Let your AI agent read and process PDFs locally on your Mac

Joy Truepath · 4 July 2026

You've got a signed contract, a bank statement, a 120-page report someone needs split into sections. You'd love to point an AI at it — "find every mention of the renewal clause," "pull pages 40 to 60 into their own file." What you don't love is the first step every web tool asks for: upload it. Some PDFs you just don't hand to a server you don't control.

So don't. TruePath PDF ships an MCP server that runs on your own machine, and it installs with a single command. Once it's wired to your agent, Claude, Cursor, or Codex can open the PDF, read it, and reshape it — with every page staying exactly where it is.

What MCP is, briefly

The Model Context Protocol lets an AI host call outside tools mid-conversation. An MCP "server" exposes a handful of actions the host can use. Here the server is a small local program that knows how to work with PDFs, so the tool your agent reaches for is running on your Mac, not in someone's cloud.

What the agent can actually do

Nine tools in the free tier, and they fall into four jobs:

• Read — get_info for page count and metadata, extract_text with per-page markers so the model cites the right page, and search, a grep-style scan that returns hits with surrounding snippets.
• Reshape — split at a page, merge several files, and pages for page-level operations.
• Rasterise — to_images to turn pages into images, and extract_images to pull out the pictures already embedded.
• Hand off — open_in_truepath opens the file in the TruePath PDF app when you want to finish something by hand.

The page markers on extract_text are the small touch that pays off: because each chunk is labelled with its page, the agent can answer "which page says X" instead of just "the document says X."

Setting it up

There's nothing to build. The server is on npm, so you install it with one command — you just need Node.js 20 or newer.

npx -y @truepathpdf/mcp-server

For Claude Desktop, add it to the config:

{
  "mcpServers": {
    "truepath-pdf": {
      "command": "npx",
      "args": ["-y", "@truepathpdf/mcp-server"]
    }
  }
}

Restart the host and ask in plain language:

Use truepath-pdf to find every mention of "indemnity" in ~/Documents/contract.pdf.

The privacy part, because it's the whole point

The reason to run this locally is the same reason you hesitated to upload in the first place: the PDF is often the sensitive thing. So the server makes no network calls at all — no telemetry, no account, and the free tools work offline, forever, with no key. It reads and rewrites the files you name, and its reach stops there.

Why this shape

If you already work inside an agent, a local PDF server means the document becomes something the assistant can actually handle — read a clause, split a report, pull the figures — without a cloud dependency you'd later have to explain to whoever owns that document. Not dramatic. Just one fewer place your files have to travel.

A few things people ask

Does anything get uploaded?
No. The server runs locally and makes no network calls. The free tools work offline with no key.

Do I need the app?
Not for the eight core tools — the server installs on its own via npx. Only open_in_truepath needs the Mac app.

Which agents work with it?
Any MCP client — there's config for Claude Desktop, Claude Code, Cursor, and Codex.