markitdown-vellum

A Vellum Assistant plugin that lets the model read office documents, PDFs, spreadsheets, presentations, notebooks, archives, and similar files as markdown — without you opening or copying them by hand.

Built on @cognipeer/to-markdown.

What it does

Adds one tool, markitdown, to the assistant. The model can point it at a file in the workspace and get back extracted text: headings, tables, slide content, sheet data, email bodies, and image descriptions where applicable.

Scanned PDFs and images are OCR'd via the workspace vision model (visionMode: "vellum") or bundled Tesseract.js (visionMode: "tesseract").

Use it when a conversation needs content from office documents, PDFs, spreadsheets, presentations, notebooks, archives

Plain text and audio are not in scope (excluded from cognipeer's FileExtension list)

Supported extensions follow @cognipeer/to-markdown FileExtension, except .txt, .mp3, and .wav.

Install

From the Vellum plugin catalog or GitHub. For local development: copy into <workspace>/plugins/markitdown-vellum/, run npm install, reload plugins.

Configuration

Edit config.json in the plugin directory (preserved on upgrade). Omitted keys use the defaults below.

Field	Default	Description
documentTimeoutMs	120000	Max time for a single conversion (ms)
visionMode	"vellum"	"vellum" — OCR via workspace vision (cognipeer custom-vlm + fetch patch); "tesseract" — OCR via bundled Tesseract.js
visionSystemPrompt	see config.json	System prompt for vision OCR
visionDefaultUserPrompt	see config.json	User prompt sent to the vision model (also cognipeer vlm.prompt)

Optional system dependency

None for Tesseract mode — @cognipeer/to-markdown ships . PDF page rendering uses the native module (installed automatically).