Keep the colour, lose the megabytes— and read any book or manga while you commute, hands-free, offline, all in one app.

The grayscale checkbox is gone

Shrinking a scan used to mean ticking grayscale and accepting that everything went grey — the teal headings, the blue stamp, the signature, the QR code. It saved about a quarter of the file and cost you every coloured pixel on the page.

The reducer now reads the page in horizontal strips and drops colour only where there is none worth keeping. A strip carrying a stamp keeps its stamp. A strip of black text on white paper loses chroma it was never using. The gate was audited against 3,447 real strips from live documents and missed none of them.

Two other things happen underneath. Every JPEG inside the PDF is re-encoded with optimised entropy coding at the source's own quantisation tables — that step is mathematically exact, the pixels do not change. And the reducer builds two candidate files, one rasterised and one not, then ships whichever came out smaller. That is why it never hands back something bigger than what you gave it.

The same page, three ways

Across page classes, measured on real files rather than samples chosen to flatter:

Two presets. High is the one to send to a client. Email is the one that gets under an attachment limit. Genuinely dirty scans — where the paper photographed grey rather than white — can also opt into background flattening, which forces paper-only blocks to true white and is worth up to another 14%. It is a checkbox rather than the default because on clean office scans it earns nothing at all.

Drop a file on the reducer to see your own numbers. No key, no signup; the result comes back in the browser, and a large file is processed as far as it can be and you are told exactly how far, rather than refused.

A note for anyone who has read about MRC compression, the textbook approach of splitting a page into a 1-bit ink mask, a colour foreground and a blurred background. We built it and measured it across 62 pages against ground truth. It is dramatically smaller — −81% — and it damaged the text on two thirds of those pages, while simply spending fewer bits got −53% at no measurable cost. So this reducer is not MRC. It is the boring thing that beat it.

Hear and watch your stories in motion

Auto-Reader is the reader app this API grew out of. It turns any book, manga or PDF into something you can listen to and watch hands-free — on the commute, at the gym, on a walk — with lifelike on-device voices. On-device by default, works offline, no account.

More than an app: the browser tools

Everything the app does is also on the website, free to try, and most of it does not need an account:

The boxes: every tool on one page, none of it needs a key

The API site is built so you can find out whether the output is good before deciding anything: every box runs without a key, and there is a free tier the moment you add an email. Current limits and plans are always on the site itself, so they are not repeated here.

What people actually come here to do

One line makes it a tool your agent can call

There is an MCP server, so the whole pipeline becomes tools inside Claude Code, Claude Desktop, Cursor, or anything else that speaks MCP:

claude mcp add --transport http auto-reader-ocr https://api.auto-reader.com/mcp

After that an agent can read a scan, shrink a PDF or translate a chapter without you writing an integration. If you would rather call it directly, it is an ordinary HTTP API:

curl -X POST https://api.auto-reader.com/v1/reduce \
  -H "Authorization: Bearer $KEY" \
  -F "[email protected]" -F "quality=email" -o small.pdf

• Keys provision themselves. An agent can sign up for its own key and start working; no dashboard visit required.
• There is an /llms.txt written for models rather than for people, alongside an OpenAPI spec and a live playground.
• Nothing is kept by default. Pages are processed in memory; only a hash is logged unless you explicitly opt into training storage. Built for Saudi PDPL.