Tile alignment is the mechanism: vision APIs charge by tile, and an image that does not land on those boundaries pays for padding it does not use. sandbox_execute is the more aggressive option — crop and binarize down to the region that carries the answer, and the rest never becomes tokens at all.
An image preprocessing MCP server for vision models. It resizes to tile boundaries, strips padding and re-encodes so an image costs fewer tokens to look at, and it works with any agent that speaks MCP rather than a single provider.
- optimize_image — snap an image's dimensions to tile boundaries, remove padding and re-encode it for a vision API
- optimize_image_batch — up to 64 images in one call, each taking the same arguments; a failure returns an error entry for that image rather than sinking the batch
- sandbox_execute — run a sequence of atomic operations (crop, grayscale, binarize, resize, contrast, brightness) to extract only the region that matters
- get_savings_stats — the cumulative token and bandwidth savings from optimizations so far
Nothing to configure — clients run npx -y vision-squeezer over stdio, or npx vision-squeezer install to pick the client, method and scope interactively. In Claude Code the plugin path bundles the server together with the /vision-stats, /vision-doctor and /vision-upgrade skills. There is also a Rust crate and a Python package of the same name.
One command — npx vision-squeezer install