Weekly Update: V3 PDF Intelligence

PDF Reader MCP V3 is a major step toward agent-grade PDF understanding with a much cleaner MCP surface.

The headline change is simple: agents can now start with one smart call. read_pdf can inspect the document, choose a high-value extraction route, and return an Agent Document Twin with Markdown, chunks, document maps, table signals, layout routing, trust signals, accessibility routing, OCR provenance, and visual evidence hooks.

The public tool surface is also smaller and easier to learn:

• read_pdf for smart Agent Document Twin extraction.
• search_pdf for cheap source-backed text evidence.
• pdf_evidence for focused inspect, render, crop, OCR, and visual-analysis operations.

This keeps the agent context lean while preserving the advanced evidence paths needed for scanned PDFs, tables, charts, formulas, figures, citations, trust reviews, and accessibility work.

V3 continues to be TypeScript-first, local-first, and benchmark-gated. The release process verifies the package contract, docs, deterministic quality checks, corpus coverage, provider evidence, crop evidence, and the SOTA release gate before publication.