Privacy Policy
Relais turns a phone you own into a private AI inference node: it runs language models on the device and serves an OpenAI-compatible API on your local network. It is built so that your prompts, documents, audio, and images never leave hardware you control — with one exception you have to opt into: reporting AI output to us, described below.
The short version
- All AI processing happens on your device. Prompts, chat history, transcribed audio, analyzed images, generated images, generated speech, and retrieval documents are processed by models running locally. None of it is sent to us or to any cloud AI service.
- We collect nothing you don't choose to send us. Relais has no analytics, no telemetry, no crash reporting, no ads, no tracking, and no account system. The one exception is reporting AI output — if you flag something and check "also send to developer," that report reaches us. It's off by default and decided per report. See "Reporting AI output" below.
- You control the network surface. The API is served only on your local network, protected by an access key generated on your device. You start and stop it.
What the app stores on your device
- Models you download (multi-gigabyte files, in the app's private storage).
- Configuration, including the node's access key and — only if you provide one — a Hugging Face access token. Both are stored in Android's encrypted preferences.
- Content you put into features: chat/session state, prompt templates, documents you ingest for retrieval (RAG), and batch jobs. All local, all deleted when you clear the app's data or uninstall.
When the app talks to the internet
Relais makes no background connections to us — there is nothing to connect to. It reaches the internet only for these specific, user-visible purposes:
- Downloading models. When you pick a model, the app fetches it from
Hugging Face (
huggingface.co) or, for the built-in Gallery catalog, from Google's model hosting (dl.google.com). The optional text-to-speech voice is fetched from the sherpa-onnx project's GitHub releases (github.com). Model searches you type in the model selector are sent to Hugging Face as search queries. - Your Hugging Face token (optional, for license-gated models) is sent
only to
huggingface.coas authentication on those downloads, and nowhere else. - Model catalog + update check. The app fetches its curated model list from
the upstream project's GitHub repository (
raw.githubusercontent.com) and checks GitHub (api.github.com) for a newer app release. These are plain HTTPS fetches; they carry no personal data beyond what any web request exposes (your IP address). - Destinations you configure. If you set up a webhook for batch-job results, or load an agent skill from a URL, the app sends/fetches exactly what you configured, to the destination you chose. These features are off until you configure them.
Requests from your own apps and devices to the node's LAN API stay on your network; the node also announces itself on your local network via mDNS (service name and port — no personal data) so clients can discover it.
Reporting AI output
Every AI reply has a REPORT option, so you can flag output you think is wrong or offensive without leaving the app. Reporting always saves the flag on your device — you can review everything you've flagged in the app's control panel — and that alone sends nothing anywhere.
If you also check "also send to developer" (unchecked by default; a separate
choice every time), that one report is sent to us at report.ventouxlabs.com, over
HTTPS. What's sent:
- The flagged output and any note you typed explaining what was wrong. If you type a name, email address, or other personal detail into that note, it is sent as you typed it — we don't scan for or strip personal information. Please avoid including it unless it's relevant to the report.
- Which reason you picked, and which part of the app the output came from.
- The model and processing mode that produced the output, if available.
We keep a report for 180 days, then it's deleted automatically. We also keep a scrambled (hashed), non-reversible version of your IP address, only to stop one device from flooding the report system — never your actual IP address, and never linked to what you reported. It's deleted an hour after your last report; if you report again within that hour, the hour restarts from that report, so it lasts no longer than you keep reporting. We use reports only to understand what's going wrong and improve moderation; we don't use them for anything else, and we don't sell or share them, though our hosting provider (Cloudflare) processes reports and the identifier on our behalf to run this service.
Permissions, in plain language
- Microphone / Camera — used only when you actively use voice input, audio transcription, or image analysis; the audio/images are processed on-device and are not uploaded anywhere.
- Notifications — so the always-on node can show its status.
- Network/Internet — for the model downloads and LAN serving described above.
- NFC — only if you enable NFC workflow tags.
- Start at boot / keep awake — so the node can run unattended, if you enable it.
Data sharing and selling
We don't sell your data, and we don't share it with anyone else. A report you choose to send us stays with us, for the moderation purpose described above — it never goes further. The app sends data to third parties only in the user-directed cases listed above (Hugging Face downloads with your own token; webhooks and skill URLs you configure). Those services' own privacy policies apply to what they receive.
Data retention and deletion
Almost everything lives only on your device. Delete any of it at any time via the in-app controls, Android's Clear data, or by uninstalling the app. The one exception is a report you've chosen to send us: it's retained for 180 days as described above, and if you want it deleted sooner, contact us (below) with enough of the report's content for us to find it — we don't keep any way to look reports up by who sent them.
Children
Relais is developer/infrastructure tooling and is not directed at children.
Changes
We'll update this policy if the app's data behavior changes, and note changes in the app's
changelog. The canonical source of this page is docs/privacy-policy.md in the
app's repository.