pdf-mcp

Surgical PDF access for AI agents: search, read, and extract without flooding context.

An MCP server that lets Claude Code and other AI agents search a PDF by meaning or keyword, read only the pages that matter, and cleanly pull out tables, images, and scanned text, even from multi-column and Japanese layouts.

mcp-name: io.github.jztan/pdf-mcp

Try it in your browser

See what your AI agent sees →

Drop in any PDF, or a whole folder of them, and watch an agent triage the corpus, search across every document at once, and read only the pages that matter, using a fraction of the tokens. 100% client-side, no install required.

Why pdf-mcp?

Features

• Hybrid search: find relevant pages with a question, not a page range. Combines BM25 keyword and semantic search via Reciprocal Rank Fusion
• Corpus search: point the server at a folder of PDFs: warm them into the cache, get per-document triage cards, and search across all documents at once with ranked, document-attributed hits
• Paginated reading: fetch only the pages your agent needs; large documents don't blow your context window
• OCR: scanned and image-based PDFs are fully readable and searchable via Tesseract, parallelized across pages for ~2–3x faster extraction on typical scans
• Structured extraction: tables, embedded images, and table of contents returned as structured data, not text soup
• Chart data extraction: pull exact (x, y) tables from vector charts, read from the plot geometry rather than guessed from the image; declines with a rendered image when a chart can't be read reliably
• Vertical-script reading order: Japanese tategaki (縦書き) reconstructed from glyph geometry into correct top-to-bottom, right-to-left order; article segmentation for dense magazine layouts; mojibake filtered
• Persistent cache: SQLite-backed; re-reads are instant and survive server restarts
• Secure URL fetching: HTTPS-only with SSRF protection; local network ranges are blocked
• Content-trust / hidden-text detection: flags text a human reader can't see (invisible render mode, sub-point fonts, transparent or white-on-white fill, off-page) so an agent treats it as untrusted rather than vetted. Flag-only: nothing is stripped

Contents

• Installation
• Quick Start
• Tools
• Example Workflow
• Remote / HTTP transport
• Configuration
• Roadmap
• Contributing
• Security
• License

Installation

pip install pdf-mcp

Semantic search is included by default (hybrid auto search is built on it; ~67 MB embedding model download on first use). The former [semantic] and [cjk] extras remain as no-op aliases. Platform note: the bundled onnxruntime has no wheels for Intel macOS on Python 3.14+ or Alpine/musl; use Python ≤ 3.13 there.

For correct reading order on multi-column PDFs (adds pymupdf4llm, which pulls pymupdf_layout/onnxruntime):

pip install 'pdf-mcp[multicolumn]'

Without it, multi-column pages fall back to positional-sort extraction, which can interleave columns.

Japanese/Chinese/Korean PDFs work out of the box: keyword search uses a char-split FTS index that matches unspaced CJK terms, and semantic CJK search is covered by the default install.

For OCR on scanned PDFs (requires system Tesseract):

# macOS
brew install tesseract

# Ubuntu/Debian
apt install tesseract-ocr

# On Windows, download the installer from:
# https://github.com/UB-Mannheim/tesseract/wiki
# Then add the install directory to your PATH.

Quick Start

Choose your MCP client below to get started:

claude mcp add pdf-mcp -- pdf-mcp

{
  "mcpServers": {
    "pdf-mcp": {
      "command": "pdf-mcp"
    }
  }
}

{
  "mcpServers": {
    "pdf-mcp": {
      "command": "pdf-mcp"
    }
  }
}

code --add-mcp '{"name":"pdf-mcp","command":"pdf-mcp"}'

{
  "servers": {
    "pdf-mcp": {
      "command": "pdf-mcp"
    }
  }
}

codex mcp add pdf-mcp -- pdf-mcp

[mcp_servers.pdf-mcp]
command = "pdf-mcp"

{
  "mcpServers": {
    "pdf-mcp": {
      "command": "pdf-mcp",
      "args": [],
      "disabled": false
    }
  }
}

{
  "mcpServers": {
    "pdf-mcp": {
      "command": "pdf-mcp"
    }
  }
}

{
  "mcpServers": {
    "pdf-mcp": {
      "command": "uvx",
      "args": ["pdf-mcp"]
    }
  }
}

Verify Installation

pdf-mcp --help

Tools

The typical pattern: call pdf_info first to plan, then pdf_search to locate; its paragraph excerpts are often enough to answer directly. Use pdf_read_pages or pdf_read_all when you need deeper context. For a folder of PDFs, start with pdf_corpus_overview to triage, then pdf_corpus_search to search across documents.

Example prompts:

"Read the PDF at /path/to/document.pdf"
"Which pages discuss supply chain risks?"
"Find sections about the training process"
"Show me what page 5 looks like"
"OCR pages 3-5 of the scanned PDF"

See docs/tool-reference.md for the complete reference: every parameter, response shape, security contract, and example. For semantic-search model selection, see docs/embedding-models.md.

Example Workflow

For a large document (e.g., a 200-page annual report):

User: "Summarize the risk factors in this annual report"

Agent workflow:
1. pdf_info("report.pdf")
   → 200 pages, TOC shows "Risk Factors" on page 89

2. pdf_search("report.pdf", "risk factors")
   → Matches with structural paragraph excerpts: each excerpt
     is the bullet, paragraph, or heading that matched, not a
     fixed-width window. Often enough to answer directly.

3. If excerpts are sufficient → synthesize answer

4. If more context needed:
   pdf_read_pages("report.pdf", "89-95")
   → Full page text for deeper reading

Remote / HTTP transport

STDIO remains the default and is what every example above uses. A second entry point serves the same tools over HTTP, but the two transports suit different jobs:

export PDF_MCP_AUTH_TOKEN="$(openssl rand -hex 32)"
pdf-mcp-http

Because paths resolve on the server, an agent connected over HTTP reads what is already there: files under an allow-listed root, or a URL the server fetches. It cannot hand over a file from its own machine. Call server_info to discover the roots a server will open. See Getting documents to the server.

It is single-tenant and fails closed: without an auth token and a [paths] allow list, the process exits rather than starting an open endpoint. Before you deploy it, read docs/remote-access.md for the trust boundary, the threat model versus stdio, and the token revocation runbook.

Docker

./deploy.sh              # generates .env with a token, pulls the image, starts, health-checks
cp your.pdf documents/   # the intake path: this folder is the server's /data/pdfs

The image is published to GHCR for amd64 and arm64, so nothing is compiled locally. Everything is baked in (OCR, column-aware extraction, embedding model), so all tools work on the first request. The container runs as a non-root user and publishes to host loopback only; put a TLS proxy in front for public access.

./deploy.sh --help lists the lifecycle commands, including --build to build locally instead of pulling. For the environment variables (host port, image tag, auth token) and the deployment guards, see docs/configuration.md.

Configuration

pdf-mcp works out of the box with no configuration. To restrict which paths and URL hosts the server can access, tune cache and worker settings, or understand what's cached, see docs/configuration.md.

• Access control: ~/.config/pdf-mcp/config.toml allow/deny rules for paths and URLs, plus response byte caps
• Content-trust phrases: extend the hidden-text injection_in_hidden hint with your own (including non-English) phrases via [content_trust].injection_phrases
• Environment variables: cache directory, TTL, and parallel OCR/render worker count
• Caching: SQLite-backed persistence, what's cached, and invalidation

Roadmap

See ROADMAP.md for planned features and release history.

Contributing

Contributions are welcome. See docs/contributing.md for setup, checks, the coherence eval harness, and quality-loop guidelines.

Security

Found a vulnerability? See SECURITY.md for the threat model, reporting channel, and expected response timeline. Please do not open a public GitHub issue for unpatched security reports.

License

MIT. See LICENSE.

Links

• pdf-mcp on PyPI
• pdf-mcp on GitHub

Blog posts

Background, benchmarks, and design notes from building pdf-mcp:

Getting started

• How I Built pdf-mcp: The problem with large PDFs in AI agents and a working solution
• How Claude Code Actually Reads PDFs: How AI agents use pdf-mcp tools to read and navigate PDF documents
• How AI Agents Should Read PDFs: 5 Patterns That Survived Production: Five production-tested patterns for how agents should navigate PDFs at scale
• A Knowledge Base Is Just a Folder: Turning a folder of PDFs into an agent knowledge base with the corpus tools, no ingestion pipeline or vector store

Search & retrieval

• Semantic vs Keyword Search for AI Agents: Benchmarks and a dual-search routing pattern: FTS5 for exact identifiers, embeddings for natural language
• Hybrid Search vs Query Routing for AI Agents: Why pdf-mcp uses hybrid RRF instead of query routing: benchmarks showing RRF wins across query types
• Section Chunking vs Page Chunking for AI Agents: Why section-aware search delivers full section content in one call while page-mode costs 2–6 extra tool calls per query
• Section-Level RAG: Why BM25 Beat Hybrid Search in My Benchmark: Why pdf-mcp's section-grain search is BM25-only: hybrid RRF caused a 33% lexical regression at section grain, so granularity decides the search technique
• How One Search Change Eliminated an Entire Agent Step: Switching pdf_search from fixed-width snippets to paragraph excerpts turned it from a pivot tool into a terminal tool: 97% vs 80% answer containment across a 30-query benchmark

Engineering & security

• MCP Server Security: 8 Vulnerabilities: What we found when we audited an MCP server for security holes
• Your LLM Is Free QA for Your MCP Server: Four Payload UX bugs in pdf-mcp that schema tests missed but Claude Desktop surfaced during real use
• Why Multi-Column PDFs Scramble Reading Order in RAG: Fixing two-column extraction (0.564 → 0.816 fidelity), the title-page author-grid regression it caused, and the aggregate metric that stayed blind to both