Less friction · for AI agents

AgentShot

The screenshot tool built for AI agents.

Snip, auto-compress to the vision sweet spot, on your clipboard. Up to 81% fewer image tokens, no loss in reading accuracy.

Before and after: same screenshot, 62% fewer image tokens

Install

One line in your terminal. Click to copy.

macOS · one line
Downloads AgentShot, installs it to /Applications, and launches it. First capture asks for Screen Recording permission — allow it, then re-open AgentShot.

Why compress at all?

Vision models bill by pixels, not bytes. A screenshot's token cost is roughly its area divided by a constant — no amount of clever PNG encoding changes that.

AgentShot resizes the long edge down to 1568px — Claude's vision sweet spot — before the image ever reaches your clipboard. Fewer pixels, fewer tokens, same readable text.

tokens ≈ width × height ÷ 750 long edge → 1568px

The numbers

Reading accuracy holds. Token cost drops by up to 81%.

Reading accuracy before and after compression
Screenshot reading accuracy is unchanged — Claude 0.86 → 0.88, GPT 1.00 → 1.00.
Image token count before and after compression
Up to 81% fewer image tokens — docs 3793 → 1908 (−50%), full 4K screen 4601 → 878 (−81%).

Where it actually pays off

AgentWhat you save
Claude Code Saves bandwidth on upload — smaller payloads, faster round-trips.
Kiro Large images can error out — real token savings and no more failed uploads.
Codex & others Real token savings, straight off the clipboard.
Copied to clipboard