Introduction

Give your coding agent the words on your screen, not the pixels. Cheapshot reads screenshots, screen recordings, and PDFs with Apple’s Vision framework, redacts secrets before the text leaves your Mac, and keeps a ledger of the image tokens your agent did not have to pay for. No model in the loop. MIT. The source is on GitHub.

Words, not pixels

A screenshot costs roughly (width x height) / 750 tokens, and rides along on every later turn. A 4 megapixel capture, one Retina laptop screen, is about 4,800 tokens. The words are usually all the agent needed. On a real screenshot, 1,018 image tokens became 37 text tokens, 96% saved.

A model has no memory between turns. Every time the agent speaks, the whole conversation so far is sent again as input, and that includes every image you ever pasted. A screenshot is not paid for once. It is paid for on the turn you paste it, and again on every turn after that, until the session ends.

That compounds fast. One 4 megapixel screenshot is about 4,800 tokens. Over a 50 turn session it is sent 50 times, roughly 240,000 input tokens for one picture. The 37 words cheapshot would have read from it cost under 2,000 tokens over the same 50 turns. Paste five screenshots in an afternoon of debugging and the pixels alone are more than a million input tokens of resend.

Prompt caching lowers the price of the resend, but it does not shrink the context the images occupy. A long session hits the context window sooner, gets compacted sooner, and loses earlier work sooner, and the images are the heaviest thing in it. The words are a few lines of text that cost almost nothing to carry.

Works with Claude Code as a plugin, and with Cursor, Cline, Codex, OpenCode, pi and oh-my-pi through the MCP server.

Requirements

• A Mac on macOS 13 or later. The binary is universal, Apple silicon and Intel; Intel is best-effort since Homebrew moved x86_64 to Tier 3 in September 2026.
• Nothing else for screenshots and PDFs. OCR is Apple’s Vision framework, already on the machine. No account, no API key, no network.
• ffmpeg on the PATH for --video (brew install ffmpeg).
• Node 22 or later for the MCP server, which runs through npx. The Claude Code plugin needs it too, since the plugin bundles that server.
• Xcode, only if you build from source.

Install with Homebrew

brew install all-caps-dev/tap/cheapshot

Universal binary (Apple silicon and Intel), notarized, macOS 13 or later. Intel is best-effort: Homebrew moved x86_64 to Tier 3 in September 2026.

Build from source

git clone https://github.com/all-caps-dev/cheapshot
cd cheapshot
make            # builds ./cheapshot and checks it is universal with a macOS 13 floor
sudo make install   # copies to /usr/local/bin

Needs Xcode; the universal build uses the Xcode build system, which the command line tools alone do not provide.

make test runs the suite. --video needs ffmpeg on the PATH (brew install ffmpeg); everything else is Apple frameworks only.

Use covers the commands; Redaction explains what never leaves the Mac.