GUI Computer Use
FERAL can see your screen and interact with any desktop application using Anthropic-style computer-use primitives: screenshots, mouse clicks, typing, key combos, scrolling, and window management.Permission Requirements
macOS
- Open System Settings → Privacy & Security → Accessibility
- Add your terminal app (Terminal, iTerm2, or the app running FERAL)
- Also grant Screen Recording permission for screenshots
Linux (X11)
No special permissions needed on X11. Wayland requiresxdg-portal or running FERAL in an X11 session.
Install one of: gnome-screenshot, scrot, or imagemagick (for the import command).
Windows
Run FERAL as Administrator if clicking in elevated windows. No special setup otherwise.Rate Limits
FERAL enforces a configurable rate limit on GUI actions to prevent runaway automation:{"success": false, "reason": "rate_limit_exceeded"}.
Coordinate Scaling (Retina / HiDPI)
VLMs see a screenshot image (max 1920px wide). On Retina/HiDPI displays, the physical screen is larger. FERAL automatically detects the DPI scale factor and converts coordinates:- macOS: Queries
NSScreen.backingScaleFactor()(typically 2.0) - Linux: Reads
GDK_SCALEenvironment variable - Windows: Falls back to 1.0
Troubleshooting
- “Screenshot capture failed” — Check that Screen Recording permission is granted (macOS) or that
scrot/gnome-screenshotis installed (Linux). - Clicks land in the wrong spot — Usually a DPI mismatch. Check the
dpi_scalevalue in screenshot responses. Override withGDK_SCALEon Linux. - “rate_limit_exceeded” errors — Increase
FERAL_GUI_MAX_ACTIONS_PER_Sor slow down the automation loop. - Typing non-ASCII fails — Install
pyperclip. FERAL uses clipboard paste for non-ASCII text. - Window focus doesn’t work — macOS needs Accessibility permission. Linux needs
wmctrlorxdotool.
