Uses pdfvision for cross-agent PDF extraction. pdfvision returns text with line breaks preserved, document metadata, per-page density, and optional rendered PNGs for multimodal models.
Prerequisite
npx pdfvision --version
Or install globally: npm install -g pdfvision. Requires Node.js >= 22.13.
Run npx pdfvision --help once before invoking — the flag set evolves (page selection, render output dir, layout / geometry / image-boxes opt-ins, etc.) and the help text is the source of truth for what the installed version supports.
Steps
1. Get a local PDF path
If the input is a URL, download first:
curl -L -o /tmp/doc.pdf "{URL}"
If the input is already a local path, skip this step.
2. Extract text + metadata
npx pdfvision /tmp/doc.pdf
Default output is markdown with one ## Page N section per page and a density Overview at the top.
3. Switch format when needed
# Structured JSON for downstream programmatic use
npx pdfvision /tmp/doc.pdf -f json
# Tag-shaped XML, easier for some LLMs to parse than JSON
npx pdfvision /tmp/doc.pdf -f xml
4. Render pages for multimodal review
If the Overview shows low character coverage on a page (i.e. the page is largely an image / scan), render PNGs and read them:
npx pdfvision /tmp/doc.pdf --render -p {pages}
The output reports the PNG paths. Read each image with the agent's image-reading capability.
5. Page subsets
npx pdfvision /tmp/doc.pdf -p 1-5
npx pdfvision /tmp/doc.pdf -p 1,3,5
Why pdfvision over a built-in PDF reader
- Works in any agent runtime that can spawn a CLI, not only Claude Code.
- Reports per-page density so the agent can detect rasterised pages and re-extract with
--render. - Caches by content hash, so repeated reads of the same PDF are instant.