Invoices in PDF, scanned contracts, itemized spreadsheets in Excel — there are plenty of situations where you want to turn the documents on hand into something an AI or a program can work with directly. Docstrange is an open-source document processing tool that converts these “readable but unstructured” documents into Markdown, JSON, CSV, or HTML. It is developed by Nanonets, a US company specializing in document AI. Released in August 2025 as an MIT-licensed Python library, its defining trait is a two-track design: it can run through a cloud API or entirely on your own machine. A browser-based version is also available, so you can drag and drop a file and check the result right away.
Key Features
- Broad input format support: Handles PDF, DOCX, PPTX, XLSX, CSV, images such as PNG, JPG, and TIFF, plus HTML and plain text. Scanned images are run through OCR first, then their structure is reconstructed
- Four output formats: Converts to Markdown, JSON, CSV, or HTML. Use Markdown when feeding an LLM, JSON for system integration, CSV for spreadsheets — pick whichever fits the job
- Strong table extraction: Designed to preserve row and column relationships even for borderless tables and tables with merged cells, targeting the part of document processing that breaks most easily
- Choice of cloud or local: Cloud mode is accurate with no setup required. Local mode (CPU / GPU) runs entirely on your own machine, so it can handle confidential documents that must not leave your environment
- Field-level and schema-based extraction: List the field names you want — “invoice number”, “total amount” — and it pulls the matching values, or pass a JSON schema to lock down the output structure itself
- Bundled MCP server: A Model Context Protocol server module ships with the project, so clients like Claude Desktop can read documents through it
Pricing
| Plan | Price | What you get |
|---|---|---|
| Cloud (not signed in) | $0 | Trial use with a daily call limit |
| Cloud (signed in / API key) | $0 | Up to 10,000 files per month |
| Local execution (CPU / GPU) | $0 | No file cap under the MIT license; throughput depends on your own hardware |
| Nanonets commercial plans | Contact sales | For high-volume processing or business system integration, contact the provider |
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- The free tier of 10,000 files per month is generous enough that personal use and small-scale business work run at effectively no cost
- Local mode means documents that cannot leave the organization never have to be sent to the cloud
- MIT-licensed open source, giving wide latitude to embed it in internal systems or modify it
- Four output formats cover LLM input, database ingestion, and spreadsheets alike
- Installs with a single
pip install docstrangeand can be called from the command line or from Python
⚠️ Cons
- Serious use requires a Python environment and command-line familiarity, which raises the bar for anyone wanting a GUI-only workflow
- Local mode speed depends on your machine; CPU-only execution can be slow
- Accuracy drops on handwriting, low-quality scans, and extremely complex layouts
- Going beyond the cloud free tier means contacting the provider
- No Japanese UI or documentation; reading English is assumed
Comparison with Similar Services
| Criteria | Docstrange | LlamaParse | Mistral OCR | AWS Textract |
|---|---|---|---|---|
| Provider | Nanonets | LlamaIndex | Mistral AI | Amazon Web Services |
| Delivery model | OSS library + cloud | Cloud API | Cloud API | Cloud API |
| Local execution | Yes (CPU / GPU) | No | No | No |
| Main outputs | Markdown / JSON / CSV / HTML | Mainly Markdown / JSON | Mainly Markdown / JSON | JSON (element-level) |
| Free tier | 10,000 files per month | Some free credits | Free tier available | Initial usage tier |
| Best suited for | RAG preprocessing, confidential documents | RAG preprocessing | Multilingual document reading | Embedding in business systems |
Who Is It For
- People who want to clean up PDFs and images into Markdown before feeding them to RAG or an LLM
- Business users who need to mechanically extract fixed fields from invoices, receipts, and application forms
- Organizations handling documents that cannot leave the premises and therefore need everything to run locally
- Developers who want to embed document parsing into their own systems (the MIT license permits modification and redistribution)
- Anyone who wants to gauge accuracy on a free tier before committing to a paid document parsing API
Summary
Docstrange is Nanonets carving out its document AI expertise as open source. Its greatest strength is letting you choose between the convenience of the cloud and the peace of mind of local processing within the same library, and the 10,000-file monthly free tier is practical in real work. That said, it is built around Python and has a different character from GUI-only products. A sound approach is to throw a few PDFs at the web version first, check how faithfully tables come through, and move on to installing the library once you are satisfied it holds up in production.