litert-ocr

OCR for photos and screenshots in the browser and in Node.js, with the tesseract.js API. It runs PP-OCRv6-Small on LiteRT.js. On 2026-11-10, version 0.2.1 read an image in 2,982 ms on a Pixel 8a in Chrome 156, with the model already loaded. Images never leave the device.

Try the demo on your own photo, scan or PDF. It reads the file in the browser tab, draws a box around each line of text, and gives you the text, Markdown and JSON.

Install

npm install litert-ocr

import { createWorker } from 'litert-ocr';

const worker = await createWorker('eng');
const { data } = await worker.recognize('https://example.com/receipt.jpg');
console.log(data.text);
await worker.terminate();

From a CDN, with no build step:

<script type="module">
  import { createWorker } from 'https://cdn.jsdelivr.net/npm/litert-ocr@0.2.2/+esm';
</script>

One model reads Latin-script, Greek, Chinese and Japanese text, so 'eng' reads them all. Korean, Cyrillic, Arabic, Hebrew, Thai and Indic scripts are not covered: createWorker('kor') rejects.

The first run downloads about 34 MB of model files and caches them.

In a browser, WebAssembly needs two headers to use more than one thread.

Related MCP server: document-intelligence-mcp

Switching from tesseract.js

Change the import:

- import { createWorker } from 'tesseract.js';
+ import { createWorker } from 'litert-ocr';

The calls and the result shape stay the same. A few options and output formats differ. Each detected line comes back on its own, so a table row can come out as separate lines. For tables, use convert.

PDFs, JSON and agents

Turn a PDF or a scan into Markdown with tables, JSON or text (Documents → Markdown):

import { convert } from 'litert-ocr';

const { content, document } = await convert(file, { to: 'md' });

From the command line:

npx litert-ocr report.pdf > report.md

As an MCP server for coding agents:

claude mcp add litert-ocr -- npx -y litert-ocr mcp

extract fills a JSON Schema from a document with an on-device language model. It needs a browser with WebGPU, and its first call downloads 2.0 GB.

Speed and accuracy

Time per image with the model already loaded (litert-ocr 0.2.1, 2026-11-10):

The receipt in the picture is a long one: 54 lines in 604 ms on the same Mac.

Character error rate on 2026-10-19 is lowest on screenshots: 0.6 % in English and 1.4 % in Japanese. It rises to 10 % on receipts from the CORD set, 26.7 % on English photos from the TextOCR set and 30.5 % on Japanese photos of signs from the JaWildText set. All the numbers and how they were measured.

Status: v0. Tested in Node.js 24, Deno 2.9 and Bun 1.4, in Chrome, and in Playwright's WebKit build. Safari itself (on a Mac or an iPhone) and Firefox are not tested yet.

Reference

• API

• Errors

• Offline

• Model and runtime files

• Safari and WebGPU

• Tests

Examples

• examples/html: plain HTML with the CDN import (node serve.mjs serves it with the two headers).

• examples/react-vite: React with Vite.

• examples/nextjs: Next.js App Router, with the import inside useEffect of a client component.

• examples/pdf-to-markdown: documents to Markdown, JSON or text, in the browser and in Node.js.

• examples/mcp: the MCP server's configuration for Claude Code and Claude Desktop.

License

Apache License 2.0. See NOTICE for the parts copied from google-ai-edge/litert-samples and for the model and runtime licenses. Independent project, not affiliated with Google or PaddlePaddle.