Pdfmux

What NameetP/pdfmux MCP server does

NameetP/pdfmux MCP server provides PDF extraction through a router that evaluates pages individually instead of applying one parser to an entire document. It distinguishes common page types, including digital text, scanned content, tables, and complex layouts, then selects an available backend such as PyMuPDF, RapidOCR, Docling, OpenDataLoader, Surya, Marker, or an optional LLM provider.

The extraction pipeline checks its own output. Pages with weak confidence, missing text, scrambled columns, or broken tables can be sent through a stronger extraction path. Results include confidence information so applications can identify documents or pages that need review rather than treating every successful process exit as reliable.

A separate verification workflow can inspect output created by another extraction engine. It compares the supplied JSON, Markdown, or plain-text result with the source PDF and reports coverage, confidence, page verdicts, alignment, hallucination risk, and table or heading integrity when available.

How it works

NameetP/pdfmux MCP server can process a single file, a directory, or a stream of pages. Its CLI supports ordinary Markdown extraction, token-limited chunks for RAG, schema-guided extraction, cost estimation, profiles, directory watching, extraction diffs, caching, and strict confidence checks. The Python API includes single-file helpers and a batch iterator that returns each path with either a result or an exception.

The default installation handles digital PDFs. Optional extras add OCR, table extraction, complex-layout processing, neural academic-paper extraction, watching, and LLM providers. LLM use is optional and can target Gemini, Claude, OpenAI models, Ollama, or Mistral OCR when the corresponding extras and credentials are configured.

Setup and configuration

Install the base package with pip install pdfmux on Python 3.11 or newer. For scanned documents, the README recommends pdfmux[ocr]; other extras are available for tables, OpenDataLoader, Marker, individual LLM providers, watching, or the complete set of dependencies.

Basic conversion requires no provider credentials. LLM fallback modes use the user's own provider access, while Ollama supports a local model. The README does not specify MCP configuration entries or environment-variable names, so client setup details should be taken from the package's current documentation.

Tools and capabilities

• Route pages across multiple extraction backends.
• Audit extraction quality and assign confidence scores per page.
• Retry pages that produce low-quality or incomplete output.
• Produce Markdown, structured JSON, and token-limited chunks.
• Verify extraction output from pdfmux or another engine.
• Process batches with manifests, streaming, caching, and strict confidence gates.

The built-in MCP server makes the package available to MCP clients, and the README specifically identifies Claude Desktop as a target client. The provided material does not list the MCP tool names or their parameter schemas.

Limitations and notes

OCR, table, neural, and LLM features are not all included in the base installation. Some backends add substantial dependencies, and OpenDataLoader requires Java 11 or newer. LLM and Mistral OCR paths may incur provider costs or require separate credentials. Verification quality depends on the structure available in the supplied extraction: page-level comparison is used when page structure is exposed, while single-blob input uses content-presence checks.

NameetP/pdfmux MCP server is MIT licensed. The repository describes the method as patent-pending and presents the project as an open-source alternative for PDF extraction workflows.