Last updated: 2026-07-11
Zerox is a free, MIT-licensed OCR library with 12.2k GitHub stars that converts PDFs into markdown using vision models like GPT-4o, Claude and Gemini.
Zerox is an open-source, MIT-licensed document extraction library with 12.2k GitHub stars that converts PDFs and 26+ file formats into markdown using vision models such as GPT-4o, Claude and Gemini instead of traditional OCR. Built by OmniAI, it ships as both a Node.js and a Python package for RAG and document-ingestion pipelines.
Zerox is an open-source document extraction library built by OmniAI, the San Francisco company founded in 2023 by Tyler Maran and Anna Pojawis that went through Y Combinator's Winter 2024 batch. Instead of running a traditional OCR engine, Zerox converts each page of a document into an image and sends it to a vision model such as GPT-4o, Claude, or Gemini, then returns clean markdown. The project has 12.2k stars on GitHub, making it one of the more widely adopted open-source document parsing tools built specifically for AI ingestion pipelines. The library supports more than 26 file formats including PDF, DOCX, RTF, HTML, XLSX and PPTX, and it can output either full-document markdown or structured JSON when you pass a JSON Schema with the extractOnly option. It ships as both a Node.js package (npm install zerox) and a Python package (pip install py-zerox), though the Python version is missing several features the Node version has, including schema extraction, per-page output and orientation correction. Both versions require graphicsmagick and ghostscript installed on the host machine, and both require you to bring your own API key from a supported vision provider (OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI or Vertex AI). Zerox is aimed at developers building retrieval-augmented generation (RAG) pipelines, document search tools, or any system that needs to turn scanned or complex-layout documents into text an LLM can reason over. Its sister repository, getomni-ai/benchmark, publishes an open-source methodology for scoring OCR and extraction accuracy across vision models and competing tools, though the original results pages on the getomni.ai blog now return 404 after the company's later product pivot. Zerox itself costs nothing to use since it is MIT-licensed, but running it is not free: every page you process is a paid API call to whichever vision model you configure, so cost scales with document volume and the provider you choose. The most important thing to know before adopting Zerox in 2026 is that it appears to be in maintenance limbo. Its last tagged release, v0.1.06, shipped on December 18, 2024, and OmniAI itself rebranded to Monumint in July 2026 to build a conversational AI product for financial institutions, a completely different focus. The code remains public under the MIT license and can still be used or forked, but there is no indication of active paid maintenance from the original team going into 2026.
Zerox itself is free and MIT-licensed with no paid tier. Running it costs money indirectly: each page processed is a paid API call to whichever vision provider you configure (OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI or Vertex AI), so cost scales with document volume and your chosen model's per-token pricing.
Zerox is an open-source document extraction library built by OmniAI, the company founded in 2023 by Tyler Maran and Anna Pojawis that went through Y Combinator's Winter 2024 batch. It converts PDFs and other documents into clean markdown by turning each page into an image and sending it to a vision model such as GPT-4o, Claude or Gemini, rather than using a traditional OCR engine. The project has 12.2k stars on GitHub. It supports more than 26 file formats including PDF, DOCX, RTF, HTML and XLSX. It is distributed as an npm package for Node.js and a PyPI package for Python. Developers building AI pipelines that need clean text from documents are the primary users.
Zerox is completely free because it is licensed under MIT, with no paid tier, subscription or usage limit imposed by the library itself. The real cost comes from the vision model you configure to run it, since every page processed is a billed API call to a provider like OpenAI, Anthropic or Google Gemini. There is no bundled or discounted API pricing; you pay each provider's standard per-token or per-image rate directly. Cost therefore scales with document volume and which model you choose, not with anything Zerox itself charges. Teams already paying for a GPT-4o or Claude subscription can reuse that spend rather than adding a separate paid OCR vendor. There is no free hosted demo currently available since the getomni.ai demo page now redirects to an unrelated product.
Zerox's core feature is converting document pages into images and running them through a vision model to produce markdown or structured JSON output. It supports more than 26 file formats, including PDF, DOCX, RTF, HTML, XML, XLS, XLSX, PPT and PPTX. Its structured extraction mode, available in the Node.js package, accepts a JSON Schema and returns typed fields instead of raw text. It supports six vision providers: OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI and Vertex AI. It processes pages concurrently with a default concurrency of 10, which can be tuned for larger batch jobs. The Node.js package additionally supports per-page extraction, orientation correction and edge trimming that the Python package lacks.
Yes, the Zerox library itself is entirely free under the MIT license, with no paywall, subscription or usage cap. What is not free is the vision model API call each page requires, since Zerox does not include or subsidize any model access. You must supply your own API key from a supported provider, and that provider bills you according to its own per-token or per-image pricing. There is no free hosted version of Zerox to try without your own API key. Anyone evaluating cost should budget for the underlying vision model's pricing, not for Zerox itself, since the library adds no markup or fee of its own. Because the source code is MIT-licensed, you can also fork and modify it freely, including stripping out features you don't need. This makes total cost highly variable, since a lightweight GPT-4o-mini setup will run far cheaper than routing every page through a top-tier model.
LlamaParse, from LlamaIndex, is a hosted document parsing API that scored 84.9% on the independent ParseBench evaluation and is a strong choice if you want a managed service instead of a self-hosted library. Reducto is a hosted document extraction API that claims roughly 20% higher accuracy than LlamaParse and Unstructured on real-world documents and offers SOC 2 and HIPAA compliance, making it a better fit for regulated industries. Unstructured.io offers both an open-source library and a hosted API with broad format support, useful if you want a similar bring-your-own-infrastructure model with more active recent releases. Docling, from IBM, is a self-hostable open-source Python library with GPU-accelerated layout parsing, worth considering if you want an actively maintained open-source option instead of Zerox's stalled release history. Marker is another open-source Python library, best suited to academic and technical PDFs that need local inference without calling an external vision API. Which one fits best mostly comes down to whether you need an active vendor relationship or are comfortable maintaining an unmaintained but functional MIT-licensed tool yourself.
Zerox is best for backend and data engineers building retrieval-augmented generation (RAG) pipelines or document search systems who want to reuse an existing vision-model API key instead of adopting a separate OCR vendor. It suits teams comfortable installing system dependencies like graphicsmagick and ghostscript and working directly with an npm or pip package rather than a hosted dashboard. Open-source contributors looking for a fork target with real developer mindshare, evidenced by its 12.2k GitHub stars, are also a good fit. It is not a good fit for enterprise buyers who need a vendor-supported SLA, since the maintaining company shifted its focus to an unrelated fintech product in 2026 and has not shipped a release since December 2024. Non-technical users who want a no-code hosted document parser should look elsewhere. Startups with tight budgets who already pay for a vision-model API and just need occasional document parsing get the most value, since there is no separate subscription to add. Teams needing guaranteed uptime or support tickets should weigh the maintenance gap carefully before committing to it in production.
Zerox itself is not a hosted API; it is a library you install and call directly in your own code, either as the zerox npm package for Node.js or the py-zerox package for Python. There is no standalone REST endpoint or SaaS API maintained by the project. A basic call looks like zerox({ filePath, openaiAPIKey }), which returns markdown, or you can pass extractOnly: true with a JSON Schema to get structured JSON back. It has no Model Context Protocol (MCP) support documented anywhere in its repository or docs as of this review. Developers who want a hosted API instead of a self-managed library should look at Reducto or LlamaParse, both of which offer managed REST endpoints. Zerox's design assumes you already have infrastructure to run Node.js or Python code and call out to a vision provider directly. Because it is a library rather than a service, you also control where your documents are processed, which some teams prefer for data residency reasons even without formal compliance certifications.
LlamaParse is a hosted, actively maintained API from LlamaIndex that scored 84.9% on the independent ParseBench evaluation, while Zerox has no independently verified accuracy score and has not shipped a release since December 2024. If you want a managed service with ongoing support and a company actively iterating on accuracy, LlamaParse is the safer choice for production use. If you want full control over which vision model powers extraction, and are comfortable with an unmaintained but functional open-source library, Zerox's MIT license and multi-provider support remain usable today. Cost structures differ too: LlamaParse bills you directly for its managed service, while Zerox only costs whatever your chosen vision provider charges per page. For teams that need active vendor support or an SLA, LlamaParse is the stronger pick; for teams that just need a free, self-hosted starting point and can tolerate stalled development, Zerox still works. LlamaParse also handles complex layouts like nested tables more reliably according to third-party benchmarks, which matters for dense financial or legal documents. Zerox's advantage is that you are never locked into LlamaIndex's infrastructure or pricing, since you supply your own vision-model key.
Zerox is a free, MIT-licensed library that turns PDFs and 26+ document formats into clean markdown using vision models like GPT-4o.
OmniAI ยท Free tier available