skyvern

browserground adapter for Skyvern

Replace (or augment) Skyvern's frontier-vision grounding step with the local browserground specialist. Two patterns:

1. Replace — every grounding call goes local

from browserground_skyvern import ground

result = ground(screenshot_path, target)
bbox = result["bbox_2d"]
# pass bbox into Skyvern's DOM-mapping stage

2. Augment — local first, cloud fallback on low confidence

from browserground_skyvern import ground_with_fallback

def cloud_ground(img, tgt):
    # your existing Skyvern cloud-grounding entry point
    ...

bbox = ground_with_fallback(
    screenshot_path, target,
    confidence_threshold=0.55,
    cloud_fallback=cloud_ground,
)

The model returns a sequence-level confidence score (geometric mean of per-token probabilities). Tune confidence_threshold against your own eval set — start at 0.55 and adjust.

Daemon mode (recommended for production)

browserground serve --http :8401 &
export BROWSERGROUND_HTTP=http://127.0.0.1:8401

The adapter now POSTs to the daemon (no subprocess spawn per call).

What this saves

	pure Skyvern (cloud vision)	with browserground
Vision tokens / step	1500+ multimodal	0 (local)
Cost / step	$0.01–0.05	$0
Screenshots leave machine	yes	no

Recipe + numbers: renezander030/browserground.