Introduction

PDF Reader MCP gives AI agents eyes for PDFs.

It is a full-fidelity Model Context Protocol (MCP) server that turns PDFs into agent-readable evidence — not a plain-text dump. The core product surface is an Agent Document Twin: text, tables, structure, crops, OCR, citations, trust, and accessibility signals linked through stable IDs.

The production package is sole-Rust and local-first (@sylphx/pdf-reader-mcp@4.1.1). Core extraction works without heavy model downloads. Scanned OCR and visual table/chart/formula/figure understanding are enabled through configured local providers when you need them.

What It Does

AI agents often need to access information from PDF documents - reports, invoices, research papers, manuals, and more. This server provides tools to inspect, verify, enrich, and extract:

• Smart PDF reads - read_pdf can profile unfamiliar PDFs, choose an extraction route, and return the Agent Document Twin in one response
• PDF profiles - Detect text-rich, low-text, mixed, or scanned/image-like PDFs before extraction
• PDF search evidence - Locate literal text matches with snippets, match offsets, character-derived or text-item bounding boxes, and provenance
• Visual page evidence - Render selected pages as bounded PNG MCP image parts with provenance
• Region crop evidence - Crop PDF-coordinate bounding boxes as focused PNG evidence
• Visual region analysis - Send focused crops to a configured local provider and normalize table, chart, formula, figure, and image-description results
• Configured OCR text layers - Run selected rendered pages through a local OCR provider and normalize text, confidence, words, language, and provenance
• Agent Document Twin - Read one linked document map with pages, elements, text-layer coverage, chunks, layout diagnostics, OCR evidence, visual enrichment indexes, trust routing, accessibility routing, and page geometry
• Full text content - Get all text from a PDF
• PDF text layers - Return direction-aware run records, line records, word records, character records, estimated boxes, provenance, and metadata coverage diagnostics
• Page-specific text - Extract text from specific pages or page ranges
• Metadata - Author, title, creation date, and other document properties
• Page count - Total number of pages
• Embedded images - Extract images as base64-encoded PNG data
• Agent document maps - Link pages, elements, text-layer and metadata coverage, chunks, layout confidence, safety findings, trust report routing and signal indexes, accessibility report routing and issue indexes, visual evidence routing, and geometry
• Trust reports - Summarize page risk, scores, signal counts, redacted evidence snippets, and optional document-map trust signal routing
• Accessibility reports - Summarize tagged-PDF coverage, tag-to-visible-content coverage, heading roles, images, forms, links, accessibility permissions, issue summaries, page-grade routing, and optional document-map issue indexes
• Citation chunks - Return stable source references for retrieval workflows
• Safety signals - Surface deterministic findings before agents trust PDF text

Key Features

Multiple Sources

Process PDFs from local files or URLs in a single request. Mix and match sources as needed, but each source must provide exactly one locator: path or url.

Batch Processing

Send multiple PDF sources in one request. The server processes them concurrently for optimal performance.

Flexible Extraction

Choose exactly what data you need - visual page evidence, region crops, full text, specific pages, metadata only, an agent document map, or everything including images.

Smart Default Read

Call read_pdf with only sources when an agent does not know the document shape. V3 profiles the PDF, chooses a useful extraction route, and returns the selected arguments with the response so the agent can see what happened.

PDF Search

Use search_pdf to find relevant pages and source snippets before deciding whether an agent should read, render, OCR, crop, or cite a region. OCR-layer search is opt-in so fast selectable-text search stays the default.

Focused Evidence

Use pdf_evidence when the agent needs one specialist operation after reading or searching: inspect, render_page, extract_regions, ocr_pages, or analyze_regions.

PDF Text Layer

Use include_text_layer when agents need direction-aware run, line, word, and character records with page-level ranges, estimated bounding boxes, and metadata coverage counts.

Image Extraction

Extract embedded images from PDFs for AI vision analysis. Images are returned as base64-encoded PNG data.

Page Rendering

Render selected pages as visual evidence for layout inspection, OCR routing, and agent verification. Rendered PNGs are returned as MCP image parts while JSON carries only metadata and provenance.

Region Cropping

Crop PDF-coordinate bounding boxes into focused visual evidence for tables, figures, charts, formulas, annotations, and citation verification.

Visual Region Analysis

Analyze focused crops with a configured local command or HTTP provider for visual table recognition, chart-to-data extraction, formula recognition, figure descriptions, and image captions. Provider commands and endpoints are environment-configured, so request payloads cannot choose arbitrary executables or URLs. Outputs can preserve rich table cell geometry, formula formats, and chart axes/series.

OCR Provider Pipeline

Run selected rendered pages through a configured local OCR command when scanned or sparse pages need a text layer. OCR commands are environment-configured, so request payloads cannot choose arbitrary executables.

read_pdf can also opt into OCR text layer fusion with include_ocr_text_layer, keeping external OCR text separate from selectable PDF text while linking it into the agent document map. When include_tables is enabled, OCR word boxes can also produce OCR-derived table structure for scanned pages.

Accessibility Report

Use include_accessibility_report when an agent needs page-level accessibility routing for tagged structure, tag-to-visible-content coverage, headings, image alt-text verifiability, form labels, link labels, copy-based accessibility permissions, issue summaries, and page-grade buckets.

Supported Clients

• Claude Desktop - Add to your claude_desktop_config.json
• Claude Code - Use claude mcp add command
• Cursor - Configure in MCP settings
• Any MCP Client - Standard MCP protocol over stdio