Product proof

This is the acquisition-facing evidence board for PDF Reader MCP.

Give your AI agent eyes for PDFs.

One local MCP server returns structured text, tables, OCR paths, visual evidence, and page-level citations.

Install now

npm install -g @sylphx/citra
claude mcp add pdf-reader -- npx @sylphx/citra

Three flagship workflows

1. Financial report / tables

{
  "sources": [{ "path": "/absolute/path/to/10-k.pdf" }],
  "include_tables": true,
  "include_markdown": true,
  "pages": [14]
}

Outcome: table cells with page + geometry — not a prose guess about revenue.

2. Research paper / citations

{
  "sources": [{ "path": "/absolute/path/to/paper.pdf" }],
  "include_full_text": true,
  "include_document_map": true
}

Outcome: reading-order structure and page-linked quotes agents can cite.

3. Scanned document / OCR

{
  "sources": [{ "path": "/absolute/path/to/scan.pdf" }],
  "include_ocr_text_layer": true,
  "include_tables": true
}

Outcome: OCR kept separate from selectable text, with page evidence for verification.

More copy-ready examples: examples/demo.

Why agents need more than text

Without evidence	With PDF Reader MCP
“Revenue was about $12M”	Page 14 · Table 3 · cell (4,2) = $12.4M
Tables become paragraphs	Rows, columns, cells, geometry
Scanned PDFs become noise	OCR path with page-linked evidence
Hidden text ignored	Trust signals when requested

Install footprint (measured)

Clean install on linux-x64 (Node 24):

Metric	TS 3.0.14	Sole-Rust 4.1.x
Full node_modules	~82.3 MiB	~24.4 MiB (~3.4× smaller)
Files	4,101	20 (~205× fewer)
Production deps	PDF.js + MCP SDK + …	{} + one native package

Native binary is multi-megabyte because it is the engine — and still yields a cleaner install than a JS dependency tree.

Performance (bounded, authorized)

Controlled same-host linux-x64 dual-mode A/B vs @sylphx/pdf-reader-mcp@3.0.14 using registry-installed 4.1.0/4.1.1 natives:

Mode	Measures	Result
persistent_warm	long-lived server, repeated identical local read_pdf after warm-up	≥ ~10× on all 8 required fixture classes
startup_inclusive	spawn + initialize + one task	large advantage

persistent_warm includes a process-local cache for identical local path+options. First request still pays full parse cost.

Not a multi-host / RSS / OCR-provider guarantee.

Details:

• 4.1.0 performance report
• claims policy

Engine

Version 4+ uses a native Rust engine via a thin Node launcher.

• Five platforms
• Fail closed if native binary missing
• No TypeScript PDF runtime shipped in production

What stays secondary

Migration ledgers, residual parity families, CI SHA archaeology → migration notes.