Privacy and permissions

What stays on your computer, what can leave it, and why each permission is asked for.

What runs on your computer

  • Recording screenshots, voice notes, meetings and walkthroughs, and the screen context that goes with them
  • Transcription: Apple’s own on-device transcriber on macOS 26 and later, Whisper on Windows, Linux and earlier macOS. Either way the audio never leaves your computer.
  • Telling speakers apart and recognising voices: built in on macOS, and on Windows and Linux once you add the optional Speaker names download
  • Reading the text in screenshots: Apple Vision on macOS, Windows’ own text recognition on Windows, and on Linux the screen text models downloaded in setup
  • Sorting by rules and by similarity: Apple’s language model on macOS, and a small similarity model, downloaded in setup, on Windows and Linux
  • Short titles from a model on your computer, if you use one: Apple’s on-device model on a Mac that has it, or the optional Smart titles and sorting download on Windows and Linux
  • Search, activity lines, redaction, context packs and the MCP server

Where a part needs a download, the app offers it in setup or Settings, and shows what your computer has. Nothing on this list sends what you capture anywhere, on any system.

Your captures, settings and speech models are stored in:

macOS~/Library/Application Support/Startwell Capture
Windows%APPDATA%\Startwell Capture
Linux~/.config/Startwell Capture

Packs go to ~/Startwell Capture on macOS and Linux and %USERPROFILE%\Startwell Capture on Windows. There is no account and no sync. Startwell cannot see them.

Your OpenRouter key

If you add one, it is stored encrypted in the folder above, and the key that unlocks it is held by your system: the Keychain on macOS, Windows’ data protection for your account (DPAPI) on Windows, and your keyring (Secret Service, such as GNOME Keyring or KWallet) on Linux. On a Linux desktop with no keyring running, the key is only obscured, not encrypted.

What can leave your computer

Only these, and only the first two without you setting something up:

  • Update checks. The app fetches new versions from capture-updates.startwellventures.com. That server sees your IP address and app version, like any web server.
  • Model downloads. Where your system has nothing of its own for a job, the app downloads an open model once: Whisper for speech, the similarity model for sorting, the screen text models on Linux (and on Windows without a text recognition language), and any optional download you add, such as Smart titles and sorting or Speaker names. They come from their publishers (Hugging Face, or the ocrs project’s storage on Amazon S3) or from Startwell’s copy beside the app updates. On Linux, walkthrough video also needs Cisco’s OpenH264 codec, which only comes from Cisco; the row for it is in Settings, Capture. These are downloads only: no audio, text or images are sent, and the models need no network after them.
  • AI requests, if you add an OpenRouter key. What the feature needs goes from your computer to OpenRouter and the model you picked, under their terms: text such as a transcript to summarise, and, when a screenshot is summarised, the image itself. Never audio.
  • Walkthrough uploads you choose to make for a share link. The reviewed pack goes to Startwell’s storage on Cloudflare and anyone with the link can open it. See Privacy for how long it is kept.
  • Topic outputs you set up: a webhook that receives a topic’s material, or a folder it is written to.

Permissions, and why

On macOS, each of these is asked for once, the first time a feature needs it:

PermissionUsed for
MicrophoneVoice notes, meetings and walkthrough narration you start.
Screen RecordingScreenshots, walkthroughs, the other side of a call, and the screenshots saved during voice notes and meetings. macOS applies it only after the app is quit and reopened. Without it, recordings keep their words and activity lines and simply have no images.
Accessibility⌘⇧E clipping the text you have selected, window titles, and the names of what you click during a recording. It never reads what you type.
AutomationAsking your browser for the current page address, per browser, when you capture a link or record.
Speech RecognitionOn-device transcription.

Check them in the app under Settings, Permissions, and change them any time in System Settings, Privacy & Security.

Windows asks for nothing up front. The microphone is a switch in Settings, Privacy & security, Microphone, which has to allow desktop apps. On Linux nothing is asked on X11. On Wayland the desktop asks which screen to share, and you choose what it sees. For the screenshots saved during voice notes and meetings, choose it once in Settings, Capture; after that it doesn’t ask again, unless your desktop can’t remember the choice.

Activity lines

While something records, the app notes which app and window is in front, the page address and title, the document open in an editor, and what you clicked, by its name. Typed text is never recorded.

On Linux with Wayland, the system keeps one app from watching another. Window names come through where the desktop offers a way (KDE and wlroots desktops such as Sway list their windows; on GNOME it is the accessibility service). Clicks in other apps, and the names of what you click, are never seen there. Page addresses come through the browser add-on. On X11 all of it is recorded, with the accessibility service running.

Every capture carries context

This is a tool for working with AI, so a capture keeps as much context as it can, on your computer. These are always on, with no switch in Settings:

  • Screenshots during voice notes and meetings. While one records, a screenshot is saved when the screen changes, without video. Email addresses, card numbers and phone numbers are blurred wherever the screen’s text can be read; where it can’t, you approve each screenshot before it goes into a pack. A screen with a password field in use is left out where the system can tell. Startwell Capture’s small panels are kept out on macOS and Windows; on Linux they can appear.
  • Recent activity. The last two minutes of activity lines are held in memory, so a screenshot, note or link can say what led up to it. They reach the disk only as part of a capture that uses them, and are gone when you quit. Nothing is polled while you are idle.
  • Click names and the green click ring, live words and spoken markers, tidy transcripts, fixes for mis-heard names, and packs on disk. Typed text is never noted.
  • The clipboard. Every finished capture is copied as one paste for an AI agent, in text: the words, clicks, pages and links, the text read from the screen, and the path of every screenshot. It stays on your computer and goes only where you paste it.
  • A small card offers to record a call it detects. It never records by itself unless you turn on Start recording calls without asking in Settings, Capture.

You can still limit what is kept:

  • The small screen icon on a voice note’s or meeting’s recording panel stops screenshots for that recording and drops what was saved.
  • Withhold a permission. Without Screen Recording there are no screenshots; without Accessibility there are no click names. The Inbox says so and has an Allow button if you change your mind.
  • A new walkthrough’s panel has a Record video too switch for that walkthrough, and in review you can leave steps out or blur them.
  • Decline a call card and that call is not recorded.