Skip to content
superPDF

PDF MCP server

Connect compatible agents to PDF operations.

Runs on your computer. Files stay on your device; the agent or script can only use the folders you allow.

Let AI agents use superPDF tools on PDFs in a folder you choose. superPDF processes the files on your computer and uploads nothing. The agent brings its own model: if it reads a document’s text, that text goes to the agent’s model, which stays on your computer only if the agent runs a local model.

  1. Install the command-line tool

    The superpdf command runs the tools on your computer. It needs Node.js 22 or later. Install it once with npm:

    Available October 14, 2026. The install command below works from that day.

    Terminal
    npm install -g @superpdf/cli

# Check it works
superpdf help

  2. Connect your AI agent

    Add superPDF to your agent's MCP settings. The agent starts superpdf mcp --allow-dir ~/Documents/superPDF on this computer and can only use files in that folder. Change ~/Documents/superPDF to any folder you like.

    Terminal
    claude mcp add --scope user superpdf -- superpdf mcp --allow-dir ~/Documents/superPDF

  3. Try it

    Put a PDF in ~/Documents/superPDF, restart your agent, and ask:

    Use superpdf to split sample.pdf into single pages, then add page numbers to a copy.

Tools the agent gets

ToolWhat it does
pdf_list_filesList PDFs in an allowed folder
pdf_infoPage count, sizes, form fields and properties
pdf_read_textEmbedded text by page, with a character limit
pdf_extract_structureMarkdown, HTML or JSON with headings, paragraphs and a section outline
pdf_mergeCombine 2 or more PDFs into a new file
pdf_splitOne new file per range or per page
pdf_from_officeWord, Excel, PowerPoint or OpenDocument to PDF, with LibreOffice
pdf_extract_pages · pdf_remove_pages · pdf_rotate_pagesPage operations into a new file
pdf_add_page_labels · pdf_add_watermark · pdf_remove_metadataFinishing steps
pdf_protect · pdf_unlock · pdf_repairAES-256 passwords and repair
pdf_run_pipelineUp to 10 steps in one call
pdf_run_batchThe same steps on up to 50 PDFs, one new file each

High-accuracy reading

pdf_read_text and pdf_extract_structure take high_accuracy: true to read the page images with a model on this computer: scanned pages get their text, tables come back as HTML tables and math as LaTeX. Add model: "4b" for the larger, slower model. It takes about 10 to 30 seconds a page with the default model, one document at a time, and sends progress after each page. The server never downloads a model: run superpdf model download in Terminal first (871 MB, it asks before downloading); until then the call fails with that instruction. See PDF CLI.