MARSHAL STACK What's new How it runs Console Studio · soon Docs ENRURO Contact

← Marshal Stack

Marshal AI — Privacy Policy

Last updated 2 November 2026

Marshal AI is a voice assistant with apps for iPhone, Apple Watch, Android and Windows. The apps do not answer on their own: they talk to an assistant server (the “core”) whose address and access token you enter in the app — either our core at ai.marshalstack.com, hosted by Marshal Stack in the EU (Hetzner, Germany), or a core you run yourself. Your requests and plans go only to that server, over TLS; the few other services an app or our core contacts are named below. This page describes what the apps and our core do with your data. If you use a core run by someone else, that operator decides how the data on their server is handled. On the iPhone and Android you can also use the app without any core, with your own AI provider key — see “Your own AI key” below.

Microphone and speech

On every device the microphone is on only while you are talking to the assistant, and the app shows when it is listening. How your speech becomes text depends on the device:

  • iPhone — on the device itself, by Apple’s Speech framework (the default) or, if you choose it in Settings, by a Whisper model. The Whisper model is optional: the app asks first, then downloads it (about 140 MB) once from Hugging Face (huggingface.co) and from then on runs it on the device, offline. Only the model files are downloaded — Hugging Face sees an ordinary download request with your IP address, and no audio, text or other data of yours is sent. Only on devices or languages where Apple’s on-device recognition is not available, Apple’s speech recognition service processes the audio under Apple’s privacy policy. Apart from that case, audio from the iPhone is never sent to our servers or to any other party.
  • Android — by the speech recognition service built into your phone (usually Google’s), which may process the audio on the device or on its provider’s servers under that provider’s privacy policy. The app itself never sends audio to our servers.
  • Apple Watch and Windows — the app records only the phrase you are saying (16 kHz mono, up to 30 seconds) and sends that recording over TLS to the core you connected. The core turns it into text in memory with open-source speech recognition models running on the core itself — on our core GigaAM v3 (by SberDevices, MIT licence), with Whisper (faster-whisper) for English — no third party is involved. The audio is not written to disk or stored, and it is discarded as soon as the text is ready; only the recognised text goes on, exactly as a typed request. The watch uses the same server address and token as the iPhone. In the “Own AI key” mode the watch sends the recording to your iPhone instead, which recognises it as described for the iPhone above.
  • Spoken replies — when the app is connected to a core, Russian replies on iPhone, Apple Watch, Android and Windows are voiced by the core: the app sends it the text of the reply and plays the audio it returns. To make this audio our core may send the reply text to Google Cloud Text-to-Speech (the “Leda” voice). Google receives only that text — no recording of your voice and no account or device data — and processes it under the Google Cloud Privacy Notice. When Google is not used or does not answer, the core puts the reply together from pre-recorded Leda clips (fixed phrases, numbers and smart-home device names, recorded in advance — never from your requests) or voices it with open-source models running on the core itself: Silero “Albina” (MIT licence) for Russian and Supertonic 3 for English. English replies, and every reply when there is no core, the core does not answer or the “Own AI key” mode is on, are spoken on the device by its built-in system voice.

What the app sends to the core

Text and settings — plus, from Apple Watch and Windows only, the recording of the phrase described above, which the core turns into text and discards, and, for spoken replies, the text of the reply to be voiced:

  • Your requests — the recognised text of what you asked, with a conversation id and the device type.
  • Plans — events, tasks and shopping items you add, with their time, place and notes.
  • Settings — your assistant preferences, such as which briefings and alerts you want.
  • A push notification token — on iPhone and Apple Watch, issued by Apple, so the core can send you the morning briefing and alerts.
  • Your device’s time zone — its name (for example, Europe/Chisinau), sent with each request so that “tomorrow at 2” means 2 p.m. in your local time. The core uses it only to answer that request and does not store it; your plans keep their time as described above.

How answers are made

Simple commands — adding or reading plans, the time, weather, waves, project status — are handled by the core itself. Any other request is passed as text, through our Marshal Stack relay, to the Claude language model by Anthropic; to answer, the model may look up your plans and project status on the core. Anthropic processes this text to produce the reply under its privacy policy.

What our core stores

Your plans; the last 20 turns of each conversation, so the assistant can follow context; a log of requests and answers with their timing, used to fix errors and improve speed; answers waiting to be delivered; your settings; and the push tokens of your devices. A push token is removed as soon as Apple reports it invalid. Voice recordings are not stored — only the recognised text, as part of the conversation and the log. The audio of spoken replies is kept only briefly in memory, to reuse it, and is not saved.

Calendar, weather and other connections

The core talks to other services only for features that are switched on:

  • Calendars — iCloud Calendar (CalDAV) or Google Calendar and Tasks (OAuth), if connected. The core reads and writes events and tasks there with the credentials given to it; they are kept on the core and used for nothing else.
  • Weather and waves — from Open-Meteo, for a place set in the core’s settings. Your location is not used for this.
  • Project status — the state of your projects and servers from Marshal Stack, if you use it.
  • Notifications — on iPhone and Apple Watch, Apple Push Notification service delivers the briefing and alerts; their text passes through Apple. On Windows, notifications are shown by the app itself.
  • Voice of replies — Google Cloud Text-to-Speech receives the text of replies to be spoken, as described under “Spoken replies” above.

On your device

The access token is kept encrypted on the device only: on the iPhone in the system keychain, excluded from backups; on Android encrypted with a key from the Android Keystore; on Windows encrypted with Windows DPAPI. The server address is kept in the app’s settings. On the iPhone both are shared, through a keychain group, only with other apps from the same developer on the same iPhone, so they can connect to the same assistant. If you allow access to Reminders, the iPhone app copies your plans into a “Marshal” list in Apple Reminders so iOS can remind you; that list stays in your Apple account. Deleting the app removes its local data.

Smart home and speakers

If you connect a smart home, Marshal controls lights, climate, blinds, scenes and similar devices on your command:

  • Through the core — a home command goes to the core as text, like any other request, and the core controls the Home Assistant connected to it, with credentials kept on the core.
  • From the phone — in the “Own AI key” mode, or when the core cannot be reached, the app controls your home directly: Home Assistant (you sign in on Home Assistant’s own screen, so Marshal never sees your password; the sign-in token is kept encrypted on the device), Apple Home on the iPhone (only if you allow access to Home) and Google Home on Android (only if you connect it). Your device list and commands then go only between the phone and your home system, not to us.
  • Marshal speakers — when you add a speaker, the app passes the name and password of your Wi-Fi network to the speaker over Bluetooth; they are not sent anywhere else. To fill in the network name, the iPhone asks for location permission (iOS reveals the Wi-Fi name only with it), and Android 11 and older require it for Bluetooth scanning; your location itself is not used or sent. The app then adds the speaker to your Home Assistant with the key from its label. After its wake word, a speaker sends the phrase it hears to the core, through Home Assistant or directly, and the core recognises and answers it the same way as for Apple Watch above: the audio is processed in memory and not stored.

Health data (iPhone)

If you allow it, the app reads last night’s sleep, today’s step count and your resting heart rate from Apple Health, only to answer questions such as “How did I sleep?”. It only reads, never writes. This data is used on the iPhone and never leaves it: it is not sent to the core, to an AI provider or to us, and it is never used for advertising.

Your own AI key (iPhone and Android)

In the “Own AI key” mode the app does not use a core. The phone sends the text of your request, with the recent conversation (and, on the iPhone, the plans needed to answer), directly to the provider you chose — Anthropic (Claude), Google (Gemini) or OpenAI (ChatGPT) — using your own API key. No audio is sent to the provider. Before the key is saved, the app asks for your permission to send this data to that provider; without it nothing is sent. That provider processes it under its own privacy policy and your agreement with it; the data does not pass through us. The key is kept on this device only — in the iPhone’s keychain, or on Android encrypted with a key from the Android Keystore. On the iPhone your plans in this mode are kept in Calendar and Reminders, which the app reads and writes only if you allow access; they stay in your Apple account. On Android, plans need a core and are not available in this mode.

What we do not collect

The apps contain no analytics, advertising, attribution or crash-reporting SDKs and no third-party trackers. They do not read your contacts or photos, do not use your location (the only location permission is the one described under “Smart home and speakers”, used to read a Wi-Fi network name), and do not track you across other apps or sites. No data is sold, and none is shared for advertising.

Retention and deletion

There is no sign-up in the app: access to our core is given by an access token. Data on our core is kept while your access is active. To have it deleted — plans, conversations, the request log, settings and device tokens — write to the address below from the email you used to get access; we delete it and revoke the token within 30 days. To stop using the app on a device, delete the app; this removes the token and the server address from that device.

Children

Marshal AI is not directed at children under 13.

Changes

If this policy changes, the updated version will be posted on this page with a new date above.

Contact

Questions and deletion requests: [email protected]

© 2026 Marshal Stack
Studio · soon Docs Privacy Terms Contact