pdf-mcp

pdf-mcp is an open-source server and Python package designed to help AI agents efficiently extract, search, and analyze content from PDF documents without loading the entire file into an AI model's context. It addresses the challenge of handling large PDFs, which can overwhelm token limits for AI assistants, by allowing targeted access to only the relevant pages or sections needed for a given task. This approach enables agents to skim, search, and read specific parts of a document, much like a human would, rather than processing the whole document at once.

The tool provides a suite of nine distinct features, including document metadata extraction, keyword and hybrid semantic search, and selective page reading. With pdf_info, agents can retrieve metadata such as page count, table of contents, text coverage per page, and estimated token counts before deciding what to read. The pdf_search function supports both keyword and hybrid semantic search (using reciprocal rank fusion and BM25 or fastembed), returning structural excerpts like paragraphs or headings that match the query. When deeper context is required, pdf_read_pages allows extraction of full text, images, and tables from specified pages. Additional tools cover full-document reads, table of contents extraction, visual rendering of pages as PNGs, cache management with persistent SQLite storage, and discovery of installed features.

pdf-mcp supports advanced features such as CJK keyword search for Japanese, Chinese, and Korean text, OCR for scanned pages using Tesseract, multi-column and vertical layout handling (including tategaki Japanese), structured extraction of tables and images, bounding-box evidence for search hits and content, and detection of hidden or invisible text. The server can render pages for vision models and maintains a persistent cache to improve performance across restarts.

The platform is delivered as a Python package installable via pip and is compatible with various AI development environments and MCP (Model Context Protocol) clients, including Claude Desktop, Claude Code, Cursor, VS Code, Codex, and Kiro. A browser-based demo simulates three of the main tools entirely client-side, ensuring that uploaded PDFs do not leave the browser tab. pdf-mcp is released under the MIT license, with its source code available on GitHub and distribution via PyPI.