeps nothing about the question. Its answer is the same for everyone who asks: the newest version's number and name, and the address of the page to download it from. That page is our server's too: it loads nothing from anywhere else, and our server keeps nothing about a visit to it.

When that version is newer than yours, the gear on Drutak's home screen wears a dot and the Control screen offers New version. Only when you tap it does your phone open that page, as it opens any link. Drutak downloads nothing and installs nothing itself: you download the new version in your browser, and Android asks you before it installs it, and refuses an update not signed with the same key as the Drutak you have. What Drutak keeps of the answer is in §3.

2. What never leaves your phone

Stated as specifically as the list above, because the absences matter:

• Your precise location. Drutak never requests precise (GPS) or background location. The only location it can use is the approximate one described in §1.5, for the weather, and only if you allow it.
• Your messages. Drutak can compose a WhatsApp message or SMS and hand it to that app to send — and, with Assist mode on, press WhatsApp's Send for you (§6). It cannot open your conversations, and requests no SMS permission of any kind. If you switch Notification access on (§6), Drutak sees message notifications the same way your notification shade does — in memory, while the app is running. None of that text is kept on disk (§6), sent to us, sent to the language model, or turned into something Drutak remembers. Since 2026-10-05, when you ask, Drutak can also answer such a message through its notification's own reply button, as a smartwatch does: your reply goes to the app that posted it, to send as it would from your shade, and the message itself is not read for that (§6, Answering a message).
• Your photos and videos themselves, or your files. Since 2026-10-04 Drutak asks for photo access, to count your photos and videos when you ask about them (§1.2, §6), and it reads only the list your phone keeps of them: how many, which albums, which days. It never reads a picture, a video, a file's name or path, or a caption. It never reads where a photo was taken, which needs a permission it does not request. It never writes to your photos, and it requests no permission to your other files. Asked to open one it listed, it hands that one to your phone's gallery app to show (§1.2).
• Your calendar's guests, notes and reminders. Since 2026-10-05 Drutak asks for calendar access, to tell you what is on your calendar when you ask (§1.2, §6), and from each event it reads only its title, its times, where it is, and whether you said no to it. It never reads who is invited to an event or organised it, its notes, its reminders, or which account or calendar it is in. Drutak never writes to your calendar, and it requests no permission to: an event you ask it to add is added by your calendar app, on that app's own new-event screen, which Drutak opens with the event filled in. Reminders you give Drutak are Drutak's own and stay on the device.
• Your contacts list. Drutak does not upload your address book, and never sends it to us. At the moment you ask it to call someone, it searches your contacts for that one name on the device and uses the number locally. One exception, and it is a real one: so that Drutak can hear names correctly — especially names that are not English words — it can tell the transcriber which names to expect. Only contacts that have a phone number, never phone numbers, never any other field.

It happens in two places:

• Android's own speech recogniser, which is Google's and works over the network. When Drutak asks which of several people you meant, the recogniser is given the names it just offered you as words to expect, and no others. Every other time it listens, on your screen or not, it is given only Drutak's own name: since 2026-09-28 while Drutak is not on your screen, and since 2026-10-03 while it is. Until then, each command Drutak listened for while it was open on your screen gave the recogniser the names of up to 600 of your callable contacts; with your contacts' names in that list, "press 7" came back as a person's name. Google's handling is governed by their terms, not ours.
• Groq, through our server — the same service that already receives your command audio (§1.2, §5). Groq is sent names only when you ask to call someone and the name Drutak heard does not match anyone clearly: the same recording is transcribed once more with the names in your contacts that sound closest to what it heard, capped at 300 characters (about two dozen names). Ordinary speech is sent to Groq with no names.

If you would rather no name left your phone, turn the Contacts permission off in Android settings: Drutak will still call people you name, and will simply be worse at hearing which one you meant.

• Your phone number, or anyone else's.
• The list of apps installed on your phone. Drutak reads it to understand "open WhatsApp", and asks your phone which of your apps can take a payment, place a call, send a message or grant a permission, so that it knows which presses could do those things (§6, Assist mode). It answers both entirely on the device. From 2026-09-28 until 2026-10-04 it also listed the apps that have an icon on your phone below a Final check switch in Drutak's Settings, on your phone only, so that you could set one apart (§3, §6); that switch is gone, and that list with it.
• Your web searches. "Search for X" opens your own browser. We see nothing of it.
• Diagnostic logs and crash reports. Drutak keeps a detailed diagnostic log so you can see what it did and why. It is written to the app's private storage, shown to you inside the app, and never transmitted to us or anyone else. Drutak's action journal and the call trace (§3) are kept the same way, but no screen shows them.

3. What Drutak remembers, and where

Drutak builds an understanding of you over time. All of it is stored only on your device, in the app's private storage:

Your conversation transcript is not saved. It is held in memory and never written to storage: what Drutak's screen shows while it is open and, since 2026-10-02, the last twenty messages of the one conversation you have with Drutak, on its screen or not, for as long as Drutak is running — so its screen shows what you said with the app closed when you open it. Since 2026-10-03 they include what Drutak said with nothing asked of it, on its screen or not — a sentence in one of its windows, a warning about your phone, a reminder come due, or what a watch you asked for said as it ended (§1.1, §1.2) — so your answer to it follows on from it. When Drutak stops running, the conversation is gone. What persists is only what was extracted from it — the table above, including the one-line daily summary, which is a summary and never the conversation itself. The one exception is a request of several steps that Drutak is still carrying out: its words are in the table above, and are deleted with the rest of that row.

With the wake word on, Drutak keeps one more copy, in the same private storage: the part of its memory it sends when the app is closed (§1.2), so it can answer you then. It is rewritten while you use the app, deleted when you turn the wake word off, and deleted by Delete all my data (§5). The messages of the one conversation (above) go to the AI with a question you ask while Drutak is not on your screen up to three at a time, and none of them goes to the AI again once you have stopped talking for more than five minutes (§1.2). They are never written to storage — apart from the words of a request of several steps while it is carried out (the table above) — and are erased by Delete all my data (§5) and by Clear chat history in Settings. An answer that read out what came in on your phone (§6) is never among them.

None of this is stored on our servers. We could not show it to you if you asked, because we do not have it. Only two things are kept on our server. The day's question count (§1.7): an install number and question numbers, gone after two days. And each report you send (§10): the reply you reported and your message that produced it, kept for 90 days and then deleted automatically.

4. Who receives your data, and what they may do with it

We use these providers. Except where a section below says otherwise, each acts as service providers processing data on our instructions — neither is free to use your data for its own purposes. Hugging Face and, for now, Google's Gemini are the exceptions, and their sections say why.

Groq — transcription and AI responses

Groq receives your command audio and your conversation text, and — when Drutak speaks first rather than answering you — the short unprompted prompt described in §1.2. It is also the first to be shown a photograph you asked Drutak to look at (§1.6), with the fixed description question and none of your words, and it receives the search words of a web search and the problem of a hard question (§1.6). Its web search is run with its search partner, Exa, which receives those search words from Groq; what Exa may keep is governed by Groq's agreement with it, which we have not reviewed, so we make no claim about it. Their Services Agreement binds them, in their words:

"For clarity, Groq is not permitted to use Inputs or Outputs for training or fine-tuning any AI Model Services or other models, unless explicitly granted permission or instructed by Customer."

We have not granted that permission and will not.

On retention, Groq's documentation states that by default it does not retain customer data for inference requests. It may temporarily log requests when investigating platform reliability problems or suspected abuse, and such logs are kept for up to 30 days.

Hugging Face — backup AI responses

Receives your conversation text only when Groq could not answer it, or had not answered within four seconds (§1.2) — the same request Groq received, from our server, so it sees our server's address and not yours. The same applies to a photograph you asked Drutak to look at (§1.6): Hugging Face is shown it only when Groq could not describe it, or had not within four seconds. Hugging Face does not run every model itself: it forwards the request to one of its inference partners, which does.

What Hugging Face and its partners may do with that text is governed by their terms, and we have not yet confirmed those terms in the way we have for Groq above. Until we have, we do not claim that they are forbidden from training on it or that they keep nothing, and you should not assume it.

Google — Gemini, the AI that chooses a request's next step

Receives, for each step after the first of a request of several steps (§6, Assist mode), a picture of your screen and the same text Groq would have received for that step — your request, the screen's text and the conversation so far — from our server, so it sees our server's address and not yours.

We use Gemini's free tier while Drutak is being built, and on that tier Google may use what it receives to improve its products, and people at Google may read it. Do not ask Drutak to carry out a request of several steps on a screen you would not show Google. Before Drutak is offered to anyone else, this moves to Google's paid tier, and this section will quote its terms the way Groq's are quoted above.

Google Cloud Text-to-Speech

Receives the sentence Drutak speaks aloud, and the language it is in, so it can record it in a voice (§1.4) — which includes any name Drutak says in that sentence. Never your recorded voice and never your question. Also receives the fixed preview sentence when you press the preview button.

Groq — PlayAI voice

Receives the same thing as above, and only when Google is not configured (§1.4). It reads English; a reply in a language it cannot read is spoken by the phone's own voice instead and is never sent.

Open-Meteo — weather

Receives a place name or a rounded position (§1.5), from our server, never from your phone directly — both when you ask about the weather and when the home screen shows it to you unasked while the app is open. No information about you is sent to it. It also receives a place name when you ask the time somewhere else (§1.6).

Frankfurter — exchange rates

Receives two currency codes and an amount (§1.6), from our server. No information about you is sent to it.

Wikipedia (Wikimedia Foundation) — facts

Receives a short topic (§1.6), from our server. No information about you is sent to it.

Cloudflare

Operates the server that relays these requests, and stores the only two things that server keeps (§3): the question count and reports you send. It processes traffic and stores them on our behalf.

We share your data with no one else. Not with advertisers, not with data brokers, not with analytics companies. There is no such recipient because there is no such software in the app.

5. Deleting your data

Because your data lives on your phone, you control it directly.

Inside the app: Settings → Delete all my data. This erases everything in §3 — memory, traits, plans, reminders (including any still waiting to go off, and, since 2026-10-03, the words kept to say one after a restart), cached answers, settings, your voiceprint, reports not yet sent, the record of a request still being carried out and of how the last one ended, the list of what Drutak did (§3), since 2026-10-03 the record of when Drutak last spoke first (§3), since 2026-10-05 the record of when Drutak last looked for a new version (§3), your routines, your notes, what you asked Drutak to remember with the app closed that its screen has not taken in yet, since 2026-10-03 the watch over your phone you asked Drutak for, since 2026-10-05 the checks you set for a time, diagnostic logs, and, since 2026-10-04, the call trace and Drutak's action journal, apart from an action too recent to delete (§3) — and drops the notifications Drutak was holding in memory (§6) and, since 2026-10-02, the conversation it was holding in memory (§3).

Uninstalling Drutak removes all of it as well. Android deletes the app's private storage when the app is removed.

Reports you already sent are the one thing neither of these reaches. Nothing in a report says who sent it — no name, account, install number or device identifier — so nothing links it to you or to your phone. Our server deletes each one automatically 90 days after it arrives (§10).

After uninstalling, we hold no account and no copy of anything in §3. The question count in §1.7 deletes itself within two days, and each report you sent 90 days after it arrived. If you want written confirmation of that, or wish to ask about a report you sent or the transient provider logs described in §4, contact us at the address in §8 and we will respond.

6. Permissions, and why each one exists

Assist mode (the Accessibility Service)

Drutak declares an Android Accessibility Service called Assist mode. It is off by default, it is never switched on without you doing it yourself in Android's own Settings, and everything else in this app works without it.

What it is used for, exhaustively — there is no eleventh thing:

• Leaving an app hands-free. When you say "close WhatsApp" while WhatsApp is in front and Drutak is in the background, Assist mode sends your phone to the home screen. Android does not let an app with no window on screen change what is on screen; this is the only way to do it without putting a badge in front of you first.
• Going back hands-free. When you say "go back" or "press back" while another app is in front and Drutak is in the background, Assist mode presses Android's own Back once, on that app — the same Back you would press yourself. It reads nothing to do it and presses nothing else. Added 2026-09-23. With Drutak's own screen in front, "go back" still moves between Drutak's screens, as it always has.
• Scrolling the screen hands-free. When you say "scroll down" — or up, left, or right — Assist mode scrolls whatever is in front of you in that direction. This is the one gesture Drutak draws on the screen, and it is deliberately the narrowest kind: you name a direction and nothing else — never a place on the screen or a target — and any scroll is undone by scrolling back the other way, so there is nothing here you cannot immediately reverse.
• Using the phone's own controls hands-free. When you say "lock the phone", "take a screenshot", "show my notifications", "open quick settings", "power menu", "go home" or "recent apps", Assist mode asks Android for exactly that one action — the same as pressing the button or pulling down the shade yourself. Each is one of Android's own fixed actions, named in full below; it reads nothing to do it, and each is yours to undo: the shade and the menus close with Back, a locked phone opens with your own unlock, and a screenshot is saved where your phone saves the ones you take, for you to keep or delete. Since 2026-10-02, while your phone is locked, the shade, quick settings, the power menu and a screenshot wait for your unlock: none of them happens before it, and Drutak asks you to unlock — or, partway through a longer request, stops there and says so. Locking the phone, going home and recent apps do not wait. With the shade or quick settings open, what they show is a screen like any other: it is read only when you ask about it or ask for a press there, and then — the text of the notifications on it included — it goes to the language model with your request, as the read below describes. A press there is decided like any press that could grant (below), because System UI is also where Android asks you to allow things like USB debugging or recording your screen: made straight away, as you asked, unless the screen chose it and you did not (below). Added 2026-09-28; until 2026-10-04 it was asked about first on Drutak's own screen while the main switch of a Final check in Drutak's Settings was on.
• Naming the app in front of you, when you talk to Drutak with it open. When you speak to Drutak while another app is on your screen — "press 7" with the Calculator open — Drutak asks Android which app that window belongs to and sends that app's name, the one under its icon, with your request, so the AI knows a screen is there before it decides whether to read it. This leaves your phone, with every such request, and it is the name alone: not what is on the screen (that is the next item, and only when you ask), not how long you have used the app, and nothing at all while your phone is locked or its screen is off. The same name goes once more, with your words and nothing else, in a one-word question asked at the same moment: does this request need what the screen shows? Drutak uses that answer only when the AI's first reply neither did anything nor looked anything up. A yes reads the screen as the next item describes, and any press that follows is decided like every other press (the item after); a no, a late answer or none changes nothing. Added 2026-09-23. That yes is the AI's judgement, and it can be wrong. Measured the day it was added, with the phone's own question and each request asked twice: every request that needed the screen got a yes (56 of 56), and 5 of 54 that did not got one too — "swipe up", "go back" (once in Telugu) and "type pune weather" both times. When the first reply did nothing, such a yes reads your screen and sends it for a request that did not need it. Drutak could not type then; since 2026-09-24 it can type into a box on your screen (below), so a request to type something now needs the screen and is read for it.
• Reading what is on your screen, when you ask about it or ask for a press or typing. When you say "what does this say", "read this", "summarise this" — anything that points at what is in front of you instead of naming it — or ask Drutak to press something on it or type into a box on it (the next two items), or to open something inside the app that is already in front of you ("open Display & brightness" with Settings open, which the AI answers by opening the app you are already in; since 2026-09-23 that is read as a press to find on this screen), Drutak reads the text of the window on screen at that moment and sends it to the language model together with your request, because that is the only way anything can answer it or find the button or box you meant. This, the picture of the screen a step of a request sends (below), and the app's name above are the only uses on this list that leave your phone, and it works in whatever app you are looking at, not only WhatsApp. It happens on that turn and for that request only: there is no background reading, no copy is kept once the answer is spoken, and a request that neither points at your screen nor asks for a press or typing on it triggers it only when the one-word question above wrongly answers yes, when its words name something on the screen that can be pressed (the end of this item), or when it asks for more than one step — or the AI takes it to — and a step has run (below). Password fields are never read — a login screen contributes its labels and not the secret. At most the first 3,000 characters are taken. Things on the screen that can be pressed are sent marked — "▸ Play" — and numbered only where several say the same words — "▸ [3] OK" — so that you can ask for one of them by what it says (next item). Until 2026-09-23 every one was numbered. A box you can type into is marked as one — "▸ ✎ Search" — since 2026-09-24. Since 2026-09-27 the screen can go with your first request. "Press 7" with the Calculator open reached the AI with no screen at all, and came back as words instead of a press. So, with another app in front, your phone first looks at what on the screen can be pressed — on your phone only; that look sends nothing and keeps nothing, and the log notes only how many things there were, how long it took and whether your words named one. When your words name one of them — every word of its label is a word you said, and a number on it a number you said — the screen is read as above and sent with your first request, instead of after the AI asks for it, so the AI can choose the button on its first answer; the press is then decided like every other press (next item). Your phone cannot tell a request for a press from one that only uses the same words, so the screen goes whatever you asked: "play some music" with a Play button on the screen sends it too. A label with no word on it, like "×" or "=", never counts as named, and neither does a number you said as a word ("press seven"); those go as they did before, with the AI asking to read the screen first. Since 2026-09-28, when Drutak is not on your screen, the other ways the recogniser heard your words (§1.2) count too: when what it wrote first names nothing there and one of the others does, the screen is read for that one, so "press 7" heard as a person's name can still press the 7.
• Pressing a button you ask it to press. When you say "press play", "press 7" or "tap the second one", Drutak reads the screen as above and presses the button you meant in the app in front of you. An ordinary press happens straight away — "press 7" on the Calculator puts a 7 on the Calculator, with no question in between. This is changed from 2026-09-22; from 2026-09-21 until then it asked first, every time. A press that could pay, grant, call or send is made as you asked, straight away, too — unless the screen chose it and you did not (below), and Drutak's own action journal notes that it went unasked. Changed 2026-10-04; until then such a press was asked about first on Drutak's own screen — every time until 2026-09-28, and from then wherever a Final check switch in Drutak's Settings asked for that app (it was off after install); that switch is gone. Since 2026-10-01 one such press is asked about first: one the screen chose and you did not. That is a press that shows signs — the internal name its app gave the button, or what the AI says that press would do (below) — of sending or calling, deleting, paying or granting that what you asked for does not do, when you did not name the button yourself. What your request does is written down by the AI from your words alone, in the same question that writes down what your screen will show once it is done (§1.2; Doing several things you ask for at once, below), never from the screen: a page that says "pay to continue" cannot make a payment yours. If that question was not asked, has not answered within four seconds, or failed, nothing says that the screen chose the press, and it is made as any press you asked for: your own request is never asked about. Changed 2026-10-04; until then such a press was asked about too. Where a press is asked about, Drutak's own screen shows the button's name and the app it is in, and says that you did not ask for it and that it came from the screen; nothing is pressed until you say yes, and saying no or looking away presses nothing. A press that could pay, grant, call or send means, since 2026-09-23: any press at all in an app that can take a payment (one that answers a UPI payment link, one that can act as a contactless card, an app store, or one on a fixed list of payment and banking apps), and on the system screens that grant a permission, install an app or approve a connection; a switch or a button in Settings; a button or a switch in an app that can place calls, send messages or email, or add a contact — and in a dialer a row as well, because a row in your recent calls is a call; and, in any app, the button that confirms one of Android's standard dialogs, whatever it says, or a button its app has named, internally, for sending, texting, calling, dialling, paying, buying, checking out, installing, allowing or granting. In a messaging app, opening a conversation or pressing a text box is still an ordinary press, and so is moving between pages in Settings — since 2026-09-24 that includes opening Settings' own search, a button its app has named, internally, for search, which until then was asked first like any other button there. What your phone decides it from is what it says each app can do, what kind of thing the button is — a switch, a row in a list, a text box, a button — and the internal name its app gave it; never the words written on the button, because a "Play" and a "Pay" differ by one letter and no rule can tell them apart without reading their language. The AI that read your screen does read those words, and since 2026-09-23 it says, with each press, what that one press would do: send, post or call; delete or erase; spend money; give a permission or install an app; or none of these. What it says can only make Drutak ask where it otherwise would not, never ask less: a press it says would do any of those is treated as one that could pay, grant, call or send — asked first on Drutak's own screen when the screen chose it and you did not — and a press it calls harmless is still treated as one if your phone's own answer says so. If your phone cannot answer those questions, every press is treated as one that could pay or grant, and if a button is classed more strictly when it is pressed than when you asked for it, nothing is pressed. What this cannot tell yet, and you should know it: a press whose effect your phone has no way to see — a Pay button, on a web page or in an app, that does not name itself, or deleting something inside an ordinary app that does not confirm it with Android's standard dialog — is asked only when the AI says what it does. If the AI misses it, or words on the screen mislead it, it is an ordinary press today. It never presses anything on Drutak's own screens. It presses only from a read of the screen made for that same request: if it has not read the screen yet it reads it first, and if the read fails it tells you so and presses nothing — it never presses from an older read. Before pressing, Drutak finds the same button again, in the same app: if it has changed, moved, or become one of two identical ones, nothing is pressed. It presses a button the way Android's own accessibility tools do — a "click" on that element — never a point on the screen, and never a password field. The numbered buttons of the last few reads (four at most), and the rest of their text, are held in memory for two minutes, so the press can find the same one and typing can be checked against the screen (the item after), and are then forgotten; they are never written to your phone or sent anywhere beyond the read itself.
• Typing words you ask it to type. Since 2026-09-24, when you say "type lo-fi in the search box", or a request needs words in a box on your screen, Drutak reads the screen as above and puts the words into the box you meant, in the app in front of you. It replaces what the box held, and it presses the box's own Enter only when the AI asks for it with the words, and only once the box reads back exactly those words. Changed 2026-09-30; until then it never pressed Enter, Search or Send, and "search for Bahubali songs" stopped at a box holding the words. The AI asks for that Enter to search for the words, or to send them where the screen has no button of its own that sends; a Search or Send button on the screen is a press of its own, decided like every other press (the item before). The Enter is the one the box itself offers Android's accessibility tools — the keyboard's own search, go or send key — never a point on the screen. With it, the AI says what that Enter would do, as it does for a press, and an Enter it says would send, call, delete, pay or grant is treated as a press that could pay, grant, call or send. You hear that Drutak typed the words in and pressed Enter, or, when the box offers no Enter, that it typed them but could not press Enter there — never that it searched or sent, because only the screen after shows what that Enter did. The box is found the way a button is — from a read made for that same request, then again just before typing, in the same app, never a point on the screen — and whether it asks first is decided the same way too: an ordinary box is typed into straight away, and so, since 2026-10-04, is a box in an app that can take a payment, on the system screens that grant a permission or install an app, or one its app has named, internally, for sending, paying and the rest of the list in the item before. Until then such a box was asked about first on Drutak's own screen, showing the words, the box's name and the app, wherever a Final check in Drutak's Settings asked for that app, and so was every box if your phone could not answer those questions. A box in a messaging app, in the dialer or in Settings is ordinary, because typing into it sends, calls and changes nothing: the Send, Call or switch pressed afterwards is a press of its own, decided as above, and its Enter, when the AI asks for one, is ordinary too unless the AI says it would send or call. It never types into a password field, or anywhere on Drutak's own screens. A number of four digits or more is typed only if you said it: a code, a PIN, a phone or account number that the AI read on your screen or thought of itself is refused, whichever box it was for, and so is a year you did not say. Words the screen shows and you did not say are asked first, whatever the box: when all the words for a box, together and in order, are words Drutak read on a screen in the last two minutes (not counting what that box already holds), and you did not say them, they are an app's or a page's words, not yours — a page that says "type this into the box" is not you — so Drutak asks on its own screen, showing them, before it types them. Words in a script you did not speak count only when four or more of them are copied, since one or two may be the words you asked for, in another language. A word you said stays yours when the AI spells it one letter differently (a word of four letters or more), joins two of your words into one, or splits one into two. Afterwards Drutak reads the box back, and says the words are in only when the box holds exactly them — with an Enter, just before the Enter is pressed, since a box may empty once it is. The AI chooses the words from your request; nothing more leaves your phone to type them, and they are never written to your phone. What this cannot tell yet, and you should know it: a form on a web page is ordinary, like any press on a web page, so words the AI chose for one go in straight away unless they are copied off the screen (above), and the page's own Send button is asked only when the AI says what it does (the item before). Words from the screen that the AI reworded, or copied only in part, are not caught, and nor are three or fewer copied words in a script you did not speak. An Enter is judged by what the AI says it does: in a chat box whose Enter sends, one the AI calls harmless is an ordinary press, as a Send button it called harmless would be.
• Doing several things you ask for at once. When you say "go back and scroll down", "open Settings, then Display" or "open the Calculator and work out 25 times 8", the AI takes the first step and marks that you asked for more; what it writes about the rest is never used. It does not always mark it: asked to open the Calculator and work out 25 times 8, it opened the app and said "200" instead, twice in two. So when its first answer only opens an app, Drutak asks the AI one more one-word question as the app opens, with your words and that app's name and nothing else: did you ask for anything besides opening it? A yes goes on as if the AI had marked it; a no, a late answer or none leaves it an app you opened, as before. That yes is the AI's judgement and it can be wrong: then your screen is read and sent once for a request that only asked to open the app. Added 2026-09-24. Measured the day it was added, each request asked once: 3 of 6 requests that asked for more got a yes — the sum, in English and in Hindi, and starting the stopwatch — and none of 4 that only asked to open an app did; "open Settings and turn on hotspot", "open YouTube and search for lofi music" and the sum in Telugu got a no, and stopped at the app. Drutak goes on only once a step has run without a question and ended where it should: a press once the screen changed — an ordinary one, or one that could pay, grant, call or send that ran unasked (since 2026-09-28 where a Final check let it, since 2026-10-04 every one you asked for) — typing once the box reads back exactly its words, an app once it is in front of you, going back, a scroll or opening the notification shade or quick settings once Android has taken it. Then it reads your screen again — whatever the next step turns out to be, even a scroll, because the rest can only be chosen from where the last step left you — and sends that text to the language model with your same words and, for each step so far, your phone's own words for what it did: "went back", "scrolled down", "opened Settings", "pressed one line of the screen they were on" or "pressed 5 lines of the screen they were on, one after another", never the words on what it pressed. Since 2026-10-01 your phone's fixed sentences for what it found after the step before go with it too, with nothing joined into them: "The phone saw nothing on the screen change after that move."; one sentence on what Android shows playing, for a request to play (below, Whether it started playing); "The phone read this app before anything was done in it, and the result its screen shows now was already there then." with a line on what that means; or "The phone checked the screen: it already shows what they asked." — never what it looked for. They are facts, not a verdict: the AI chooses from them and the screen what is left. The AI answers with the next step, or tells you it is all done, with the answer when the screen shows one. That word is not taken alone (added 2026-09-24). As the first step goes — or, when the AI's first answer only opened an app, once the one-word question above says you asked for more — Drutak asks the AI once more, with your words and nothing else: no screen text, no earlier messages, no memory. The question is what your screen will show once it is all done: the number that answers a sum, a few words of what you asked to find or play, or the app it ends in — and, since 2026-09-26, which tab or item on it is chosen, which switch is on or off, and whether something on it keeps changing, like a stopwatch that runs. When its answer names one of those but not the app it ends in, the same question is asked once more, with your same words and nothing else (added 2026-09-27): without the app, your phone cannot tell where to look for it. Since 2026-10-01 the same question also asks which of four kinds of thing doing it does — sends or calls, deletes, pays, grants — never to whom or how much; your phone never looks for that on the screen, and uses it only to ask you before a press the screen chose (Pressing a button you ask it to press, above). Drutak looks for those only in what Android itself marks on the screen: a switch's own on or off, never the words an app writes beside it. Something that keeps changing is looked for in two reads a moment apart, on your phone only, and neither sends anything; one read alone never counts. Before each step after the first, the screen that step would be taken on is looked at for all of it too, on your phone only, and that look sends nothing. What it finds goes to the AI with that step as one fixed sentence — "The phone checked the screen: it already shows what they asked." — never what it looked for, and the AI decides from it and the screen what is left: a stopwatch that already runs is the AI's to see, and to leave alone. Since 2026-10-01; before, your phone ended the request there itself, with nothing pressed, on its own reading of the screen. When your screen was read for your first request — because the AI asked to read it or, since 2026-09-27, because your words named something on it (above) — this question is asked beside that request, whether or not the request turns out to have more than one step. From 2026-09-26 the button the AI chose for a first press was checked against its answer before it was pressed; since 2026-10-01 it is not: Drutak waits up to four seconds for that answer, so that a press for something to play is followed by a look at what plays (Whether it started playing, below), and the press goes whatever the answer says, or with none — but for one case, added 2026-10-02. When that answer names only switches or tabs, and the screen the press was chosen from, read before any step and of the app the answer names, already shows each one so — looked for as above, on your phone only, and that look sends nothing — the press is not made: pressing a switch that is already on turns it off. The AI is asked once more instead, never twice in a row, with your screen read and sent again, your same words, its answer written back as its kind alone (a press, never what it named), and one fixed sentence of your phone's: "The phone did not do that, and did nothing: before any step, the screen that answer was chosen from already showed what they asked. If it still shows what they asked for, the request is done and nothing needs pressing." — never what it looked for. The AI decides from that and the screen: that it is all done, or the step it names. A number, words to find, something to play and something that keeps changing are never looked for there: before any step nothing on the screen is yet of your request, what plays is not on it, and one read alone never counts. One case is still checked before each step: a request picked up after Android closed Drutak, when nobody can say whether its last step went through (below). Until a step goes on what your phone sees that last step left, the screen is checked against the answer before each step is taken: when it already shows everything you asked, nothing is pressed, and you hear that there was nothing more to do (for a request to play, that it plays, when Android shows a new play); when the answer names a switch, a tab or something that keeps changing and your phone cannot tell from the screen whether it is already so, even after one more look on your phone that sends nothing, a press that could change it is not made on a guess: you hear that your phone could not tell. When the AI says it is all done, your screen is read once more, on your phone only, and that read sends nothing: Drutak looks on it for what that question said, and for any number in the AI's words that neither you nor the screen said. When the screen plainly shows something else — another number, or the home screen instead of the app — the AI is asked again instead: once for a request, twice for a request to play (Whether it started playing, below), never right after it was asked again about a step it named (below), and only while the request still has time, with your screen read and sent again and your phone's words that the screen does not show it is done, never what it looked for. Then, only when that read finds it done, you hear the AI's own words — unless they hold a number neither you nor the screen said, or leave out the answer the screen shows; then your phone's own line that the screen shows it is done, with the answer it found there. When the read finds it plainly not done, you hear that it tried and the screen does not show it done; when the read cannot tell — nothing it could look for, no screen to look at, or a switch or state it could not see — you hear "I couldn't verify that it completed." Either way, never the AI's words that it is done, and the request is kept as not done; the diagnostic log notes only whether the read found it done and whose words you heard. This is since 2026-10-01. Earlier the same day, the AI had the last word even where the screen did not bear its words out; before that, you heard your phone's line there, and where it could not check, the AI's words followed by "I couldn't check that on the screen." Several buttons on one screen, like the keys of a sum, can be chosen at once: every one of them must be on that screen, once, before the first is pressed, and each of the others is found again just before it is pressed, on a new read made on your phone that sends nothing. Every step and every press goes through exactly the same checks as a request made on its own: an ordinary press happens straight away, and so does a press that could pay, grant, call or send unless the screen chose it and you did not, and a press is made only from a read made for it. At most thirty steps, a hundred presses and five minutes for one request, not counting the time it waits for your answer on your phone's own screen (below); a step that would repeat the one before — but for a scroll or going back — is asked back once (below), and taken only when the AI answers it again; and going back, a scroll or opening the notification shade or quick settings that leaves the screen exactly as it was is told to the AI, in your phone's words "The phone saw nothing on the screen change after that move.", and three such moves in a row end it. Since 2026-10-01; before, the same step twice in a row, but for a scroll, or one such move ended it. Since 2026-09-29 each step after the first also sends a picture of your screen. The list of what can be pressed reads some screens badly (a video's card, an advertisement, a dialog drawn as a picture), so, on Android 11 and later, your phone takes one picture of the screen in front for that step, made small (720 pixels wide), and sends it with the screen's text to Google's Gemini through our server, which chooses the step (§4). One picture per step of a request you made, never on a timer, never kept on your phone and never written to your storage or gallery; the log notes only its size and how long it took. A window the app marks as secure — banking apps and password screens usually do — Android leaves out of the picture: it comes out black. Since 2026-10-02 some screens are never pictured. No picture is taken while your phone is locked or its screen is off, while a call rings or is under way, or while the app in front is Drutak itself, one Android says takes payments (UPI or a contactless card) or sells apps, or one that asks for permissions or installs apps; Android's own system screens, like the notification shade, count among these. Nor when your phone cannot tell which app is in front or what Android says it is, or once the time your request is given is up. Your phone asks again once the picture is taken, and drops it if a different app came to the front meanwhile. That step goes on with the screen's text alone, and the log notes which of these reasons it was, never the app. Other apps — a messenger, your dialer, Settings — are pictured as before. When Gemini is busy or slow to answer, our server asks it once more — its lighter model, with the same picture and inside the same wait — and takes whichever answer comes first; for a minute after Gemini is busy or slow (or as long as Google says it will be busy, at most ten minutes), steps are asked of the lighter model first. A step is never asked of Gemini more than twice. When neither answers, the step is asked of Groq as before, with the text alone: the picture goes to Google and nowhere else. A picture presses nothing: what the AI chooses is decided exactly like every other step. One case is asked about again instead of ended: an ordinary press after which your phone watched the screen and saw nothing on it change, with more of your request still to do — like the key that clears a calculator that already shows nothing. The buttons it had chosen to press after that one are dropped, and your screen is read and sent again, with the same words as any other step and your phone's line for that step ending ", and saw nothing on the screen change after it". What that press did is still not known, and Drutak does not say it worked. If the next press changes nothing either, it stops. Added 2026-09-24. That button is not pressed again next: if the AI's next answer starts with it and names more buttons after it, it is left out and the others are pressed in the AI's order, exactly as buttons chosen at once always are (above). Nothing is added. Added 2026-09-24. The other case: the AI answers a step after the first with a button or a box that is on no line of the screen just read, with only the button that just changed nothing, by opening the app already in front or, since 2026-10-01, with exactly the step before, or with a button or a box that several lines of the screen say when it named none of them by the number the screen writes before each. Nothing is done, and it is asked once more — never twice in a row — with your screen read and sent again, the question it answered, its answer written back as its kind alone (a press, words for a box, an app to open, a scroll, going back, a search or music — never what it named), and your phone's own words for why nothing was done. Added 2026-09-24. When it answers once more with exactly the step before, that step is taken: the AI was told it would repeat the step before, and chose it. Never a press or typing that is not ordinary — one that could pay, grant, call or send, or words copied off a screen — and never several buttons at once: the request stops there. Any other step repeated — a switch asked for again, or the notification shade or quick settings opened again — still stops it, as before. Added 2026-10-01. One more case: a press or typing your phone did not make because, when it went to make it, the button or box was no longer on the screen as it had been read. Nothing was pressed or typed, and with the app it was read from still in front of you, the request is asked about again instead of ended, with your screen read and sent again. When it was one of several buttons chosen at once, but not the first, the ones pressed before it stay pressed and your phone does not press them again itself, the ones after it are dropped, and your phone's line for that step ends ", then found the next line it was to press no longer on the screen as it had been read, and pressed no more" — which can come again in the same request, but only ever after a press that landed. Otherwise it is asked once more, as in the case above and never twice in a row, with your phone's words that the button or box it named was no longer on the screen as it had been read. Before the request goes on, that press or typing is taken back off the request's record (below); if that cannot be written, it stops there, saying so. A press or typing that would have been the first thing the request did, another app in front, a phone that is locked or whose screen is off, and a request out of time still stop there, as below. The press or typing that was not made is not counted against the limits above; whatever comes after it is. Added 2026-09-25. If a step asked you anything (but for the case below), failed, or could not be seen to land, if a button is no longer on the screen (but for the case above) or is on it twice (but, since 2026-10-01, for the AI's answer to a step after the first: the case above), if your phone is locked or its screen is off, or if the AI's answer comes back empty or would repeat the step before (but for the case above), Drutak stops there and says so: what happened to the step it took and, if anything ran, that it did not do the rest. All the steps count as one question (§1.7). Added 2026-09-23; more than two steps since 2026-09-24. Since 2026-09-30 a request you make on Drutak's own screen goes on this way too when its first step opens an app, searches the web or YouTube, or plays music: with Voice replies on, it is handed to the part of Drutak that listens in the background, which carries it on in the app it opens, as above (§1.2). Before, Drutak's screen stopped once that app was open. Until 2026-10-02 the wake word had to be on as well; with it off, that part now starts for the request alone, with no microphone open, and stops when the request is over. A press or typing that your phone's own screen asks you about first — a press the screen chose and you did not, or typing of words copied off the screen — does not end the request. Drutak says it is asking you there and, with more of your request left, that it will carry on with the rest after; the request then waits for your answer, its record kept as described below. Your answer is never given for you: only your tap on that screen says yes, and Drutak takes no step of the request while it waits. Since 2026-10-03 you can also call it off by telling Drutak, while it is not on your screen, to stop your request: Drutak cancels the step it asked you about in its own action journal, ends the request, kept as called off (§3), and tells you that step will not happen; its question can stay on your screen, and answering it then does not carry that step out. If the journal does not say within three seconds that the step was cancelled, the request ends all the same, kept as stopped, and Drutak says it cannot tell whether that step went through and asks you to check your screen. The AI decides whether your words mean this request; it is told that stopping anything else on your phone, such as music, is not. It goes on only once your phone saw that press or typing land, from a fresh read of your screen, like any step after the first: the buttons it had chosen before the question are dropped, and your phone's line for that step ends ", once they confirmed it on the phone's own screen". It goes on only while your phone is awake, unlocked and showing the app that step was taken in or the app the request was in before it, and only in a quiet moment — not while you are talking to Drutak, the app is open or the microphone is switched off — that comes within ten seconds. Otherwise it stops there: the step was done and, if more was left, the rest was not. If you tap Cancel, the request ends without a word. If the question could not be shown, went unanswered (your phone stops asking after two minutes), was refused, or the step failed, Drutak stops there. If nobody can tell whether the step went through, Drutak stops there and asks you to check your screen: a step you were asked about is never gone on from on a guess. Whatever ends it is said at the first quiet moment, if one comes within ten seconds. A new request of several steps takes its place, and one waiting ends when Drutak stops. Since 2026-10-02, so does a request of yours that Drutak's own screen opens an app for, and the one waiting is kept as stopped, like any other that ended (§3). Its question can stay on your phone's screen until your phone stops asking: answering it then does that one step, and nothing more of the old request. Added 2026-09-25. A wait the AI chooses. Since 2026-10-03, when what is left of such a request has to wait for something still to come — a download to finish, a code sent by message, a page still loading — the AI can answer a step after the first with a wait of its own choosing, and Drutak takes no step meanwhile: at most ten minutes a wait, and at most thirty minutes of waits in all for one request (after Android closes Drutak, all the time the request had waited before counts, whatever it waited for). Its record says it waits and until when (§3) before anything else is done; if that cannot be written, the request stops there, saying so. Drutak says it will wait, then look at the screen again, and keeps your screen on, dimmed, for the wait and at most half a minute more, so that the app it waits in stays in front. Nothing is pressed or typed and your screen is not read while it waits, and no microphone is opened for it. When Drutak listens for a few seconds after saying it will wait, that is the follow-up listening after an answer (§1.1). Your screen going off ends the request there, said. A newer request takes its place; telling Drutak to stop your request, Drutak stopping, and Delete all my data (§5) end it. Once the time is up it goes on only while your phone is awake, unlocked and showing the app the request was in, with no call on, inside its five minutes, and in a quiet moment, as above, that comes within ten seconds, from a fresh read of your screen: the AI is told that your phone waited, then read the screen again, and decides what is left. Since 2026-10-04 it does not go on while Drutak is listening for your words, or is about to: it looks at your phone again a second later, and that time is not counted in the ten seconds. Otherwise it ends there, and Drutak says it stopped, at the first quiet moment, if one comes within ten seconds. The time waited is not taken from its five minutes; each wait counts as one of its steps. A wait with no time your phone can read, or past the request's steps, its five minutes or its thirty minutes of waits, ends the request there, said. A wait in the AI's first answer to a request is cut out. Added 2026-10-03. If Android closes Drutak in the middle of such a request — to free memory, say — Drutak can pick it up again when it restarts, and never does a step twice. While the request goes on, your phone keeps a record of where it stands in Drutak's private storage (§3), and every step is written into it before it is handed to Drutak's own action journal, so the record always names the one action that may have been under way: by its kind and the key the journal filed it under, never by what it pressed or typed. A step that cannot be written down is not taken, and the request stops there, saying so. A request longer than a thousand characters is not kept at all, and is not picked up again. When Drutak starts listening again it asks its journal what became of that action, and nothing in the record can be pressed again. If the action did not happen, or happened and was a press, typing, opening an app, going back, a scroll or opening the notification shade or quick settings, the request goes on from a fresh read of your screen, like any step after the first (§1.2). If nobody can say whether it happened, the request goes on only when it was an ordinary press or typing, opening an app, going back, a scroll or opening the notification shade or quick settings, with the AI told it is not known whether it went through, so that your screen decides; anything else — a press or typing that could pay, grant, call or send, whether you were asked about it or not, a message, an alarm — stops there, and Drutak asks you to check your screen. A request whose last step was refused or failed, or was one that ends a request by itself (a message, an alarm, going Home, locking the phone, a screenshot) and may have been carried out, is not picked up either. A question Drutak had asked you about a press is never answered for you: if the request goes on and still needs that press, it is decided afresh, as any press is (above). The request goes on only inside its five minutes, on the same start of your phone, at most twice, and only while your phone is awake, unlocked and still showing the app the request was in, with nothing else going on — you talking to Drutak, the app open, the microphone switched off. Otherwise it ends there, and Drutak says it was interrupted and stopped — or, when it cannot tell whether the last step went through, asks you to check the screen — at the first quiet moment, if one comes within ten seconds: never while you are talking to it, the app is open or the microphone is switched off. A record from before your phone last restarted or more than two minutes old, one found when Drutak is started straight into a new request (the assistant button on a headset or, since 2026-10-05, a hold of the power button with Drutak as your digital assistant), and one Drutak finds when it stops again without having started listening are deleted without a word. Since 2026-10-02, so is one found when Drutak is started for a request you made on its own screen, or when that screen opens an app for a newer request of yours (§1.2). Since 2026-10-03, a request that was waiting for a call to end before its next step counts its two minutes from when that wait was to end — fifteen minutes after the wait began — and the time since that wait began, up to its end, is not taken from its five minutes. If a call still has your phone's audio when Drutak starts listening again, Drutak waits for the call to end — no later than that wait was to end — and then decides as above; if the call goes on longer, the request ends there, and Drutak says it stopped because of the call, as above — never while the call goes on, and not at all if it goes on another fifteen minutes. Since 2026-10-03, the same holds for a request waiting out a time the AI chose (above): its two minutes count from when that wait was to end, and the time it waited, up to that end, is not taken from its five minutes. While that time is not yet up and your phone is awake and unlocked, Drutak waits it out, your screen kept on, dimmed, as above, then decides as above; your screen going off ends it, said. Added 2026-09-25.
• Tapping Send in WhatsApp, and only after you have asked Drutak to send that message. Since 2026-09-28, when you ask Drutak, while it is on your screen, to send a WhatsApp message to someone in your contacts, it opens that conversation in WhatsApp or WhatsApp Business with your message in it and presses Send — only in a conversation whose name is that contact's name, only while the box holds exactly your message, and only within fifteen seconds; it says the message went only once the box has emptied. The message goes as you asked, with no question first. Changed 2026-10-04; from 2026-09-28 until then, wherever a Final check in Drutak's Settings asked for WhatsApp, you were asked first on Drutak's own screen, which showed who it went to. Since 2026-10-05 one such message is asked about first, on that screen, which also says why: one to a person only something Drutak read for that request named — a web page, your screen or your calendar, say — and you did not, in that request, an earlier one, or any other way your phone heard it. A person you named is never asked about. Without Assist mode, or with Drutak not on your screen, the message is typed out in the conversation and you tap Send yourself. Until 2026-09-28 this was switched off in every version.

What Assist mode does NOT do, by construction:

• It no longer has a package scope, and you should know that plainly. Until 2026-09-20 this section said Android delivered Assist mode screen events from WhatsApp and WhatsApp Business only. That restriction was removed, because it did not do what this policy claimed it did: with it in place, the on-request read described above could not actually reach the app in front of you — it returned Drutak's own screen instead. The choice was to drop the read or drop the scope, and the read is the useful half. So Android now delivers Assist mode window events from every app. What that does not mean: those events are not a feed of anyone's screen. Assist mode does nothing with an event unless a WhatsApp send is already pending — every other event is discarded the instant it arrives, nothing is buffered, nothing is stored, and nothing is sent anywhere. The only time any screen text leaves your phone is the read above, on a request you make: when you ask about what you are looking at, ask for a press or typing on it, or name something on it that can be pressed. Going home, going back, scrolling and the phone's own controls are not reads at all, though a request for more than one step reads the screen between them, as described above. A picture of the screen leaves your phone only with a step of such a request, to Google's Gemini (above), since 2026-09-29. Pressing a button reads the window once more, on your phone only, to find the same button, and sends nothing. So does the check of a request the AI says is done, and so does the look at what can be pressed that comes before each request you make with another app in front (above): it keeps nothing, and the screen is read to be sent only when your words name one of those things. An earlier version of this policy said your banking app's contents are never sent to Drutak. That sentence stopped being true when the read above was added, so it was removed rather than left standing.
• It cannot tap a coordinate or drag. Until 2026-09-21 this line said it could not tap a button at all, and until 2026-09-24 that it could not type; each stopped being true when the press and the typing above were added, so the line was rewritten rather than left standing. What holds now: it presses only a button it has just read for you — after your yes on its own screen when the screen chose a press that could pay, grant, call or send and you did not — or WhatsApp's Send for a message you asked it to send (above), and types only into a box it has just read for you, replacing what the box held and pressing nothing but, since 2026-09-30, that box's own Enter when the AI asks for it, once the box reads back exactly the words (above); never a place on the screen. It draws one gesture and one only — a reversible scroll in a direction you name — and it is built so that no place on the screen can be passed to it from outside. The permission that lets it scroll (canPerformGestures) is requested for exactly that: the code that forms a gesture lives in a single place and can be reached only through those four fixed scroll verbs, so "scroll" can never quietly become "tap anywhere".
• Its navigation is exactly eight fixed actions — home, back, recents, locking the screen, a screenshot, the notification shade, quick settings and the power menu — plus the four scrolls above, and nothing else. It cannot dismiss or clear a notification, and it cannot switch the phone off by itself: the power menu only opens, and anything in it is a press you ask for, decided like any other (the phone's own controls, above). Until 2026-09-28 this line said three fixed actions; it was rewritten the day the other five were added. The language model that understands your sentence can choose one of these and can do nothing more with it. The one thing it can name beyond them is a button or a box from a read, by the words on it, and by its number only where several say the same words — a number alone presses nothing — and, for a box, the words to put in it and whether its Enter follows them. Your phone, not the model, decides which button or box that is and whether the press or the typing is asked first.
• Nothing it reads is stored or logged — not the screen text, not the name of a conversation, not a message, not a button it pressed, not the words it typed or the name of the box it typed them into, not the picture a step takes of the screen (the log notes its size and how long it took, never what it shows). The only thing held at all is the last few reads (four at most), their numbered buttons and boxes and the rest of their text, in memory, for two minutes, to find a button again and to check words for a box against the screen. The WhatsApp match compares the name you asked for against the one on screen, on your phone, and discards it. Diagnostic logs record how many characters were read and never what they said.
• Sent anywhere has exactly two exceptions. The read you asked for, that your words naming something on the screen made, or that a request for several steps makes between them: the screen text goes to the language model with your question, on that turn, and is not written to your phone or to ours. With a step of a request for several steps, a small picture of that screen goes too, to Google's Gemini alone, and is not written to your phone or to ours either (above). The permission that takes it (canTakeScreenshot, since 2026-09-29) is requested for that picture alone: it is taken in one place, reached only from such a step, and never on a timer. And the name of the app in front of you, with a request you make while it is open (above). Everything else Assist mode reads stays on the device.

Turning it off: Android Settings → Accessibility → Drutak → off. Drutak keeps working; "close app" then tells you it cannot, instead of pretending it did.

Force stop: Drutak cannot force-stop another app and does not claim to. "Close app" means leave that app. When the app is not on your screen, Drutak opens its Settings page for you, where Android's own Force stop button is.

Notification access (the Notification Listener)

Drutak declares an Android Notification Listener Service. It is off unless you turn it on, on Android's own screen (Settings → Notifications → Device & app notifications → Special app access → Drutak), and Android does not let an app grant this to itself. Until you grant it, asking "what came in on my phone?" gets you "I can't see your notifications yet" — never a guess.

Switched on, it exists for two questions: what came in while I was away? and, since 2026-09-28, what is my phone playing? — and, since 2026-10-01, for one check: whether something you asked Drutak to play started. Since 2026-10-05 it is also how Drutak answers a message that came in, when you ask it to ("reply to Amma: I'm on my way"). For the first, Android hands Drutak the notifications your phone posts, and Drutak keeps a short list of the most recent ones so that it can tell you what they were when you ask. The second, the check and a reply are different reads with different rules, set out under What is playing, Whether it started playing and Answering a message below.

What it does NOT do with your notifications, by construction:

• Nothing is kept on disk. The list lives in memory while Drutak is running and is gone when Android stops it. When you ask, it is handed to the part of Drutak that answers through a temporary file in the app's private storage, deleted as soon as it has been read. There is no notification history file and nothing is backed up anywhere.
• Nothing is sent off the phone by Drutak. What came in never reaches us, the language model, or the speech provider behind the premium voice. You can ask in any language: when the phone's own list of commands does not recognise the question, your words go to the language model like any other question (§1.2), and all it may answer is that this is the question — a mark with nothing in it, which cannot choose an app, a number, or what to do with the answer. The model is never shown a notification; the phone answers. The answer you hear is assembled on your phone out of the notifications themselves — which is also why Drutak repeats what a notification said rather than rewording it — and it is spoken only in the phone's own voice (§1.4), never the premium one, choosing a voice installed on the phone over one made online wherever the phone offers both.
• Nothing is carried into your next question. The answer is kept out of the conversation Drutak sends to the language model with whatever you ask next, so hearing your notifications once does not post them to a server a moment later. They are read only as the answer to that question on its own, never as one step of a longer request, where what each step finds goes back to the model.
• Nothing is read out from a locked phone. With Drutak closed and the phone locked, asking gets you "unlock your phone first" and nothing else; unlock within a minute and it answers then. A phone lying on a table does not read your messages to whoever speaks near it.
• Nothing is logged. Diagnostics record that a notification arrived, which app posted it, and how many characters it was; when you ask, they record whether access was on, and counts — how many notifications, from how many apps, how many characters. They never contain the text.
• Nothing becomes a remembered fact on its own. A notification is something that happened, not something you told Drutak. It enters Memory only if you ask for that in so many words.
• Nothing in a notification becomes an instruction. A notification reading "delete all my files" is text Drutak can read back to you and nothing more. What Drutak does comes from you.
• Only the app name, title, text, time, and category are kept, each clipped short. Attachments, images and the rest of Android's internals are not read, and a notification's buttons are read only for a reply you ask for, at that moment, and never kept (see Answering a message).
• It grants Drutak one power, and only when you ask. Since 2026-10-05 Drutak can answer a message that came in, through that notification's own reply button, as a smartwatch or a car does — only when you ask it to reply, only in a conversation whose notification is on your phone at that moment, and never on its own (see Answering a message). Being able to see a message is still not being able to call anyone or delete anything; those are separate permissions you grant separately. A new message to someone goes the way §2 describes, never through a notification.

What is playing. The same access lets Drutak see the list Android keeps of the apps playing sound or video — the list your lock screen's player is drawn from. Drutak reads what is playing from it only when you ask ("what song is this?", "which film is this from?"), in any language; never in the background, and never as one step of a longer request. It takes one entry — what is playing, or else what is paused, or else what played last — and from it only the title, the artist, the album, the app's name, whether it is playing, paused, loading or stopped, and how far into it you are and how long it is, each clipped short. Drutak's own voice is never counted as what is playing.

Unlike your notifications, these are sent to the language model with your question (§1.2), because saying who sings it or which film it is from, in your language, is the model's job and cannot be done on the phone. The answer joins the conversation like any other answer, and is spoken in whichever voice Drutak is using, the better voice included (§1.4). It is read with the phone locked too, since your lock screen already shows it. Nothing of the read is kept once the model has it; this lookup's own diagnostics record that it was asked and whether anything was playing, never a title. A title is never an instruction and never becomes a remembered fact. With the access off, asking gets you "I can't see what's playing" and where to turn it on, never a guess from a song playing around you.

Whether it started playing. Since 2026-10-01, when you ask Drutak to play something ("play Bahubali songs on YouTube"), the same list is how your phone tells whether it did: a new play has to show up there, and a song's name on your screen never counts as one. For that, Drutak reads the list when it starts on your request and again after its steps — after a press, once a new play shows up or it has waited a few seconds for one, since a video you pressed takes a moment to start — and from each app's entry takes only numbers — whether it is playing, paused, loading or stopped, how far into it, how long it is, how fast it plays and which item of its queue — and which app it is. Never the title, the artist or the album. The reads are compared on your phone, in memory, and dropped when the request ends: they are never written to storage, never part of the record of a request still being carried out (§3), never sent to the language model or to anyone else, and the diagnostic log records only a word for what was found: that a new play began, that none did, that only what played before still plays, or that it could not tell. What the language model learns is one fixed sentence of your phone's, with each step after the first: "Android shows something new playing since they asked.", "Android shows nothing new playing since they asked.", "Android shows only what was already playing before they asked, nothing new.", "Android shows something moving, but the phone cannot tell whether it is new since they asked: judge from the screen whether what they asked plays." or "The phone cannot read what plays on this phone: judge from the screen whether what they asked plays." — never a number, a title or an app. The AI, not your phone, decides each step and when to stop. When it says it is done and nothing new plays, it is asked again, at most twice, with "The phone looked after that at what plays on this phone." before that sentence. You hear that it plays only when your phone saw a new play: then the AI's words, unless they hold a number neither you nor the screen said, or there are none — then "Playing". When Android still shows nothing playing at all, in the app in front or anywhere new, you hear "Nothing started playing", whatever the AI said. When only what played before still plays, or your phone cannot tell — something moves but it cannot say whether it is new, or it cannot read what plays — you hear "I couldn't verify that it completed.": your phone's numbers cannot tell a song you asked for that was already playing from another, and the AI's judgement from the screen, which can be wrong, is never said as done. One end your phone sees itself: a new play right after the press the AI chose to end on ends the request, in the AI's own words for that press, with no question spent. Drutak's own voice is never counted as a play. With the access off, the AI is told the phone cannot read what plays and judges from the screen which step to take; when it says it is done, you hear that Drutak couldn't verify it. This is since 2026-10-01. Earlier the same day, after the second question the AI had the last word: its words were said where they held no number nobody said, even when only what played before still played or the phone could not tell, and with the access off you heard its judgement; before that, with the access off, Drutak said it couldn't tell whether it was playing — never that it was. Each version was rewritten rather than left standing once it stopped being true.

Answering a message. Since 2026-10-05, when you ask Drutak to reply to a message that came in ("reply to Amma: I'm on my way", or "reply: I'm on my way" just after Drutak told you what came in) while its own screen is not in front of you, it answers through the reply button that message's own notification carries — the button a smartwatch or a car uses for a quick reply. Drutak opens no app and presses nothing on your screen: your words are put in that button's box and handed to the app that posted the notification, which sends them as it would from your notification shade. A button that would open a screen instead, or drop your words, is never used — Android 12 and newer say which buttons those are; older Android cannot say, and there a reply button is taken at its word. When the app's notification has no such button, you hear "That app doesn't take a reply from its notification" and nothing is sent. With Drutak's own screen in front, answering someone is a new message to them, the way §2 describes.

What it reads, and when. Only when you ask to reply, Drutak reads the notifications in your shade at that moment, and from each takes only which app posted it, the name your phone shows for its conversation (a group's name before its sender's), when it was posted, whether the app marks it as a conversation of messages, and its reply button. Never the message itself. Drutak's own notifications and a group's summary are skipped. That read is not written to storage, never sent to the language model or to anyone else, and dropped once the reply is handed over or refused. The language model writes only what you said — whose message, as you named it, and your reply in your words (§1.2) — and never sees a notification.

Whose message. Your phone matches the name you said to the names of those conversations, the way it matches a name you ask it to call. When what you said is that conversation's name, or words of it, or the same name spelled another way, the reply goes. When it is only close — spelled or sounding nearly the same, or the name with a word more — or two conversations go by that name, Drutak asks you first, on its own screen, which shows the conversation it chose (of two by one name, the newer), that app and your reply, with Reply and Cancel; when it fits several different conversations, or none, you hear so and nothing is sent. Since 2026-10-05 it asks you first the same way, and its screen says why, when only something Drutak read for that request — a web page, your screen, your calendar, say — named whose message it is, and you did not. Said with no name, the reply goes to the conversation Drutak's last read-out of your notifications named, if that read-out named just one, in the last two minutes; for this, Drutak holds the keys Android gives the notifications a read-out names, in memory only, replaced by the next read-out and gone when Drutak stops. If the read-out named several and only one of them can be answered now, or there was no such read-out and only one conversation in your shade can be answered, Drutak asks you first. When none of the conversations the read-out named, or none at all, can be answered now, you hear that Drutak could not find the message; otherwise it asks you to say their name with your reply. Either way, nothing is sent.

Never from a locked phone. A reply goes out from your phone in your name, so with the phone locked or its screen off Drutak asks you to unlock it, raising the lock screen's own unlock prompt, and replies once you have, within a minute.

What you hear, and what is kept. "I sent your reply" only when, within about four seconds, the app posted that notification again or took it away, as a messaging app does once it has taken a reply in — your phone's record of that one notification, not proof that the message reached anyone. Handed over and not seen to go, you hear "I gave your reply to the app, but I couldn't see it go"; a reply that could not be handed over, you hear that it was not sent; and when its sending broke off midway, you hear that Drutak cannot tell whether it happened. To see that, Drutak keeps the key Android gives each notification it answered, in memory only, at most eight, the oldest dropped first, until Android withdraws the access, Drutak stops or Delete all my data (§5). The History screen (§3) shows Replying to a message and how it ended, never whom, which app or which words; Drutak's action journal (§3) keeps which app it went through, and the notification's key, the conversation's name and your words only as a fingerprint. Diagnostics record no name, no words and nothing of the notification. With Notification access off, Drutak cannot reply this way at all, and says that the access is off.

Turning it off: the same Android screen, off. The list Drutak was keeping is dropped the moment Android withdraws the access — not when you next ask, and not when the app is next closed. Drutak keeps working; it then tells you it cannot see your notifications, instead of pretending there are none.

Find my phone

Say "where are you?", "find my phone" or "ring my phone" — with the app open or closed, the phone locked or not — and the phone rings until you find it. Your words reach the language model like any other question (§1.2); the ring itself happens on the phone and sends nothing anywhere.

What a ring changes, and puts back. For the ring, and only for it: the alarm volume is raised to its top; if Do Not Disturb would keep an alarm silent and you have given Drutak Do Not Disturb access, it is lifted to let alarms through; the flashlight blinks, if it was off; the phone vibrates; and a black screen saying "Here I am" and "Tap anywhere to stop" comes up, over the lock screen when the phone is locked. When the ring ends, the alarm volume and Do Not Disturb go back to what they were — unless you changed one of them yourself while it rang, and then your change stays.

What stops it: a touch on that screen, Back or Home, a volume key, the power key, unlocking the phone, saying Drutak's name or "stop ringing", or two minutes with nobody stopping it.

What the screen over the lock can do: nothing but stop the ring. It shows those two lines and nothing of yours, it never unlocks the phone, and a touch on it reaches nothing behind it. While the phone rings, Drutak does not open its microphone to hear a follow-up question, which would hear only the ring.

7. Children

Drutak is for adults. It is not directed to children and is not intended for use by anyone under 18. We do not knowingly collect data from anyone under 18. If you believe someone under 18 has used Drutak and you want their data removed, deleting the app removes it; you may also contact us.

8. Contact

Questions, complaints, or requests about this policy or your data:

lokeshsiripurapu837@gmail.com

If you are in a jurisdiction that grants you rights of access, correction, deletion, or objection over your personal data, you may exercise them by contacting us. In practice, §5 already gives you complete deletion of everything on your phone without needing to ask.

9. Changes to this policy

If this policy changes in a way that affects what leaves your phone, we will update the date at the top and describe the change in the app's release notes. We will not quietly widen what Drutak collects.

10. Reporting AI-generated content

If Drutak produces a response you believe is harmful, offensive, or inappropriate, you can report it from within the app with Report this reply. Before anything is sent, that screen shows you what you said and what Drutak said, and Cancel sends nothing.

A report contains exactly these, and nothing else:

• The reply you reported, up to 2,000 characters.
• The message of yours that produced it — that one message, up to 2,000 characters. The rest of the conversation is not included.
• The reason you chose: offensive, unsafe, or incorrect.
• When you reported it.
• The name of the AI model that wrote the reply — a fact about the app, not about you.

It does not contain your name, your contacts, anything Drutak remembers about you, your location, the instructions Drutak gives the AI, your install number (§1.7) or any identifier for your phone. Sending it reveals your IP address to our server, which handles it as §1.3 describes and does not store it with the report.

A report waits on your phone until it can be sent, and Delete all my data (§5) erases one that is still waiting. Once it arrives, our server keeps it, with the time it arrived, for 90 days so we can investigate, and then deletes it automatically. It is not passed to Groq or anyone else.