📄 PDF Reader MCP
Your agent read the PDF. Did it read the truth?
The most-starred PDF MCP server on GitHub. One call turns any PDF into an Agent Document Twin — structured text, tables, trust signals, and source evidence you can search, crop, and cite.
Local-first · One smart read_pdf call · Evidence with page + bbox · 397 tests · 39/39 release-gate checks
⭐ Star this repo if agents should cite PDFs with proof, not guess from plain text. · Quick start · See it work · Roadmap · Why not plain text?
The problem
PDFs are not text files. They are layout, pixels, tables, hidden text, scanned pages, and reading order that breaks the moment you flatten them.
Most PDF tools give agents a text dump. Tables disappear. Scanned pages go blank. Hidden text sneaks in. Citations become guesses. Then the model hallucinates — confidently.
PDF Reader MCP is built for the moment your agent needs to prove an answer, not just sound plausible.
Why not a plain text dump?
| Typical PDF path | PDF Reader MCP |
|---|---|
| Dump text into context | Return markdown, chunks, tables, and a linked document map |
| "Trust the summary" | Page numbers, bounding boxes, crop IDs, and render evidence |
| Hope tables survived | Cells, geometry, confidence, warnings, continuation hints |
| Scanned pages silently empty | OCR path with word boxes and provenance |
| No idea what is risky | Trust report for hidden text, spoofing, unsafe links, injection-like content |
| Ship and pray | 39/39 SOTA release-gate checks on every version |
Full capability matrix: comparison guide.
See it work
Install once. Call once.
claude mcp add pdf-reader -- npx @sylphx/pdf-reader-mcp
{
"sources": [{ "path": "/absolute/path/to/report.pdf" }]
}
read_pdf inspects the PDF, picks the extraction route, and returns the Agent
Document Twin — no manual include_* flags required:
{
"auto_read": {
"workflow": "digital_text_route",
"selected_arguments": {
"include_markdown": true,
"include_tables": true,
"include_chunks": true,
"include_trust_report": true,
"include_document_map": true
}
},
"markdown": "# Annual Report 2026\n\n## Executive Summary\n\n...",
"tables": [
{
"page": 5,
"cells": [
{ "row": 0, "col": 0, "text": "Quarter", "bbox": [72, 650, 180, 670] },
{ "row": 0, "col": 1, "text": "Revenue", "bbox": [200, 650, 300, 670] }
],
"confidence": 0.95
}
],
"trust_report": { "risk_level": "low", "findings": [] }
}
Abbreviated shape — see full example and workflows.
Search, then verify the source region:
{
"sources": [{ "path": "/absolute/path/to/report.pdf" }],
"query": "revenue recognition",
"max_matches_per_source": 10
}
Use the returned page and bounding box with pdf_evidence (render_page or
extract_regions) when the agent needs visual proof before citing.
Why agents use it
| Need | What you get |
|---|---|
| Read the document | Markdown, JSON, HTML, page text, metadata, chunks, and semantic AST. |
| Prove the answer | Page numbers, bounding boxes, evidence IDs, region crops, and source renders. |
| Handle scanned PDFs | Rendered pages routed through configured OCR providers with word boxes and provenance. |
| Recover tables | Selectable-text and OCR-derived tables with cells, geometry, confidence, warnings, and continuation hints. |
| See what text extraction misses | Visual page evidence, focused crops, and configured visual-region provider adapters. |
| Protect the agent | Trust reports for hidden text, prompt-injection-like content, visual spoofing, unsafe links, and redaction. |
| Route accessibility work | Tagged-PDF coverage, tag-visible coverage, headings, images, forms, links, permissions, and page grades. |
| Ship with proof | CI, package smoke, deterministic quality benchmarks, provider artifacts, and release gates. |
Quick Start
Claude Code
claude mcp add pdf-reader -- npx @sylphx/pdf-reader-mcp
Claude Desktop
Add this to claude_desktop_config.json:
{
"mcpServers": {
"pdf-reader": {
"command": "npx",
"args": ["@sylphx/pdf-reader-mcp"]
}
}
}
Any MCP Client
npx @sylphx/pdf-reader-mcp
Node.js >=22.13 is required. The default package works without downloading
OCR models, vision models, Ollama, LM Studio, llama.cpp, or cloud credentials.
Docker
# Pre-built image from GitHub Container Registry
docker run --rm -i -v /path/to/pdfs:/workspace ghcr.io/sylphxai/pdf-reader-mcp
# Or build locally
docker build -t pdf-reader-mcp . && \
docker run --rm -i -v /path/to/pdfs:/workspace pdf-reader-mcp
Need Cursor, VS Code, Windsurf, Cline, Warp, HTTP transport, Docker customization, or filesystem sandboxing? See the installation guide.
MCP Tool Surface
| Tool | Use it when the agent needs to... |
|---|---|
read_pdf |
Use first. With only sources, it auto-inspects and reads the PDF in one call; with explicit include_* options, it runs precise manual extraction. |
search_pdf |
Search selectable text and optional OCR text with snippets, offsets, boxes, and provenance. |
pdf_evidence |
One focused evidence tool for inspect, render_page, extract_regions, ocr_pages, and analyze_regions operations. |
Full request and response details live in the API reference.
Agents can force auto: false for precise manual extraction, or use
auto_detail: "fast", "balanced", or "full" to control output depth without
learning dozens of switches.
Agent Document Twin
The Agent Document Twin is the main reason to use this project instead of a plain text extractor. It keeps the document readable by agents while preserving the evidence needed to verify the answer.
| Layer | Output |
|---|---|
| Lossless PDF layer | Text runs, lines, words, characters, fonts, transforms, page geometry, metadata coverage, outlines, forms, attachments, annotations, permissions, and structure signals where available. |
| Visual layer | Page renders, region crops, crop provenance, visual candidates, OCR source renders, and provider-normalized visual evidence. |
| Semantic layer | Page, section, paragraph, list, caption, header, footer, table, image, chart, formula, figure, and diagram nodes where available. |
| Evidence layer | Stable IDs, page ranges, bounding boxes, crop IDs, confidence, warnings, and extraction method provenance. |
| Agent layer | Markdown, JSON, HTML, citation chunks, routing plans, trust report, accessibility report, and document map indexes. |
Example: Read With Evidence
{
"sources": [{ "path": "/absolute/path/to/report.pdf" }],
"include_markdown": true,
"include_chunks": true,
"include_tables": true,
"include_text_layer": true,
"include_document_map": true,
"include_document_ast": true,
"include_trust_report": true,
"include_accessibility_report": true
}
Provider-Enabled Intelligence
The current package stays local-first. The roadmap target is a Rust MCP server with the same public tool contract, plus optional deployment-controlled providers for OCR and visual enrichment.
| Capability | Default behavior | Enable with |
|---|---|---|
| Selectable-text PDFs | Works out of the box | No extra dependency |
| Rendering and crops | Works out of the box | No extra dependency |
| Trust and accessibility reports | Works out of the box | No extra dependency |
| OCR for scanned pages | Provider-ready | MCP_PDF_OCR_* |
| Visual table/chart/formula/figure/image enrichment | Provider-ready | MCP_PDF_REGION_ANALYSIS_* |
Supported visual provider paths include local commands, local HTTP servers, Ollama, OpenAI-compatible endpoints, LM Studio, and llama.cpp. Request payloads cannot choose arbitrary executables or arbitrary provider URLs; providers are configured by the deployment environment.
# Example shape only. Point these at your own local OCR command.
export MCP_PDF_OCR_COMMAND="tesseract"
export MCP_PDF_OCR_ARGS_JSON='["{input}", "stdout", "tsv"]'
See the guide and API reference for provider configuration details.
Release Proof
Claims are backed by