Privacy
What leaves your device, and what doesn’t
Last updated 18 September 2026. Written to be read, not to be survived.
The short version
In the native app, your spoken review answers are transcribed on your phone and the recording never leaves it. The one exception is a deep dive, a live spoken conversation about a card that you start yourself: its audio goes to OpenAI, through our server, for as long as the conversation lasts. Your review log never reaches our servers. There is no account, no analytics and no tracking in either app. What we do send, when the walking tutor is running, is the text of the card being asked and the text of what you said, so that it can be graded and spoken back to you.
Who we are
Loudwalk is published by ANGELTECH (SAS), trading as GRAINES DE CODE, registered at RCS Brive, VAT FR39878731157. For anything on this page, write to privacy@agentkey.dev. We are the data controller for the processing described here.
Your cards and your review history
They are files on your own devices. Loudwalk stores your collection and your review log in your app’s folder, and mirrors them to your iCloud Drive so your Mac and your iPhone agree. That folder is yours: we have no access to it, no copy of it, and no ability to read it.
Your existing Anki collection is read from a copy and is never written to.
The walking tutor
When the native app’s walking tutor is running:
- Your answers are transcribed on your device. Speech recognition uses Apple’s on-device speech framework. The audio of a review answer is never uploaded, never stored, and never reaches us or anyone else.
- A deep dive is a live call. When you start a conversation about a card — from the card on screen, or by asking out loud during a walk — your microphone audio is streamed through our server to OpenAI’s live voice model while it lasts, with the card’s text and fields, a description of its picture, and answers you previously gave to it. We keep no audio and no transcript; the transcript is saved on your own devices only.
- The card and your answer are sent to be graded. The question, the expected answer, and the transcribed text of what you said go to our server, which passes them to OpenAI to produce a verdict, a correction, and sometimes a sentence of context.
- Card text is sent to be spoken. To ask a question out loud, the text of that question is sent to be turned into speech.
- Pictures, only when you ask for a memory hook. If you press the hint button, the card’s available pictures, note fields, deck name and review counts are sent so the teacher can make a relevant memory hook.
On iPhone, AI tutoring stays off until you enable it in Walking tutor settings after reviewing this data sharing. You can withdraw that permission there at any time. Reviewing on screen remains available without AI.
We do not store the transcript of your answer, the grade, or the feedback. They pass through and are gone. OpenAI processes this content as our processor and does not use it to train models.
One thing that is cached, and why
Spoken questions are cached on our server, stored under a fingerprint of the text rather than under your name, so that a card asked by a hundred people is only ever synthesized once. That cache does contain the audio of card questions, which means card text is recoverable from it — we would rather say that plainly than claim we store nothing. It is not linked to you, and it is deleted when it stops being used.
Audio of the tutor’s feedback to you is never cached, because it is derived from what you just said.
The website’s three-card demo
The demo uses sample cards, without an account or access to your own collection. Speaking is optional: microphone access is requested only after you choose “Answer out loud”. Your browser may send audio to its speech recognition provider, according to that browser’s settings and privacy policy. You can type instead. This differs from the native app’s on-device speech recognition.
Your answer text and the sample card are sent through our server to OpenAI for grading. We do not store your audio or answer text. The tutor’s spoken questions and corrections are prerecorded sample audio, served as static files.
When you start a challenge, Cloudflare Turnstile checks your browser to prevent automated abuse. Cloudflare receives connection and browser information for this check; see its privacy policy. Our demo database stores a daily keyed hash of your network address, a hashed session token, the sample card IDs, timestamps and right/wrong verdicts. These records are deleted by a daily cleanup after two days, normally within three days. Aggregate daily and monthly counters, without network or session identifiers, are retained for up to four months to enforce usage limits. Session tokens expire after 15 minutes and are held in browser memory, not local storage.
We use this processing to provide the challenge you request and to protect the service from abuse. Sharing is optional: the share button prepares your score and a public challenge link for you to send. It does not share your answers or session token.
Your subscription
To know whether the tutor should run, we hold: a random identifier your app generates for itself, a hash of a secret it keeps in your device’s Keychain, your licence details from our payment provider, and counters of how much you have used — a number of cards and an amount of processing. Not what those cards were.
Website payments are handled by Lemon Squeezy, which acts as merchant of record and holds your billing details. In-app purchases on iPhone are handled by Apple under the individual developer account of Antoine Berton. We never see your card number.
For Apple purchases, we keep verified transaction identifiers, subscription expiry and refund status to provide access. Connecting a Mac stores a hash of a one-use code and the two installation identifiers. Codes expire after ten minutes and are removed by daily cleanup. Both devices share the subscription’s usage allowance.
We do not know your name or your email address unless you write to us.
What we deliberately do not do
- No analytics or telemetry SDK in either app — no Google, no Firebase, no Sentry.
- No advertising, no advertising identifiers, no profiling.
- No cookies on this website beyond what is needed to serve it.
- No sale of personal data or sharing for advertising. The service providers described above process data to operate the service.
- No use of your content to train models.
How long we keep things
Usage counters roll off on a schedule measured in months, and exist to enforce your plan’s limits and to stop abuse. Records of a purchase are kept as long as accounting and tax law requires. Cached question audio is removed once it is no longer used.
Your rights
Under the GDPR you may ask for access to the data we hold about your installation, its correction, or its erasure, and you may object to processing or ask for it to be restricted. Write to privacy@agentkey.dev and we will act within one month.
Because we identify your installation only by a random identifier, we may need you to send it from inside the app to be sure we erase the right one — that is a consequence of not knowing who you are, and we think it is the right trade.
You also have the right to complain to your data protection authority. In France that is the CNIL.
Where the data is
Our database and cache run in the European Union. Grading and speech are performed by OpenAI, which may process the content described above outside the EU under the transfer safeguards in its data processing agreement.
Changes
If this policy changes in a way that affects what leaves your device, the app will tell you before it takes effect, rather than a date on this page changing quietly.