AI agents read Markdown well and scanned PDFs badly. Hand Claude a photographed page or a two-column paper full of equations and you often get broken tables, lost formulas and paragraphs in the wrong order. TigerOCR fixes the input instead of hoping the model copes: documents in, Markdown out. It is part of Mr Pirate's tools, and the directory lists it as TigerOCR.

What you get back

Markdown with headings, tables, LaTeX formulas and the reading order preserved, plus each detected figure as its own image crop and a result.md link. It accepts PNG, JPEG, WebP, HEIC and PDF, up to 32 MB and 100 pages per job, and you can pick a page range to split longer PDFs.

TigerOCR runs Baidu's Unlimited-OCR, an open-weight vision-language model with 3 billion parameters built on DeepSeek-OCR's image encoder. It reads with language context instead of matching characters one at a time. It has two modes: gundam reads the page in tiles, which helps with small or dense text, and base reads the whole page at once at 1024 pixels.

Built for agents

TigerOCR is a remote MCP server for Claude, Gemini CLI, Codex and Cursor, and a plain REST API for everything else. Your agent gets five tools:

  • ocr_document OCRs a file from a URL or a base64 payload and waits for the result.
  • get_ocr_job fetches a result in windows, one page or a character range at a time, so a long document never floods the model's context.
  • search_ocr_job finds a passage in a finished result and returns snippets with page numbers.
  • list_ocr_jobs shows recent jobs.
  • usage reports your balance, this month's spend and the current cost per page.

The guides to my research papers on this site were written from TigerOCR's output.

What it costs and how long it takes

You pay for the GPU time your jobs use. Each job reserves an estimate from your wallet and settles to the real time when it finishes, and failed jobs are refunded in full. The typical cost is a few cents per page. Credit is added from $5 and never expires.

In TigerOCR's October 2026 tests, each page took 18 to 80 seconds of GPU time depending on the GPU and how dense the page was. A PDF's pages run on up to 5 GPUs at once, so a dense 10-page paper finished in 3 minutes 34 seconds.

Your documents

Uploads are processed in memory on the GPU and are not kept. Results are stored for 7 days so your agent can page through them, then purged, and you can delete your account from the dashboard at any time.

Start

Sign in with GitHub, create an API key and paste one connector snippet into your agent. On claude.ai, add a custom connector with the URL from your dashboard. The TigerOCR docs have the snippet for each client and the REST endpoints.