MarkItDown

MarkItDown

MarkItDown converts documents to Markdown so agents consume less context than raw PDFs or Office files.

Input modes (use exactly one)

1. URL — url must be http(s): public pages, PDFs, Office files, or a presigned HTTPS GET (e.g. S3). The server fetches the URL; it cannot read paths on your machine.
2. Bytes — content_base64 + file_name (with extension, max 25 MB decoded). Use when the host already read a local file and passes content on the MCP wire (typical in Claude Code).

Output

• markdown — extracted text with structure (headings, lists, tables) where the format allows.
• extraction_method — text, ocr_plugin, or vision_ocr
• ocr_used — whether OCR ran
• pages_ocrd — pages processed when vision OCR ran

OCR modes

• ocr_mode=auto (default) — text extraction first; LLM vision OCR when a PDF has little extractable text (scanned/image-only statements, scans).
• ocr_mode=force — OCR every PDF page with LLM vision (use for known scanned documents).
• ocr_mode=off — text extraction only.
• max_pages — optional cap when OCR runs (large documents).

Workflow hints

• Batch local PDFs: read each file, call bytes mode per file, store markdown elsewhere.
• Already hosted file: presigned or public url only — no upload step on this bundle unless you use another storage bundle first.