juergenkoller-software/pdf-content-search-mcp

**Full-text PDF search + OCR for Claude, Cursor, and any MCP client.**
This is the official Model Context Protocol bridge for **PDF Content Search** — a native macOS app that indexes thousands of PDFs (including scanned ones via Apple Vision OCR) and finds anything in under a second.
> **You need the PDF Content Search app installed and running.** This MCP server is a stdio→HTTP bridge — the index, OCR engine, and search live in the app. Get PDF Content Search at store.juergenkoller.software/apps/pdf-content-search.
---
What you can do
> "Claude, find every invoice from Telekom in 2024 across all my PDFs and list amount + date."
>
> "Cursor, search 'data processing agreement' across my contracts folder, return the matching PDF + page number."
The app exposes search, filter, OCR, indexing, and document-metadata tools through MCP. Search results include page numbers, highlighted snippets, and bounding-box coordinates from OCR. Full tool list is announced by the server on tools/list after initialize.
---
Installation
Prerequisites
1. **macOS 12 (Monterey) or later**
2. **PDF Content Search app installed and running** — get it here (one-time purchase €89.99, 30-day free trial, includes free iOS + Android scanner companion apps)
3. **Swift 5.9+** (Xcode 15+) if building from source
Build from source
Universal binary (for shipping inside the app bundle):
Already inside the app
If you have PDF Content Search installed, the bridge ships inside the app bundle at Contents/Resources/pdf-content-search-mcp. Use the "**Set up Claude Desktop**" button in the app to auto-configure Claude Desktop with the right path and token — that's the recommended path for end users.
This repo is the **source code** for that bundled bridge, kept open so you can audit it, fork it, or run it under your own sandbox.
---
Configuration
Claude Desktop
Get PDF_CS_TOKEN from **PDF Content Search → Settings → API Server**, or use the in-app "Set up Claude Desktop" button.
Claude Code
Cursor / other MCP clients
Same pattern — stdio MCP server.
---
Environment variables
| Variable | Default | Description |
|---|---|---|
| PDF_CS_TOKEN | _(none)_ | Bearer token matching the app's API Server token (required) |
| PDF_CS_PORT | 44477 | Port of the app's API server |
| PDF_CS_HOST | 127.0.0.1 | Host to reach the app |
---
Protocol behavior
- Newline-delimited JSON on stdin/stdout (MCP stdio transport).
- Each line is POSTed to the app's /mcp endpoint verbatim.
- The first response's Mcp-Session-Id header is cached and sent with every subsequent request, so session state survives across messages.
- Notifications (requests without id) expect HTTP 202 and produce no stdout.
- Transport errors are turned into JSON-RPC error responses (code -32000 "Cannot reach PDF Content Search") so the client surfaces a meaningful message instead of hanging.
- HTTP 401 → JSON-RPC error -32001 with a hint to rerun the setup.