This page is the live, public-facing companion to the Claudia Privacy Policy. It enumerates every third-party service Claudia can be configured to use, what data flows there, and where to read each provider's own privacy policy.
The Privacy Policy itself does not embed retention windows or provider-specific terms, as those change quarterly and would make the policy stale. This page is updated whenever Claudia adds, removes, or changes a provider.
Last updated: August 12, 2026
How to read this
If you do not configure a provider, no data flows to that provider. Most rows below describe potential flows that activate only when you save an API key for that service or grant an OAuth scope. Where a row does not need a key, that is stated in the row.
AI model services (chat, voice, photo analysis)
Data sent: the message you typed (or the transcript of what you spoke), the minimum context Claudia needs to answer (recent chat history, memory facts only when relevant, attached photos or documents when included). Photos you attach are sent to the model you selected so it can analyze them. Claudia does not generate images.
Before each turn, Claudia runs a short classifier pass that decides which tools the turn may need. Since August 3, 2026, that pass goes only to the cheap model of the provider you already selected for the turn. It is never routed to a different company. If you are running Claudia fully on-device, the classifier runs on-device too and no network call is made for it.
| Provider | What goes there, and when | Provider privacy policy |
|---|---|---|
| Anthropic | Chat (Claude family) and the pre-turn classifier, when Claude is your selected model | https://www.anthropic.com/privacy |
| OpenAI | Chat (GPT family) and the pre-turn classifier when GPT is selected; optional voice transcription (Whisper) and optional voice synthesis (TTS), both off by default | https://openai.com/policies/privacy-policy |
| Google (Gemini) | Chat (Gemini family) and the pre-turn classifier, when Gemini is selected | https://policies.google.com/privacy |
| Groq | Chat inference and the pre-turn classifier for selected models; optional voice transcription (Whisper on Groq), off by default | https://groq.com/privacy-policy/ |
| Cerebras | Chat inference and the pre-turn classifier for selected models | https://www.cerebras.ai/privacy-policy |
| xAI | Chat (Grok family) and the pre-turn classifier, when Grok is selected | https://x.ai/legal/privacy-policy |
| OpenRouter | Routing layer that can dispatch chat and the pre-turn classifier to additional model providers when selected, including photos you attach for analysis. OpenRouter passes your request on to the underlying model provider you picked. | https://openrouter.ai/privacy |
| Moonshot AI (Kimi) | Chat inference and the pre-turn classifier, when a Kimi model is selected | https://www.moonshot.ai/privacy |
| NVIDIA | Chat inference for selected models via NIM. The pre-turn classifier does not run against NVIDIA; on this provider it is skipped with no network call. | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/ |
| Cloudflare (Workers AI) | Chat inference for selected models; optional voice transcription (Whisper on Workers AI), off by default. The pre-turn classifier does not run against Cloudflare; on this provider it is skipped with no network call. | https://www.cloudflare.com/privacypolicy/ |
| ElevenLabs | Voice synthesis only, and only if you turn on a cloud voice. Claudia sends the text of the spoken reply and receives audio back. On-device voice is the default. | https://elevenlabs.io/privacy |
| Hugging Face | One-time download of the on-device voice model file (Piper en_US-amy-medium, about 63 MB) on first voice use. The model is then cached locally and later voice synthesis is fully offline. Your message text is never sent to Hugging Face. | https://huggingface.co/privacy |
Voice transcription and synthesis (on-device default)
In v1, both voice transcription and voice synthesis run entirely on-device by default. Voice audio does not leave your device on that default path. No transcription provider key is required, and voice mode works with any chat-model key.
Voice transcription uses Apple's built-in speech recognition (SFSpeechRecognizer) on iPhone. You may optionally enable a cloud transcription provider in Settings → Audio, in which case the recorded audio is uploaded to that provider:
- OpenAI Whisper (via your OpenAI key)
- Groq Whisper (via your Groq key)
- Cloudflare Workers AI Whisper (via your Cloudflare key)
Cloud transcription is sometimes more accurate on rare words, accents, or noisy environments. The on-device default sends no audio anywhere, needs no network round-trip, and adds no API cost.
Voice synthesis (Claudia's spoken replies) uses an on-device neural voice model, Piper "amy" (en_US-amy-medium). Audio is generated locally and never sent to a server on the default path. The model file (an ONNX build of about 63 MB) is downloaded once on first voice use from Hugging Face (huggingface.co/diffusionstudio/piper-voices) and cached on your phone. Only the model file itself is fetched, and your message text is not sent to Hugging Face. You can optionally use a cloud voice with your own key in Settings → Audio:
- OpenAI TTS (via your OpenAI key),
gpt-4o-mini-tts. - ElevenLabs (via your ElevenLabs key),
eleven_turbo_v2_5.
On the cloud-voice path, Claudia sends the text of the reply to the chosen provider, the provider returns mp3 audio, and the audio is then played on your device. Voice provider keys are stored only in the iOS Keychain on your phone.
Android in v1: voice mode is not available on Android. We will ship Android voice after iOS launch.
Cloud connectors (removed August 10, 2026)
Removed August 10, 2026: earlier versions of Claudia could connect the Google accounts below, activated only when you explicitly connected them. Those connections were removed from the app on August 10, 2026, and the current app cannot connect a Google account. The table is kept for the historical record. If you connected an account in an earlier version, Sign out and remove API keys in Settings deletes any Google sign-in tokens still on your phone.
| Service | OAuth scope | Data sent / received | Provider privacy policy |
|---|---|---|---|
| Google Gmail (drafts only) | https://www.googleapis.com/auth/gmail.compose | Outbound: draft body, subject, recipient (after you confirm the draft card). Inbound: draft ID. Never sends mail. | https://policies.google.com/privacy |
| Google Gmail (read) | https://www.googleapis.com/auth/gmail.readonly (separate consent) | Outbound: search query when you ask Claudia to check email. Inbound: message metadata + snippets. | https://policies.google.com/privacy |
| Google Drive (read-only) | https://www.googleapis.com/auth/drive.readonly | Outbound: file search query, file ID. Inbound: file contents for text extraction. | https://policies.google.com/privacy |
Narrow public services (keyless, used by specific tools)
These run only when the matching tool is invoked. They do not receive your chat history.
| Service | What goes there, and when | Provider terms |
|---|---|---|
| Open-Meteo | Latitude, longitude, and a time range, when you ask about weather. Coordinates are either approximate device location (Claudia requests low-accuracy location, roughly city-scale, about 5 km) or the coordinates of a place name you typed. Open-Meteo also handles the place-name lookup behind a weather question. No key required. | https://open-meteo.com/en/terms |
| OpenStreetMap Nominatim | A place name or an address, when Claudia needs to turn it into coordinates. No key required. | https://operations.osmfoundation.org/policies/nominatim/ |
| OpenStreetMap Overpass | A category and a bounding area, when you ask what is nearby and Claudia looks up places from OpenStreetMap data. No key required. | https://operations.osmfoundation.org/policies/api/ |
| Link previews | When a message contains a link, Claudia fetches that URL directly so it can show a title and preview. The request goes to whoever runs that website, and it carries the URL itself. No PrentusAI server is involved. | Governed by the policy of the site you linked |
| Apple Maps | Claudia can hand a place or an address to Apple Maps as a deep link when you tap to open directions. Once you tap, the request is between your phone and Apple. | https://www.apple.com/legal/privacy/ |
Optional public services (require your own API key)
| Service | What goes there, and when | Provider privacy policy |
|---|---|---|
| Brave Search | Your search query, when a turn uses the web-search tool. Requires your own Brave Search key. Without a key the tool does not run. | https://search.brave.com/help/privacy-policy |
| Google Places | A place query, plus latitude and longitude if available, when Claudia looks up a business or point of interest. Those coordinates are either approximate device location (low-accuracy, roughly city-scale, about 5 km) or the coordinates of a place name you typed. Requires your own Google key. | https://policies.google.com/privacy |
App infrastructure and user-configured endpoints
These are not AI providers, but they are network calls, so they are listed here rather than left out.
| Endpoint | What goes there, and when | Policy |
|---|---|---|
Expo update service (u.expo.dev) | Claudia asks whether a newer version of the app's code is available, at launch, when you return to the app after switching away (throttled), and when you tap "Check for updates". The request carries the app version and platform. It never carries your messages, memories, or keys. This call happens whether or not you have configured any provider. | https://expo.dev/privacy |
| Webhooks you configure | If you paste a webhook URL into Claudia, Claudia posts to that URL when the matching action fires. You chose the destination, so you control what it receives and who operates it. | Governed by whoever runs the URL you entered |
Removed August 10, 2026: an optional "Claude connector" relay (claudia-connect.prentusai.workers.dev) was listed here between July 20 and August 10, 2026. The feature was discontinued, the server was deleted along with any data it held, and the app no longer contains a pairing flow. No PrentusAI-operated endpoint remains in the app.
What we do not do
- We do not run a PrentusAI backend that your chat data passes through. Each model call goes directly from your device to the named provider.
- We do not include analytics, crash reporting, or advertising SDKs.
- We do not use your data to train AI models. (The providers above each have their own training-data policies, so please read theirs.)
Retention
Retention is per-provider. Most providers retain inputs and outputs for a short window for safety review (typically 30 days) unless you opt out at the provider level. Please consult each provider's privacy policy linked above for the current details. We do not embed retention windows here because they change.
How to stop data from going to any provider
- Remove the API key for that provider in Settings → Intelligence → API Keys.
- Or remove ALL keys at once with Settings → Privacy & Memory → "Sign out and remove API keys."
- Google connections were removed on August 10, 2026. Use "Sign out and remove API keys" to delete any Google sign-in tokens saved by an earlier version.
- Or remove the webhook URL you configured.
After removing a key, Claudia cannot reach that provider until you re-enter the key. What a provider already received cannot be taken back by the app. To have it removed, contact that provider directly under their data-deletion process.
Updates
This page is updated whenever Claudia's set of supported providers changes. Material changes will also be noted in the Claudia release notes in the App Store.
For questions: team@prentusai.com