The screenshot tool built for Codex and GPT vision agents. Snip a region and it's auto-compressed to the vision model's optimal size, then copied to your clipboard.
Why it matters for Codex: OpenAI bills vision input by the image's pixel dimensions, tiled. A full-screen Retina grab burns thousands of tokens before the model reads a thing. AgentShot caps the long edge at 1568px and JPEG-encodes it, so you stop paying for pixels the model never uses.
Measured, not theoretical — same screenshot, real OpenRouter API calls (bench/run_openrouter_tokens.py):
| Model | Original (757 KB, 3024×1964) | AgentShot (174 KB, 1568×1018) | Saved |
|---|---|---|---|
| GPT-5.5 | 7083 input tokens | 1896 input tokens | −73% |
curl -fsSL https://raw.githubusercontent.com/interesting-vibe-coding/agentshot/main/install.sh | bashInstalls to /Applications and launches. Grant Accessibility once in onboarding. The first capture asks for Screen Recording — allow it, then reopen AgentShot.
Press ⌘⇧2 to start. Click once for full screen, or drag to select a region. A preview appears:
C/↩— copy the compressed image (label showsoriginal KB → compressed KB)⇧C— copy the originalEsc— cancel
Paste into Codex or any model billed by image resolution. Shortcut, quality tier, and launch-at-login live in the menubar (📸).
OpenAI bills vision by pixel area (tiled). AgentShot's pipeline: downscale to ≤1568px long edge → JPEG q0.82 → enforce ≤⅓ the original size → hard cap at 1000 KB. Reading accuracy is unaffected — only token count drops (benchmarks).
Note on Claude: Anthropic normalizes images server-side before billing, so for Claude the token cost is already fixed regardless of what you send — AgentShot saves no tokens there. The win is real on resolution-billed models (GPT/Codex).
Homepage & benchmarks: https://interesting-vibe-coding.github.io/agentshot-site/ · MIT License · part of interesting-vibe-coding