Ferrisgrid Use

FerrisGrid Use

FerrisGrid is single-step. Do one observe or one action per call, then stop and inspect the Markdown result.

When a task says to use FerrisGrid only, do not use browser automation, page-source fetches, scraping, external search, or host GUI control to inspect or drive the target. Use FerrisGrid observations/actions and inspect the screenshot files returned by FerrisGrid.

If ferrisgrid is not installed on the host, install the published CLI first:

cargo install ferrisgrid-cli

Container Workspace

For the Linux Docker/noVNC workspace, keep the agent terminal on the host and run FerrisGrid inside the container:

docker ps --filter name=ferrisgrid-workspace --format '{{.Names}}:{{.Status}}:{{.Ports}}'
docker exec ferrisgrid-workspace ferrisgrid doctor
docker exec ferrisgrid-workspace ferrisgrid observe --max-image-edge 1280

If the container is not running:

docker run --rm -d --name ferrisgrid-workspace -p 6080:6080 -v "$PWD:/workspace" ferrisgrid-linux-workspace

The human can view the same background desktop at:

http://127.0.0.1:6080/vnc.html?autoconnect=1&resize=scale

Container-specific checks:

• Run ferrisgrid doctor inside the container before task work; expected backend is native-linux-x11.
• If .ferrisgrid/ was deleted, the next observe recreates it automatically.
• Use docker exec ferrisgrid-workspace ferrisgrid ...; do not run host cargo run when the target is the container desktop.
• If host curl to noVNC fails from the sandbox but docker ps shows 6080 mapped, retry with approved local network access before assuming noVNC is down.
• Leave the container running when the user wants to inspect the workspace; stop it only when cleanup is requested.

Observe

Capture all screens by default:

ferrisgrid observe

Capture one screen only:

ferrisgrid observe --screen-id screen-1

Use smaller images for faster LLM vision:

ferrisgrid observe --max-image-edge 1280

Coordinates

• Use coordinate_mode: normalized-1000.
• x=0 y=0 is the top-left of the target screen.
• x=1000 y=1000 is the bottom-right.
• Coordinates are screen-local, not virtual-desktop coordinates.
• If multiple screens are listed, include screen_id.
• Read the returned screenshot= path and optional metadata= path after every call.

Act

Write compact Markdown to an action file, not JSON or prose:

# Update .ferrisgrid/action.md with one compact Markdown action, then run:
ferrisgrid act --file .ferrisgrid/action.md

For the container workspace:

docker exec ferrisgrid-workspace ferrisgrid act --file .ferrisgrid/action.md --max-image-edge 1280

Before each action, replace the whole action block. Do not patch only one field if that can leave duplicate keys such as two x: lines. If in doubt, read the file first:

sed -n '1,80p' .ferrisgrid/action.md

Example .ferrisgrid/action.md:

status: action
action: click
screen_id: screen-1
x: 500
y: 500
button: left
wait_after_ms: 500

Use wait_after_ms on UI-changing actions so FerrisGrid captures the updated UI after the action settles. It waits after executing the action and before taking the returned screenshot. Omit it for pointer moves or actions where an immediate capture is useful. Maximum: 30000.

Suggested wait_after_ms values:

• 150-300: simple desktop focus changes, opening menus, toggles, hover-triggered UI.
• 300-700: local app tab switches, list selection, expanding rows, lightweight modals.
• 700-1200: browser SPA route changes, message/thread switches, filtered lists, autocomplete results.
• 1200-2500: form submissions, sign-in transitions, search results, uploads that start background work.
• 2500-5000: full page loads, slow web dashboards, OAuth redirects, network-heavy state changes.
• 5000-15000: installers, app launches, downloads, operations that visibly show a progress state.

Useful actions:

status: action
action: click
screen_id: screen-1
x: 742
y: 611
button: left
wait_after_ms: 700

status: action
action: scroll
screen_id: screen-1
x: 500
y: 500
delta_y: -720
wait_after_ms: 300

status: action
action: type
text: hello

status: action
action: press_key
key: enter
wait_after_ms: 1200

Browser Tips

• If a page does not scroll, first click inside the page content, then retry scroll from that area.
• Browser overlays, galleries, and find boxes can capture input; close them before continuing.
• For product pages, cmd+f with a price symbol or keyword can jump to hidden details faster than repeated scrolling. Close find with escape before reading the final view.
• On Linux/X11 use ordinary web shortcuts such as ctrl+l, not macOS cmd shortcuts.
• If the address bar is already focused, type the URL directly, then press_key enter.
• On retail sites, dismiss cookie banners before country or location modals if clicks appear to do nothing.
• If a campaign/category feed hides product names or prices, use site search for a broad term to get a normal product grid.
• When grid product names are truncated, open the product detail page to confirm the full name and price.
• FerrisGrid overlays can obscure small text; rely on several screenshots, scroll slightly, or open detail pages rather than guessing.

Troubleshooting

• If observe works but act through stdin returns CoreGraphics returned no displays, use the default --file .ferrisgrid/action.md path.
• If docker exec ... ferrisgrid observe works but noVNC is not visible from the host, check docker logs ferrisgrid-workspace for the WebSockify line and confirm port 6080 is mapped in docker ps.
• If Chromium opens smaller than the Xvfb display, check the container entrypoint --window-size values and verify with xdotool search --onlyvisible --class chromium getwindowgeometry %@.
• If input lands in the wrong place, run xdotool getmouselocation inside the container after a move_mouse smoke test to confirm coordinate mapping.

Loop

1. Run observe.
2. Inspect returned screenshots.
3. Write exactly one action to .ferrisgrid/action.md and run act --file, adding wait_after_ms when the action should change visible UI.
4. Inspect the post-action screenshot path.
5. Repeat until done.

Finish with:

status: done
reason: task complete

Source & license

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

• Author: BrunoV21
• Source: BrunoV21/FerrisGrid-CLI
• License: MIT
• Homepage: https://brunov21.github.io/FerrisGrid-CLI/

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