SylphxAI/pdf-reader-mcp

npm install -g @sylphx/pdf-reader-mcp (installs the native binary for your platform only; missing native package fails closed)
claude mcp add pdf-reader -- npx @sylphx/pdf-reader-mcp (Claude Code)
JSON config with command npx @sylphx/pdf-reader-mcp for Claude Desktop / Codex / Cursor / VS Code / any MCP client

Usage

• •Three tools: read_pdf (smart default — markdown, tables, structure, OCR, citations), search_pdf (page + snippet matches before deep reading), pdf_evidence (crops, renders, focused evidence ops)
• •Minimal call: {"sources": [{"path": "/absolute/path/to/report.pdf"}]}
• •HTTP transport: MCP_TRANSPORT=http pdf-reader-mcp
• •Programmatic SDK: import { Citra } from '@sylphx/pdf-reader-mcp/sdk' — citra.read({ sources: [...] })

Key features

• ✓Agent Document Twin: markdown + structure + geometry + provenance (e.g., 'Page 14, Table 3, cell (row 4, col 2) = $12.4M')
• ✓Native Rust engine (v4) via a thin Node launcher; ~3.4x smaller install and ~205x fewer files than the prior TS version
• ✓OCR path for scanned documents with page-linked evidence; table rows/columns/cells with bounding boxes
• ✓Trust signals / prompt-injection handling when requested; handles broken ToUnicode CMaps without crashing
• ✓Dual-mode stdio/HTTP; supports dual-era hosts that send server/discover before initialize (e.g. Gemini Antigravity CLI)

Best for

Agents that need citeable, geometry-accurate evidence from PDFs — financial reports, research papers, and scanned documents.

Caveats

• ⚠Local-first and platform-specific: one native binary per platform (macOS arm64/x64, Linux x64/arm64, Windows x64); missing native package → fail closed (no silent engine switch)
• ⚠Performance numbers are same-host measurements — explicitly 'not a multi-host guarantee'
• ⚠Prerelease / transitional package name @sylphx/pdf-reader-mcp (bin pdf-reader-mcp); individual products only, no central monorepo