📄 @sylphx/pdf-reader-mcp

The PDF intelligence layer for AI agents that need source evidence, not just extracted text.

V3 smart tool surface · Agent Document Twin · Evidence-first extraction · Visual crops · OCR adapters · Tables, charts, formulas, figures · Trust & accessibility reports · Benchmark-gated releases

---

PDFs are not plain text files. They are layout, pixels, tables, hidden text, permissions, annotations, scanned pages, and ambiguous reading order.

PDF Reader MCP turns that mess into an Agent Document Twin: a linked, source-backed representation of the PDF that agents can inspect, search, verify, crop, OCR, enrich, cite, and read with confidence.

If your agent has ever hallucinated from a PDF, lost a table, trusted hidden text, missed a scanned page, or needed to cite the exact region that proves an answer, this is the MCP server for that workflow.

Why Agents Use It

Need	What PDF Reader MCP gives you
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.

Need Cursor, VS Code, Windsurf, Cline, Warp, HTTP transport, Docker, or filesystem sandboxing? See the installation guide.

One Smart Tool First

The default V3 agent path is one tool call:

{
  "sources": [{ "path": "/absolute/path/to/report.pdf" }]
}

With no manual include_* flags, read_pdf profiles each PDF, chooses the extraction route, and returns the Agent Document Twin in one response. Digital text PDFs get Markdown, chunks, tables, layout routing, and source evidence. Mixed or scanned PDFs are routed toward configured OCR and visual providers when those providers are ready. Metadata, page geometry, warnings, provider readiness, and the selected read_pdf arguments are included so the agent can see what happened.

Agents can still force auto: false and use explicit include_* options for a precise manual extraction. Use auto_detail: "fast", "balanced", or "full" when the agent wants to control output depth without learning dozens of switches.

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.

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
}

Example: 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 operation render_page or extract_regions when the agent needs visual proof before citing or summarizing.

Provider-Enabled Intelligence

The default package stays TypeScript-first and local-first. Heavy engines are optional, deployment-controlled adapters.

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

Strong README claims should be backed by shipped evidence. This repo publishes machine-readable artifacts and gates releases on them.

Artifact	Current proof
pdf_sota_release_gate.json	passed, 39/39 release-gate checks passing
pdf_quality_benchmark.json	score 1, 69/69 deterministic quality checks passing
pdf_provider_benchmark.json	strict provider evidence enabled, 4/4 final-bar provider profiles certified
pdf_corpus_benchmark.json	corpus-style PDF intelligence assertions with capability summaries
pdf_provider_manifest_crop_benchmark.json	deterministic crop-substrate proof for provider-manifest regions
pdf_provider_manifest_benchmark.json	deterministic scoring proof for table, formula, chart, figure, and image regions

Run the same proof locally:

bun run benchmark:release-artifacts
bun run benchmark:release-gate
bun run package:smoke

See performance and release evidence for the full benchmark contract.

Output Formats

read_pdf can return the same PDF in several agent-friendly forms:

• Plain text and page text
• Markdown for RAG and summarization
• HTML for rendering or downstream transformation
• Structured elements with page and geometry provenance
• Document AST for semantic navigation
• Citation chunks with page, element, table, and bbox references
• Tables with rows, cells, geometry, warnings, and confidence
• Trust and accessibility reports
• Agent Document Twin indexes linking text, visual, OCR, table, trust, and accessibility evidence

Security Model

PDFs can contain hostile or misleading content. The server treats extraction as an evidence workflow, not as a trusted text dump.

• Local-first by default.
• URL loading is guarded by host, private-IP, size, and HTTP policy controls.
• OCR and visual providers are configured by environment, not by request body.
• Trust reports surface hidden text, near-invisible geometry, off-page text, overlapping text, unsafe links, redaction signals, and prompt-injection-like content.
• Rendering, crops, OCR, and visual enrichment preserve provenance so agents can route weak evidence to verification instead of silently trusting it.

Documentation

Topic	Link
Getting started	docs/guide/getting-started.md
Installation and clients	docs/guide/installation.md
API reference	docs/api/README.md
Capability overview	docs/comparison/index.md
Architecture and design	docs/design/index.md
Performance and release proof	docs/performance/index.md

Development

git clone https://github.com/SylphxAI/pdf-reader-mcp.git
cd pdf-reader-mcp
bun install
bun run build
bun test

Useful checks:

bun run check
bun run typecheck
bun run docs:build
bun run package:smoke
bun run benchmark:release-gate

Support

• Issues
• Discussions
• npm package

If you want local-first, evidence-backed PDF intelligence to keep improving for AI agents, star the repo. It helps the project reach more builders who need PDFs to be verifiable, not just readable.

License

MIT © SylphxAI

Star History