Goldziher/sceptre

EasyOCR's accuracy, Rust's speed and footprint — CRAFT + gen2 CRNN OCR over ONNX, as a library, CLI, and MCP server.

About sceptre

EasyOCR's accuracy. Rust's speed and footprint.

sceptre is a from-scratch Rust reimplementation of EasyOCR's OCR pipeline — CRAFT text detection then gen2 CRNN recognition with CTC decoding, over ONNX. It matches EasyOCR's output on every script it supports, runs ~2.8× faster on a fraction of the memory (and a cold one-shot run is ~4.4× faster than EasyOCR warm), and ships as one self-contained binary with no Python runtime. Use it as a library, a CLI, or an MCP server.

8 scripts · CRAFT + gen2 CRNN · ONNX Runtime or pure-Rust · library · CLI · MCP · offline-first

Documentation · Install · Quickstart · Why sceptre · Benchmarks · How it works · Contributing

EasyOCR is excellent and accurate — but it's a PyTorch stack: a Python interpreter, a multi-gigabyte runtime, and a heavy process to keep warm. sceptre keeps the accuracy and drops all of that.

What you get Why it matters --------- Parity accuracy Validated against real EasyOCR output across the gen2 scripts — English, Latin, Chinese (simplified), Japanese, Korean, Cyrillic, plus Telugu and Kannada — matching text (word/char-F1) and boxes (IoU). A faithful reimplementation, not an approximation. What EasyOCR reads, sceptre reads. Substantially faster ~2.8× higher throughput than EasyOCR warm on the same corpus — and even a cold, one-shot CLI run (~4.4×) beats EasyOCR's warm, already-loaded reader. More pages per second, less waiting, cheaper batch jobs. A fraction of the memory Peak RSS around 3× lower than the Python + torch process, measured like-for-like (both whole-process peaks). Runs where EasyOCR won't — small

From the project README.

Maintaining this server?

Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:

[![On MCP Radar](https://mcp.liqiwa.com/badge.svg)](https://mcp.liqiwa.com/s/Goldziher--sceptre.html)

Related MCP servers