Screen Capture

Screen Capture (Wayland / Hyprland)

grim captures; slurp selects a region interactively; hyprctl + jq give geometry for non-interactive captures. Output goes to a file or to wl-copy.

Requires grim and slurp (both present on this machine). Wayland only — scrot /maim won't work here.

Full screen

grim ~/screenshot.png                 # focused output (or all, depending on setup)
grim -o "$(hyprctl monitors -j | jq -r '.[] | select(.focused).name')" ~/focused-monitor.png

Interactive region (you draw the box)

grim -g "$(slurp)" ~/region.png

Active window — no interaction

geom=$(hyprctl activewindow -j | jq -r '"\(.at[0]),\(.at[1]) \(.size[0])x\(.size[1])"')
grim -g "$geom" ~/active-window.png

Pick a specific window by class — no interaction

geom=$(hyprctl clients -j | jq -r '.[] | select(.class=="firefox") | "\(.at[0]),\(.at[1]) \(.size[0])x\(.size[1])"' | head -1)
grim -g "$geom" ~/firefox.png

To the clipboard (no file)

grim - | wl-copy --type image/png                    # full screen → clipboard
grim -g "$(slurp)" - | wl-copy --type image/png      # region → clipboard

Smaller / compressed (good for feeding to a model)

grim -t jpeg -q 75 ~/shot.jpg        # JPEG at quality 75
grim -t png -l 6 ~/shot.png          # PNG compression level

Reading it back

After saving, Read the image file to analyze it — that's how the agent "sees" the result. Keep captures in /tmp or a scratch dir and clean up after.

Gotchas

• Scaling: grim saves at the output's physical pixel resolution. If a

monitor has scale ≠ 1 (hyprctl monitors -j | jq '.[].scale'), image pixels = logical × scale. To turn a pixel you see in the screenshot into a cursor coordinate, divide by scale. This is the #1 mistake when clicking from a screenshot — see visual-computer-use-loop.

• Multi-monitor: without -o/-g, grim may stitch all outputs; pass -o

for one monitor or -g for an exact box.

• Timing: capture after the UI has settled; add a brief wait after launching

or focusing a window before grabbing it.

Pairs with

• hyprland-control — geometry of windows/monitors to capture.
• visual-computer-use-loop — capture → reason → act → re-capture.
• screen-recording — for motion instead of a still (separate skill).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hdprajwal
• Source: hdprajwal/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.