OCR Buddy

Faithful, fully-local OCR. Select a region, get the text — no server, no hallucinations.

OCR Buddy grabs text from anything on your screen — a code block, a paragraph, a formula, or a table — and reads it entirely on your device. Drag-select a region of any page and the text appears in Chrome's side panel, right next to the captured image so you can verify it. No pixel ever leaves your computer. WHAT'S NEW IN 2.5.6 (Minor Fix) • Added an in-panel way to rate OCR Buddy; minor polish WHAT'S NEW IN 2.5.5 (Main Features) • Page → Markdown: turn a whole page into clean, AI-ready Markdown built from the page's real structure — headings, lists, links, tables, code blocks — not OCR. Preview it, copy it, or download a .md file. Fully local. • Hybrid image OCR: text baked into readable images is recognized and inserted right after the image, clearly labelled as extracted from an image — never silently merged into the prose • Capture viewport: OCR everything on screen in one click, no region drag • Capture full page: scroll-capture an entire page and OCR it tile by tile, merged into one result • Reads coloured text on light backgrounds — red error messages, blue links and the like, which used to be skipped • Faithfulness fix: repeated same-size captures no longer risk showing a previous capture's text every capture is recognized fresh CAPTURE • Three sources: select a region, the whole viewport, or an entire scrolling page • OCR any image: open a file, paste, drag & drop into the panel, or right-click an image on a page • Capture works on a paused cross-origin video — no tainted-canvas failures THREE MODES • Text/Code — faithful text with reconstructed spacing, line breaks and code indentation, plus syntax highlighting • Formula — one equation to LaTeX, rendered beside the source crop so you can verify it • Table — one table to a Markdown grid, borderless tables included LANGUAGES • 11 languages: Latin (English, Italian, French, German, Spanish +40) built in; Chinese + Japanese, Cyrillic, East Slavic, Greek, Korean, Thai, Devanagari, Tamil and Telugu as one-time downloads, cached for offline use PRIVACY FIRST • No servers, no API calls, no telemetry, no account • Models run on-device (WebGPU, with WASM fallback); the default experience is fully offline • Page → Markdown reads the page locally and never uploads it; cross-origin images can't be read by the browser, so they keep just their alt text • Optional language packs are a one-time model download from a pinned open-source repository — no page content or image ever rides along • Per-word confidence: uncertain words are flagged, never silently trusted; blank regions return empty output, never invented text • Installs with no "access to all websites" warning — page access is granted per-site, only when you ask Free and open source (MIT) — https://ocr-buddy.com · https://github.com/Fanfulla/ocr-buddy Classic OCR, not generative OCR. Large vision-language models top the benchmarks, then invent fluent but wrong text the moment the image gets unclear — and for code, numbers, IDs, or prices a confidently-wrong transcription is worse than none. OCR Buddy makes the opposite bet: it uses detection plus CTC recognition (PP-OCRv5 on ONNX Runtime Web), has no language prior, and transcribes only the glyphs actually present. When it can't read something, it leaves a blank or flags low confidence — it never makes up a sentence. Three modes, switchable after capture: - Text / Code — prose or code, with a Code view that restores indentation and adds syntax highlighting. - Formula → LaTeX — one equation into LaTeX, rendered with KaTeX beside the source crop so you can check it. - Table → Markdown — one table into a clean grid, rebuilt from word-box geometry (works on borderless tables too). Faithful by design: the captured crop sits above the extracted text every time, low-confidence words are underlined, and an ambiguous region yields empty output rather than invented filler. Private by architecture: there is no server, no account, no telemetry, and no analytics. The OCR models are bundled inside the extension and run in a local offscreen worker, so it works fully offline. The only network use is Chrome downloading the extension itself. Free and open source (MIT). Manifest V3 · Chrome 124+ · WebGPU-accelerated with a multi-threaded WASM fallback.