Free desktop app + CLI · Mac, Windows waitlist
Privacy-first document extraction.
Pull text, tables, and data out of PDFs with OCR that runs on your own computer. Redact sensitive details before anything leaves your desk. Every result links back to the exact spot on the page.
macOS · Apple silicon — 1.0.0-rc.55
Download Okra-1.0.0-rc.55.dmg (57.4 MB) — Apple silicon Mac, macOS 13 or later. Developer ID signed, hardened, notarized by Apple, and stapled.
SHA-256 260c885d918b91c55e5f2f4991f955bc7a87630d71cdd888a8a59995e2b164b9 · release notes
- Download the DMG and open it.
- Drag Okra onto the Applications shortcut, then eject the DMG.
- Open Okra from Applications — it opens normally under Gatekeeper.
- Optional: verify the download with shasum -a 256 -c Okra-1.0.0-rc.55.dmg.sha256.
The desktop app is free and needs no account. Your files never leave your computer.
Extract without uploading. Verify every value. Redact, export, share.
Drop in a scan or a text PDF. Built-in OCR reads it on your machine, so nothing is sent to a server. Each extracted block links to its exact location on the page, so you can check a number against the source in one click. Find names, account numbers, and other sensitive details, review each one, then export Markdown, JSON, or CSV, or save a redacted copy as a new PDF.
Find the model that fits your performance and privacy requirements.
19 parsers. Scores recited from the published ParseBench leaderboard (arXiv 2604.08538) and our reproduction of its harness. Cost is vendor list price per page, not a measured spend. Our own dataset is in progress.
ParseBench overall score, peak memory, and list cost per page, by parser. On-device parsers first.| Parser | Runs | Overall | Memory | Cost / page |
|---|
| Infinity-Parser2 Flash · Local, MLX | On your machine | 73.3 | 4 GB | Free |
|---|
| Chandra OCR 2 · Local, 4-bit MLX | On your machine | 70.1 | 3.5 GB | Free |
|---|
| Qwen 3 VL 8B · Local, 4-bit MLX or Ollama | On your machine | 62.0 | 6 GB | Free |
|---|
| Dots OCR 1.5 · Local, 4-bit MLX | On your machine | 55.8 | 3 GB | Free |
|---|
| PyMuPDF4LLM · Text layer only | On your machine | 53.5 | 0.1 GB | Free |
|---|
| Docling · CPU pipeline | On your machine | 50.6 | 2 GB | Free |
|---|
| MinerU 2.5 · 1.2B, local | On your machine | 45.9 | 3 GB | Free |
|---|
| LlamaParse · Agentic | Cloud | 84.9 | — | 1.25¢ |
|---|
| LlamaParse · Cost effective | Cloud | 76.8 | — | 0.38¢ |
|---|
| Gemini 3 Flash · Thinking high | Cloud | 75.0 | — | 2.41¢ |
|---|
| Reducto · Agentic | Cloud | 73.0 | — | 4.76¢ |
|---|
| Gemini 3 Flash · Thinking minimal | Cloud | 71.0 | — | 0.65¢ |
|---|
| Chandra OCR 2 · Datalab hosted, balanced | Cloud | 69.5 | — | 0.40¢ |
|---|
| Reducto · Standard | Cloud | 67.8 | — | 2.38¢ |
|---|
| Mistral OCR 4 · Standard | Cloud | 60.7 | — | 0.40¢ |
|---|
| Azure Document Intelligence · Layout | Cloud | 59.6 | — | 1.00¢ |
|---|
| Google Document AI · Layout parser | Cloud | 50.4 | — | 1.00¢ |
|---|
| AWS Textract · Layout + tables | Cloud | 47.9 | — | 1.50¢ |
|---|
See every benchmark · How we reproduced the local runs
Built for documents you can’t upload.
Start with the built-in OCR, add a supported local model, or connect Ollama on your machine. Models download only when you approve. Opening a PDF never starts a parse; exports and redacted copies are new files. Cancel a long parse, inspect the run history, and resume from completed pages in the app or the CLI.
Read the CLI guide
