Everything cheapshot stands on, with links. Dependencies, tools, prior art, and borrowed ideas.
Everything cheapshot stands on, with links. Add a line whenever a dependency, tool, or borrowed idea arrives.
Runs on
Section titled “Runs on”- Apple Vision for on-device text recognition (
VNRecognizeTextRequest). - PDFKit for the PDF text lane.
- CryptoKit for
source.sha256. - FFmpeg for scene detection in
--video(select,mpdecimate,metadata,-fps_mode vfr). - Swift and Swift Package Manager.
- XCTest.
- Docling, with its RapidOCR engine on ONNX Runtime, for the companion script
scripts/docling-save. Not linked into the cheapshot binary.
Built with
Section titled “Built with”- Claude Code and the superpowers skills (brainstorming, writing-plans, subagent-driven-development).
- Perplexity for the five spec research questions.
- Homebrew and mislav/bump-homebrew-formula-action for distribution.
- 1Password for SSH commit signing.
Docs site
Section titled “Docs site”- Astro and Starlight.
- lucode-starlight 1.0.0, the theme.
- ryanilano/subfolio-astro-docs for the config and Pages workflow.
- withastro/action builds and uploads the site.
- actions/deploy-pages publishes it to GitHub Pages.
- actions/checkout in every workflow.
MCP server
Section titled “MCP server”- @modelcontextprotocol/sdk, the server and stdio transport behind
cheapshot-mcp. - zod for the tool argument schemas.
Distribution
Section titled “Distribution”- GitHub Actions builds, tests, signs, and notarizes.
- GitHub Pages hosts the docs site.
- GitHub Releases carries the notarized zip.
- softprops/action-gh-release creates the release the docs link to.
- Apple notarization with
notarytool.
Prior art and neighbours
Section titled “Prior art and neighbours”- mac-ocr by privatenumber: the closest neighbour, and the source of the “save vision tokens” framing.
- ocrmac, macos-vision-ocr, ocrtool-mcp, Peekaboo.
- Maus, Supamaus Lite, Redaktr: the Mac apps in the same space.
- Presidio Image Redactor, agent-sweep, Strac MCP DLP.
- ffmpeg-skill by MastroMimmo, and ffmpeg-analyse-video-skill.
Ideas borrowed
Section titled “Ideas borrowed”- Indentation from bounding boxes: nikl, “We fine-tuned our OCR to read code” and the Tesseract discussion in tesseract#2071.
- Markdown plus a JSON sidecar as the LLM payload: Docling, Marker, MinerU, olmOCR, Unstructured, Reducto, Chunkr.
- Token-format measurements: formatarc, Markdown vs HTML for LLMs and improvingagents.com, best input data format.
- Frame dedupe by text hash: US patent 11582509.
- Apple’s document APIs: RecognizeDocumentsRequest, DocumentObservation, WWDC25 session 272, VNGenerateImageFeaturePrintRequest, AVAssetImageGenerator.
Demand signals that shaped the product
Section titled “Demand signals that shaped the product”The full source lists live in docs/research/.