Zerox

OmniAI's Zerox reduces OCR to its simplest possible form: rasterize each document page to an image, hand it to a frontier vision model, and collect markdown back. The approach handles layouts that break traditional OCR, including tables, charts, and mixed formatting, because the vision model reads the page the way a person would. It ships as both a Node package (npm install zerox) and a Python package (pip install py-zerox), with support for OpenAI GPT-4o and GPT-4o-mini, Azure OpenAI, AWS Bedrock Claude models, and Google Gemini, plus Anthropic and Vertex AI in the Python version. Provider API keys are required, so cost and privacy follow whichever model is chosen, and there is no fully offline path. Around 26 input formats are accepted, including PDF, DOCX, XLSX, PPTX, HTML, and images, with markdown returned per page for downstream chunking. The MIT license permits commercial use and the repository has about 12,300 stars, though commits slowed after mid-2025; OmniAI keeps a live demo at getomni.ai/ocr-demo. It became a default for quick LLM document ingestion where setup time matters more than per-page cost.