Image OCR & Scanned PDF to Text — 19 Languages, Line Boxes

Paste links to images (PNG, JPEG, WebP, BMP) or PDFs and get the text in them as clean JSON: one row per image or PDF page, with the OCR confidence, word count and, if you want, every line with its position. Scanned PDFs are read page by page; PDF pages that already have a text layer are taken as is (exact and free of OCR mistakes). 19 languages, including Japanese, Chinese, Korean, Arabic, Hindi and Thai. The reading happens inside the run — your files are not sent to any other OCR service.

Try it now. The input is already filled in. Press Start: one scanned English page (Tesseract's sample image) — 8 lines, 58 words in about 7 seconds, for $0.013, confidence 92 (measured on 28 Oct 2026). The first lines of text:

Mild Splendour of the various-vested Night!
Mother of wildly-working visions! hail
I watch thy gliding, while with watery light

The row also has the image size, line count and the language used.

What you get

Row	Fields
page	url, page and pageCount (PDFs), method (ocr / text-layer), text, confidence (0-100), words, chars, lineCount, lines (with Include lines with position: text, confidence, box {x, y, width, height}), width, height, fileType, fileBytes, languages, note
error	url, message — free

Use it for

• Scanned documents: contracts, letters, old reports and faxes that are only pictures inside a PDF
• Receipts and invoices: amounts, dates and reference numbers, with line boxes to find where each value sits
• Screenshots and product images: text on banners, labels, menus and slides
• AI and RAG: turn image-only documents into text your pipeline can search and embed

Input

{ "urls": ["https://tesseract.projectnaptha.com/img/eng_bw.png"], "languages": ["eng"], "includeLines": true }

Measured on 23 October 2026 (our own test files): a 1240×700 invoice image read in about 2 seconds with 90+ confidence; a 2-page scanned PDF and a 1-bit fax-style PDF read page by page; Japanese read correctly with jpn.

How it reads

• Each site's robots.txt is read first (the rules for all robots and those addressed to Claude crawlers); links it does not allow are skipped with a free error row. Links to private networks are not read.
• OCR is done with Tesseract (LSTM models) inside the run. Clear, straight text at a normal size reads best; blurry, tiny, rotated or handwritten text reads poorly — the confidence and a note tell you when to be careful.
• In a scanned PDF, the largest picture on each page is read. GIF, TIFF and SVG are not read (free error row).
• Pick only the languages that are in your files (up to 3): each extra language makes reading slower and slightly less accurate.

Pricing

$3 per 1,000 images or pages ($0.003 each), plus $0.01 per run. Images and pages where no text is found, and error rows, are free. The $0.01 run-start fee is the one charge every run pays, even one that returns nothing.

For AI agents

{"urls": ["https://example.com/scan.pdf"], "languages": ["eng"]} returns one page row per page with its text and confidence. Rows have type: page or error.

**Something broke or looks wrong?** Open an issue on this Actor's Issues tab and I'll look at it. **Found it useful?** A short review on the Store page helps other people find it.