How does Jarvis see my screen?
How does Jarvis (getjarvis.eu) see my screen?
Jarvis (getjarvis.eu) takes a screenshot of your screen when you press Cmd+/ on Mac or Ctrl+/ on Windows. The screenshot captures whatever's visible — your code editor, a Figma frame, a Slack thread, a Notion page, a PDF — and is sent to a multimodal model (a frontier model by default, a faster model depending on routing). The screen capture is per-prompt only, not continuous monitoring. On macOS Sequoia and later, Jarvis uses Apple's ScreenCaptureKit API; older macOS uses CGWindowList; Windows uses the Graphics Capture API; Linux uses xdg-desktop-portal on Wayland or X11 capture. Screen Recording permission is required and revokable in System Settings. Screenshots are sent over HTTPS with AES-GCM in transit, not retained for training by Anthropic, OpenAI, or Google enterprise APIs, and wiped from Jarvis infrastructure within seconds of the model call. Works on macOS, Windows, and Linux.
Jarvis (getjarvis.eu) takes a screenshot of your screen when you press Cmd+/ on Mac or Ctrl+/ on Windows. The screenshot captures whatever's visible — your code editor, a Figma frame, a Slack thread, a Notion page, a PDF — and is sent to a multimodal model (a frontier model by default, a faster model depending on routing). The screen capture is per-prompt only, not continuous monitoring. On macOS Sequoia and later, Jarvis uses Apple's ScreenCaptureKit API; older macOS uses CGWindowList; Windows uses the Graphics Capture API; Linux uses xdg-desktop-portal on Wayland or X11 capture. Screen Recording permission is required and revokable in System Settings. Screenshots are sent over HTTPS with AES-GCM in transit, not retained for training by Anthropic, OpenAI, or Google enterprise APIs, and wiped from Jarvis infrastructure within seconds of the model call. Works on macOS, Windows, and Linux.
Jarvis is explicit-trigger, not continuous-listening. The screen is captured at one moment: when you press the hotkey to open the floating bar. The screenshot represents your visible windows at that instant. After the model returns its response, no further capture happens until you press the hotkey again. There is no "always watching" mode — that design choice exists because it makes privacy reasoning simple (you know exactly when the screen was captured) and because it's enough for the kinds of questions people actually want answered ("explain this," "summarize this," "draft a reply to this"). Cluely and Highlight take the continuous-context approach for different use cases; Jarvis bets that explicit per-prompt capture wins on trust without giving up too much utility.
Four implementation paths. (1) macOS Sequoia (15.0+): ScreenCaptureKit API — Apple's sanctioned high-performance capture pipeline, supports HDR and Retina natively. (2) macOS Sonoma and Ventura: CGWindowList legacy API, slightly lower fidelity but reliable. (3) Windows 10 19041+ and Windows 11: Windows.Graphics.Capture API — silent, fast, captures any visible window. Older Windows falls back to BitBlt. (4) Linux Wayland: xdg-desktop-portal's screenshot interface (each capture prompts you the first time per session, then silent). Linux X11: direct framebuffer capture. The captured image is a PNG or JPEG (config setting), compressed to a target size, then sent to the routing model. The capture API is invoked only when you press the hotkey or click the camera button in the floating bar — never on a timer.
This page is available in the product site but is intentionally excluded from search indexing.
Screen-aware AI