Last updated: 2026-08-24
Zerox is an open-source document extraction library with 12.2k GitHub stars that converts PDFs and other office and markup formats into markdown or structured JSON by sending each page to a vision model, for example GPT-4o or Gemini, instead of running traditional OCR. It ships as separate Node.js and Python packages.
About Zerox
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 used open-source document-parsing tools built specifically for AI ingestion pipelines. The library handles a long list of file types, including PDF, DOCX, RTF, HTML, XLSX and PPTX, and can output either full-document markdown or structured JSON when you pass a JSON Schema through the extractOnly option. It ships as both a Node.js package (npm install zerox) and a Python package (pip install py-zerox). 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. 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.
Pricing
Zerox itself is free and MIT-licensed with no paid tier of its own. Running it costs money indirectly, since every page processed is a billed API call to whichever vision provider you configure, and that provider's own per-token or per-image rate is what you actually pay, not Zerox. Cost scales with document volume and the model you pick, from a cheap GPT-4o-mini run up to a top-tier model on a large batch.
Key Features
- Vision-model document conversion: Renders each PDF page as an image and feeds it to a vision model like GPT-4o or Gemini, then returns clean markdown instead of running a traditional OCR pass.
- 26+ supported file formats: Handles PDF, DOCX, RTF, HTML, XML, XLS, XLSX, PPT, PPTX, CSV, ODT and more, all converted through the same pipeline.
- Structured JSON extraction: Accepts a JSON Schema via the extractOnly option (Node.js only) to return structured fields instead of raw markdown.
- Multi-provider support: Works with OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI and Vertex AI, so you are not locked into a single vision provider.
- Concurrent batch processing: Processes multiple pages in parallel with a default concurrency of 10, configurable for larger batch jobs.
Pros
- No vendor lock-in or per-seat fee, since the code is MIT-licensed and the only recurring cost is whatever vision-model API you already pay for.
- GitHub stars and broad file-format support make it one of the more visible open-source document parsers available today.
- Works with six different vision-model providers instead of locking you into one vendor.
Cons
- No tagged release since v0.1.06 on December 18, 2024, and the maintaining company rebranded to Monumint in July 2026 to pursue an unrelated fintech product, raising real questions about ongoing maintenance.
- The Python package (py-zerox) is missing features the Node.js package has, including schema extraction, per-page output and orientation correction.
- Requires graphicsmagick and ghostscript as system dependencies, which adds install friction compared to a pure-npm or pure-pip package.
- The original getomni.ai benchmark results and documentation pages now return 404 or redirect to the unrelated Monumint site, making it harder to verify accuracy claims.
Frequently Asked Questions
What does Zerox actually cost?
Zerox the library costs $0: it's released under the MIT license with no subscription or seat count. What you actually pay is your vision-model provider's own rate, since every page processed is one billed API call to whichever of the six supported providers you configure. A cheap model like GPT-4o-mini keeps that per-page cost low, while routing a large batch through a top-tier model gets expensive fast.
Can you use Zerox without paying?
Yes: the library itself has no usage cap, seat limit, or signup gate, since it's a self-hosted MIT-licensed package rather than a metered SaaS account. You install it with a single npm or pip command and can process as many documents as your own API budget allows, with nothing owed to OmniAI itself. The MIT license also means you can fork or trim the code down to only the pieces you need.
What should you use instead of Zerox?
Reducto is the closest hosted alternative: it turns PDFs, scans and spreadsheets across 30+ file types into structured JSON with agentic OCR correction, a better fit if you want a managed API instead of running a library yourself. Veryfi is a stronger pick for structured extraction workloads like invoices and receipts, publishing a 99.56% field-accuracy figure that Zerox has no equivalent benchmark for. LlamaIndex is the pick when you specifically need agentic OCR built into a broader enterprise automation workflow, not just a single conversion step.
What separates Zerox from Reducto?
Reducto is a hosted, managed API, while Zerox is a self-hosted library you run against your own vision-model key, so the real trade-off is control versus convenience. Reducto adds agentic OCR correction across 30+ file types out of the box; Zerox instead depends entirely on whichever vision model you route pages through yourself. Teams that want a maintained vendor relationship and support are better off with Reducto, while teams already paying for a vision-model API and comfortable self-hosting can stick with Zerox's MIT-licensed code at no extra cost.
How do you get started with Zerox?
Getting started takes three steps: install the package for your language (Node.js or Python), install graphicsmagick and ghostscript as system dependencies on the host machine, and get an API key from a supported vision provider. From there, one function call like zerox({ filePath, apiKey }) returns markdown for a document in a single request, or you can pass extractOnly: true with a JSON Schema to get structured fields back instead. There's no signup, dashboard, or account to create, since everything runs through your own vision-model key you already hold.
Top Alternatives
- Reducto: Reducto is the better pick when you want agentic OCR correction bundled into a hosted API instead of a self-hosted library you maintain yourself.
- Veryfi: Choose Veryfi over Zerox for a managed API with a published 99.56% field-accuracy figure instead of an unmaintained open-source library.
- Mindee: Mindee suits teams that need a hosted extraction API built specifically for invoices, receipts and IDs rather than a general-purpose document-to-markdown library.
- LlamaIndex: LlamaIndex layers an agentic OCR workflow on top of enterprise automation, something this bare conversion library was never built to do.
HokAI guides covering Zerox
- Best AI Search & Retrieval Infrastructure Tools in 2026: AI search and retrieval infrastructure is four purchases, not one: web search, document parsing, vector storage, and agent memory. Compare eight tools by layer.