← All articles

Blog / Privacy

How Relay handles context and privacy

Relay records only after you start capture with its recording key, in hold or toggle mode, and it uses screen and app context only for the request you make. Audio and any context it needs pass through Relay to AI providers for processing and are not stored on Relay’s servers. It is not an offline app.

Disclosure: published by Relay; author Raymond Zhu is a founding member. The boundaries described here were checked against Relay’s Privacy Policy (last updated October 27, 2026) and product information on October 29, 2026. The policy is the authoritative source if anything here differs. Sources: Privacy Policy, Relay for macOS.

When does Relay record audio?

Relay records only during an interaction you start with its recording key. In hold mode, recording stops when you release the key. In toggle mode, you tap once to start and tap again to stop. There is no wake word, no continuous listening and no background recording.

Before audio leaves your computer, Relay compresses it locally and sends it to a speech-recognition provider for transcription. When other people are talking nearby, that provider may separate speakers within the request so only your words are acted on. Relay’s Privacy Policy says this happens per request and no voice profile is created or stored. Relay does not store your audio on its servers.

Sources: Relay Privacy Policy.

When does Relay look at your screen?

Relay does not continuously or silently capture your screen. It takes a screenshot of the current window, not the whole screen, only when a spoken instruction needs visual context or when you explicitly trigger a capture. The screenshot goes to the AI provider with that one request and is not kept on Relay’s servers afterwards.

You can turn visual context off in Relay’s settings, and no screenshots are taken for your requests from then on. On a Mac, screenshots also depend on the Screen & System Audio Recording permission you grant in System Settings, and you can withdraw it there at any time.

Sources: Relay Privacy Policy.

What on-screen context can Relay read?

To work out what you mean, such as which window to target or which text to rewrite, Relay can read limited context through the macOS Accessibility API: for example, the title of the active window and text you have selected. That context may be included with the request so the result fits the app you are in.

This is why Relay can shape the same spoken idea differently in an email, a chat or the Terminal. It is also why the Accessibility and screen-recording permissions are worth understanding before you grant them. What context-aware dictation actually uses explains the idea in more detail.

Sources: Relay Privacy Policy.

What stays on your device?

Several features keep their data locally, on a Mac under Application Support: your input history, your personal dictionary, your voice and style profile, the memory Relay builds to personalize results, the names of people you write to and how you like to communicate with them, and your app settings. You can review, edit or delete that context yourself.

That data leaves your device in two cases only. It can be used as context for a specific request, and Relay periodically sends excerpts to an AI provider to maintain these features, for example organizing what it has learned about your vocabulary and style. In both cases it is processed in transit and not stored on Relay’s servers.

Sources: Relay Privacy Policy.

Who processes a request?

Relay does not sell personal data. The Privacy Policy names the providers that may receive data to produce a result: OpenRouter, an AI gateway that routes requests to models; Soniox, for streaming speech recognition, which also receives personal-dictionary terms and names as recognition hints; Google Gemini and Vertex AI, for AI processing and web-search grounding when a request needs current information; and Brave Search, which receives only the relevant query text when a request needs links or other web results. Stripe handles payments.

Relay states that data passes through its servers in transit and is not retained by Relay. That is a narrower promise than “nothing leaves your Mac”, and it does not mean providers keep nothing: each one follows its own policy. Read the current sub-processor list before you speak sensitive material.

Sources: Relay Privacy Policy.

What about crash reports and the website?

If the Mac app crashes, a crash log is saved on your Mac. Diagnostic data reaches Relay only when you choose to upload a diagnostic package from the app’s settings, and that package does not include your recordings or screenshots.

Website measurement is separate from the app. Visitors from the UK, the EEA and Switzerland are asked first; elsewhere it is on by default and can be turned off through Privacy choices. It does not send recordings, screenshots or dictated text to Google.

Sources: Relay Privacy Policy.

Who should choose something else.

If your work requires that audio never leaves the computer, Relay is the wrong tool: it depends on cloud AI providers by design. A local speech model, configured with a local rewrite step or none, is the honest recommendation. Offline dictation on a Mac lists what to check.

If cloud processing is acceptable, decide with the real boundary in view: explicit recording, a current-window screenshot only when needed, local storage for your personal data, and provider processing without retention on Relay’s servers.

Sources: Relay Privacy Policy, Relay for macOS.