Vision Squeezer

This server attacks the vision API billing problem most developers don't notice: LLMs charge by pixel dimensions, not file size, and every provider has different tiling math. It exposes tools to preprocess images before they hit Claude, GPT-4o, Gemini, or other vision models by cropping padding, snapping to tile boundaries, and simulating each provider's token calculation algorithm. The CLI supports auto mode for text detection, quality binary search to hit SSIM targets, and batch processing with JSON output. You'd reach for this when you're burning tokens on screenshots with solid borders or images that spill into extra tiles by a few pixels. Install via the interactive npx script, Claude Code plugin marketplace, or as a standalone Rust binary.