An MCP server that gives any agent — Claude Desktop, Claude Code, Cursor, Windsurf, Zed, your own — a local, private PDF → Markdown tool backed by pdffr. Born-digital pages decompile from glyph geometry in milliseconds; scans and screenshots get on-device OCR only where the text layer can't explain the ink. Nothing leaves the machine.

Install

Claude Desktop / Claude Code / Cursor / Windsurf — add to your MCP config:

{
  "mcpServers": {
    "pdffr": {
      "command": "npx",
      "args": ["-y", "pdffr-mcp"]
    }
  }
}

Claude Code one-liner:

claude mcp add pdffr -- npx -y pdffr-mcp

Requires Node 20+. The first run downloads the OCR model for the requested language (English by default) into the OS temp directory.

Tools

`pdf_to_markdown`

Decompile a PDF to Markdown. Give it one of path, url or base64.

Argument	Type	Notes
`path`	string	absolute or relative file path
`url`	string	fetched with the server's network access
`base64`	string	raw PDF bytes
`pages`	string	e.g. `"1-3,7"`; default all
`ocr`	boolean	escalate unexplained ink to OCR (default `true`)
`lang`	string	tesseract language(s), e.g. `"deu"`, `"eng+ara"` (default `"eng"`)
`maxChars`	number	truncate the result (default 200 000) — the response says if it did

Returns the Markdown, followed by a one-line stats footer (pages, time, OCR regions).

`pdf_outline`

Headings only — level, text and page — so an agent can decide which pages to read before paying for them. Same inputs (path / url / base64, ocr, lang).

`pdf_tables`

Every table as JSON rows ([{ page, rows: string[][] }]). Same inputs.

Why not a vision model

A PDF already knows where every glyph is. pdffr reads that geometry directly and reserves OCR for the pixels the text layer doesn't explain, so a 30-page report comes back in well under a second and a scanned invoice in a couple of seconds — with no API key and no upload. See the architecture notes.