Skip to content

index.md

Latest commit

 

History

History
64 lines (54 loc) · 3.79 KB

index.md

File metadata and controls

64 lines (54 loc) · 3.79 KB
Raw

Capability Overview

PDF Reader MCP is designed as a full-fidelity PDF intelligence layer for agents. The comparison below is category-based and focuses on the agent workflow: read a smart Agent Document Twin first, search cheaply when the task has a literal query, and request focused evidence only when the answer needs source-level proof.

Capability PDF Reader MCP Text/CLI tools Cloud PDF APIs Generic filesystem MCP
MCP-native PDF tools ✅ V3 three-tool surface ❌ ❌ ⚠️ raw file access only
Preflight inspection and routing ✅ ❌ ⚠️ API-specific ❌
Literal search with evidence ✅ snippets, offsets, boxes, provenance ⚠️ text only ⚠️ varies ❌
Text layer fidelity ✅ runs, lines, words, chars, metadata coverage ⚠️ usually text only ⚠️ varies ❌
Agent Document Twin ✅ document map plus AST and evidence indexes ❌ ⚠️ vendor-specific ❌
Page rendering evidence ✅ bounded MCP image parts ⚠️ external commands ✅ ❌
Region crop evidence ✅ PDF-coordinate crops ⚠️ custom glue ✅ ❌
Scanned-page OCR path ✅ configured local provider with provenance ⚠️ external glue ✅ ❌
OCR-derived tables ✅ when OCR word boxes are available ❌ ⚠️ varies ❌
Table quality diagnostics ✅ cells, geometry, spans, warnings, continuation hints ❌ ⚠️ varies ❌
Formula/chart/figure/image enrichment ✅ configured visual-provider adapters ❌ ⚠️ vendor-specific ❌
Trust report ✅ hidden text, prompt-injection-like text, visual spoofing, unsafe links, redaction ❌ ⚠️ varies ❌
Accessibility report ✅ tagged-PDF, tag-visible coverage, forms, links, images, permissions, grades ❌ ⚠️ varies ❌
Citation chunks ✅ page, semantic, size, and table chunks ❌ ⚠️ varies ❌
Local-first default ✅ ✅ ❌ ✅
No required API key ✅ ✅ ❌ ✅
Reproducible release proof ✅ quality, corpus, provider, package-smoke, and release-gate artifacts ❌ ❌ ❌

Why It Matters

Agents need more than extracted text. For high-value PDFs they need to know where content came from, which page or crop proves it, whether the reading order looks uncertain, whether a page needs OCR, whether a table has weak geometry, and whether hidden or unsafe content should be treated as untrusted data.

PDF Reader MCP exposes that as a compact V3 tool surface:

  1. read_pdf is the default entrypoint. With only sources, it profiles the PDF, chooses useful extraction options, and returns the linked Agent Document Twin.
  2. search_pdf finds source-backed text matches before spending more context on broad extraction or visual proof.
  3. pdf_evidence handles focused follow-up operations: inspect, render_page, extract_regions, ocr_pages, and analyze_regions.

When PDF Reader MCP Is The Better Fit

  • You want agents to start with one intelligent PDF read instead of learning a long list of extraction tools.
  • You need stable page, element, chunk, crop, table, OCR, trust, and accessibility references for downstream citations.
  • You need local-first execution and want OCR or visual models configured by the deployment, not selected by each request.
  • You need source evidence for tables, charts, formulas, figures, and scanned pages.
  • You want public benchmark artifacts and a release gate that prove the shipped capability surface.

Boundaries

PDF Reader MCP does not bundle heavy OCR, vision, formula, or layout model weights. The default package stays TypeScript-first and local-first; advanced OCR and visual understanding are enabled through explicit local providers and validated through provider benchmarks.