OWLscan
How it works

From a book on your table to a file anyone can read.

Four stages. The first one happens on your phone and takes about as long as turning the pages. The rest are choices about format, and they run on your own Claude rather than on a server here.

  1. 1 · Capture

    Point the phone and turn pages. It watches for the page to settle and shoots by itself, so both your hands stay on the book. Nothing is uploaded until you tap Finish.

    Out: scan.pdf

  2. 2 · Read

    A vision model reads whole pages in context rather than one character at a time. Hyphens rejoin across line breaks, running heads and folios drop out, paragraphs stitch back together across page boundaries.

    Out: clean.md — one continuous manuscript

  3. 3 · Choose a format

    The manuscript is the words layer. Everything below is a different look applied to the same words, and you can take all of them.

    Out: book.html · book.epub · interior.pdf

  4. 4 · Keep it

    Files, on your disk, in formats that open without us. Public-domain scans can be published to the open shelf so the next person does not have to scan it again.

    Out: a copy that outlives the scan

The toolkit

Each stage is a skill you can read.

The pipeline is a set of Claude skills and plain scripts, not a black box. Install them once and your Claude knows how to run every stage. Nothing here calls an API of ours, and nothing bills you through us.

scan-to-markdown

Three ways in: a book you are holding, an image-only PDF you already have, or a Project Gutenberg text. All three land on the same clean manuscript.

book-to-html

Semantic, reflowable HTML in one self-contained file. No webfonts, no scripts, no external requests. Original print page numbers are preserved as anchors, so a citation to page 114 still lands. Verifies that no text was dropped.

book-to-pdf

A typeset interior through CSS and Paged.js — trim size, margins, running heads, real pagination. The design is decoupled from the words, so rebuilding is a script and costs nothing to run again.

book-to-epub

Hand-assembled EPUB3 with proper footnotes and page-break anchors, gated on epubcheck at zero errors — the same validator a store runs when you upload.

proofread-manuscript

A finishing pass before layout. Spelling, grammar, and house typography are caught by local tools at no token cost; only the judgment calls reach a model.

Where the work runs

On your Claude, not ours.

Your shelf speaks MCP. One paste connects it to Claude Code, Claude Desktop, or claude.ai, and from then on your own session does the reading and the formatting. There is no agent on our side to pay for, and the whole library is plain HTTP if you would rather skip MCP and fetch the files yourself.