pdf-mcp

Agentic RAG over your PDFs, one file or a whole folder, as a single MCP tool.

The agent decides when to search; pdf-mcp does the retrieval and hands back excerpts. It is an MCP server that lets Claude Code and other AI agents search one PDF or a whole folder 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, with optional CUDA acceleration for warming large corpora.

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?

Installation

pip install pdf-mcp

That is the whole install: hybrid search, corpus tools, multi-column and CJK reading order all work out of the box.

OCR on scanned PDFs additionally needs system Tesseract:

brew install tesseract        # macOS
apt install tesseract-ocr     # Ubuntu/Debian
winget install Tesseract-OCR  # Windows

GPU embedding is optional and off by default. On an NVIDIA card it makes the embedding pass one to two orders of magnitude faster; set PDF_MCP_CUDA=1 after installing the CUDA build of onnxruntime. Setup per CUDA series is in docs/configuration.md.

Quick Start

claude mcp add pdf-mcp -- pdf-mcp

Then ask Claude to read a PDF. For Claude Desktop, VS Code, Codex CLI, Kiro, or any other MCP client, see docs/clients.md.

pdf-mcp's tools are also plain Python functions, so you can import them and hand a PDF to the Anthropic SDK without running a server. Two runnable scripts, for a question and for a whole document: examples/.

Why this exists, and what broke along the way: Claude's 100-page PDF limit and how I got around it

Tools

13 specialized tools rather than one monolithic one. Typical pattern: pdf_info to plan, pdf_search to locate (its paragraph excerpts often answer the question outright), pdf_read_pages when you need more. For a folder, pdf_corpus_overview to triage, then pdf_corpus_search.

Text returned by any of these is untrusted content extracted from a PDF. pdf_info(content_trust=True) reports hidden text a human reader cannot see, and the read tools flag it per page.

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"

Full reference, every parameter and response shape: docs/tool-reference.md. Embedding model selection: 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 is the default and is what every example above uses. pdf-mcp-http serves the same tools over HTTP, for clients that cannot spawn a process (the Anthropic API MCP connector, claude.ai custom connectors) and for a warm corpus shared by several clients.

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

Paths resolve on the server, so an HTTP agent 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. It is single-tenant and fails closed: with no auth token and no [paths] allow list, the process exits rather than serving an open endpoint.

Docker images are published to GHCR for amd64 and arm64, with everything baked in, so every tool works on the first request:

./deploy.sh              # token, image, start, health-check
cp your.pdf documents/   # this folder is the server's /data/pdfs

Read docs/remote-access.md for the trust boundary and threat model before deploying, and docs/configuration.md for setup, client config, and token rotation.

Configuration

pdf-mcp works out of the box. To restrict which paths and URL hosts the server may touch, tune cache and worker settings, or add your own content-trust phrases, see docs/configuration.md.

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.

Contributors

Thank you to everyone who has helped improve this project through code, reviews, testing, and feature requests:

@Summer907 · @ebbsanchez · @VooDisss · @DerDennisOP · @deepdmk · @TheSOV

Per-release contributor credits are listed in the Changelog.

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

The story behind the releases. Building pdf-mcp keeps surprising me: benchmarks that go the wrong way, formats that break everything, features I had to remove. I write about that thinking in The Dispatch. Come along if that's your kind of thing.

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 to Send PDFs Over 100 Pages to Claude's API: Measuring the real page and token ceilings on the Claude API, and the two ways around them: search the PDF on disk for questions, window the text and render only the picture pages for whole-document tasks
• 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

Corpus & multi-document search

• 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
• Cross-Document Retrieval for AI Agents Without a Vector Database: Why BM25 scores don't merge across per-document indexes but ranks do, and how two-stage RRF puts a gold document in the top 3 on 84.8% of 184 graded queries over a 100-PDF corpus

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
• How I Fixed Vertical Japanese PDF Extraction: Tategaki pages extract scrambled because reading order is geometric, not stored; rebuilding it from glyph positions (columns right to left, characters top to bottom), with no OCR and no new dependency