Extract text from images with word-level boxes.

A free OCR tool that extracts editable text from images and scanned PDFs directly in your browser — screenshots, receipts, whiteboard photos, scanned contracts, certificates, letters, and archives. Recognized words come with confidence scores and on-page coordinates, and there is an optional on-device vision model for asking questions about a page.

The engine is Tesseract compiled to WebAssembly: nothing is uploaded, there is no account, and a batch of images or a 300-page PDF works through the same queue.

How it works

Tesseract.js — the open-source Tesseract OCR engine compiled to WebAssembly — runs the recognition entirely in your tab. The engine core and each language’s trained data download on demand and are cached by the browser, so after first use recognition works offline. You can pick from 24 language chips covering English, simplified and traditional Chinese, Japanese, Korean, the major European languages, and Arabic, Hebrew, Hindi, Thai, Vietnamese, and other scripts; Tesseract’s full catalogue covers well over 100 languages under the hood.

Before recognition, images can be preprocessed (grayscale or binarization) to lift accuracy on faint or noisy scans, and a page-segmentation mode — automatic, single block, single line, or sparse text — tells the engine how the page is laid out. A minimum-confidence slider filters weak words out of the text view.

Scanned PDFs, page by page

A PDF is handled the way it should be: every page is rasterized at roughly 144 DPI and recognized in order, so a multi-page letter or contract becomes one continuous transcript rather than a folder of screenshots. Pages are rendered lazily — only the page being recognized is in memory — which keeps even 300-page documents feasible on an ordinary laptop. Output formats are plain TXT, JSON with each word’s bounding box and confidence, and hOCR with positions and layout; in the viewer, every recognized word is overlaid on the page and color-coded by confidence.

Batch queue and ZIP export

Images and PDFs share one serial queue with per-item status, word counts, elapsed time, and a running progress bar. Add a folder of receipts or a stack of scanned letters and they process one after another; failed items stay marked so a re-run only retries those, and results for the whole queue download together as a ZIP of named text files. You can paste screenshots from the clipboard as well as drag files in.

Smart extract with a vision model

Beyond transcription, Smart extract runs an on-device vision-language model (Moondream, around 800 MB via WebGPU) to answer free-form questions about the open page — “What is the total amount on this receipt?” or “What date is stamped here?” It is a separate, optional download, requires WebGPU, and is disclosed as plainly as the OCR engine: the model runs locally and the image never leaves your device.

Free and Pro limits

The free tier processes batches of 10 source files (each PDF counts as one file) and PDFs up to 30 pages, with unlimited batches. Pro raises the queue to 50 files and PDFs up to 300 pages. Pages beyond the PDF limit are skipped with a notice rather than failing the whole job. Neither tier has a daily quota, a run cap, or a watermark, and both include commercial use.

Quality tips and privacy

Sharp, well-lit, straight images of printed text recognize near-perfectly; cursive handwriting, heavy skewing, and decorative fonts remain genuinely hard, and the confidence overlay shows you exactly where to proofread. Switching to the correct language before recognition is the single biggest accuracy lever for CJK and Arabic-script pages.

Scans of contracts, IDs, medical forms, or private correspondence stay on your device by construction — recognition never uploads a pixel, which is the whole reason to choose it over a free online OCR website.