You have hundreds of PDFs on your Mac. Contracts, invoices, manuals, reports. You know the information is in there somewhere — you just can't remember where. So you open one PDF after another, scroll, search manually. Or you give up and ask a colleague.
What if you could simply ask: "Which contract contains the cancellation terms for the server room?" — and get an answer within seconds, with a source reference?
That's now possible. Starting with version 1.16, PDF Content Search can connect directly to AI assistants like Claude — via the Model Context Protocol (MCP). All processing stays on your Mac.
What is MCP?
The Model Context Protocol (MCP) is an open standard developed by Anthropic. The idea: AI assistants like Claude should be able to access not just their training data, but also your local data and applications — but only when you allow it.
MCP works like a power outlet: the app (in this case PDF Content Search) exposes functions as "tools." The AI assistant can use these tools to retrieve information or perform actions. Communication runs over standardized JSON-RPC — a proven protocol used across many domains.
The key point: No upload to the cloud. The AI assistant accesses the app via localhost. Your documents never leave your machine.
What does this mean in practice?
Without MCP, you'd have to manually copy text from PDFs and paste it into the chat. With MCP, the assistant can independently search your documents, read text, and summarize results. A few examples:
- Answering contract questions: "What's the notice period in the lease agreement?" — Claude searches your PDFs and quotes the relevant passage.
- Analyzing invoices: "List all invoices over $500 from last quarter." — Searches across all indexed documents with amount filtering.
- Sifting through technical docs: "How is API authentication described in the project handbook?" — Full-text search across hundreds of pages in seconds.
- Comparing documents: "Compare the warranty terms in these two contracts." — Claude reads both documents and summarizes the differences.
- Quick overviews: "Summarize the key points from the March 15 meeting notes." — Without opening the PDF yourself.
You ask questions in natural language. Claude uses the PDF Content Search API in the background to find the right documents, read the OCR text, and give you a well-grounded answer.
Setup in 4 steps
The setup takes less than two minutes. You need PDF Content Search version 1.16 or later and an MCP-capable client (e.g. Claude Desktop or Claude Code).
Step 1: Enable the API server
Open PDF Content Search and go to Settings. In the API Server section, enable the local server. It runs on port 44477 and is only reachable via localhost — no external access possible.
Open in App (opens the settings in PDF Content Search)
Step 2: Set an API token
Create a secure token. The token protects access to the MCP server. Without a valid token, all requests are rejected.
Step 3: Copy the MCP configuration
In the MCP (Model Context Protocol) section of the settings, you'll see the status and endpoint URL. Click "Copy Configuration" — the ready-made JSON configuration is copied to your clipboard:
{
"mcpServers": {
"pdf-content-search": {
"type": "http",
"url": "http://127.0.0.1:44477/mcp",
"headers": {
"Authorization": "Bearer YOUR_TOKEN"
}
}
}
}Step 4: Paste into Claude
For Claude Code: Create a .mcp.json file in the root directory of your project and paste the configuration. Claude Code will automatically detect the server on the next launch.
For Claude Desktop: Open the settings (Claude > Settings > Developer > Edit Config) and add the pdf-content-search block under mcpServers.
After restarting Claude, you'll see the PDF Content Search tools in the tool list. From now on, Claude can access your local document collection.
What PDF Content Search offers as an MCP server
PDF Content Search exposes over 20 tools that Claude and other MCP clients can use. Here are the most important ones:
Search and documents
| Tool | Description |
|---|---|
search_documents |
Full-text search across all indexed documents with filters and sorting |
get_document_text |
Read a document's OCR text — page by page or in full |
get_document_metadata |
File name, path, page count, file size, processing status |
export_documents |
Export multiple documents at once (text + metadata) |
Folders and processing
| Tool | Description |
|---|---|
list_folders |
List all monitored folders |
list_folder_contents |
List documents in a folder (with pagination) |
get_processing_status |
OCR statistics: how many documents processed, pending, failed |
reprocess_document |
Run a single document through the OCR pipeline again |
App control
Particularly useful: Claude can also control the app directly. This is handy when you use the assistant as a copilot and want it to show you specific documents:
ui_search— Trigger a search in the appui_navigate— Navigate the document listui_select_document— Select a specific documentui_screenshot— Take a screenshot of the app windowopen_in_finder— Reveal the document in Finderopen_document_external— Open the document in the default app (e.g. Preview)
Privacy: What stays local, what doesn't?
Transparency matters — so here's a clear breakdown of what happens where:
What stays on your Mac
- Your PDF files: The original documents are never uploaded. They stay on your hard drive.
- OCR processing: Text recognition runs entirely locally via Apple Vision. No cloud OCR.
- Full-text index: The search index lives in a local SQLite database.
- MCP communication: The server listens exclusively on
127.0.0.1— unreachable from any other machine on the network, token-protected, sessions expire after one hour of inactivity.
What leaves your machine
When you ask Claude a question, here's what happens: Claude sends a tool request to the local MCP server. PDF Content Search queries its local index and returns text excerpts. These text excerpts are then sent by the Claude client to Anthropic's servers so that Claude can formulate an answer.
In other words: the PDF files themselves are never uploaded, but the retrieved text passages are — just like any other message you type into Claude. This is no different from manually copying and pasting into the chat, except that MCP handles the searching and copying for you.
If you work with sensitive documents, you should be aware of Anthropic's privacy policy. On the business tier (Claude Team/Enterprise), inputs are not used for training.
Apple and MCP
The support from major tech companies shows that MCP is more than a passing trend. Beyond Anthropic (the developer), Google, Microsoft, and OpenAI have all adopted the standard.
Particularly relevant for macOS users: Apple has announced native MCP integration in macOS Tahoe (version 26.1). This means even more apps and system services will be able to connect via MCP in the future. PDF Content Search is already prepared for this.
FAQ
Do I need programming skills?
No. The entire setup is done through the PDF Content Search graphical interface. You only need to paste the copied configuration into Claude — that's it.
Are my documents sent to a server?
The PDF files themselves are never uploaded. OCR processing and search run locally. However, when Claude formulates an answer, the retrieved text excerpts are sent to Anthropic's servers — just like any input you enter in Claude.
Which AI assistants work with MCP?
MCP is an open standard. Currently, Claude Desktop and Claude Code are the most widely used MCP clients. More will follow — especially with native macOS support from Apple.
Which file formats are supported?
PDF Content Search indexes PDFs, images (PNG, JPG, TIFF, HEIC), and various text formats. OCR recognition via Apple Vision supports numerous languages and can also recognize handwritten text.
Can I control which folders Claude has access to?
Claude only sees the folders you've added as monitored folders in PDF Content Search. You have full control over which documents are indexed.
Is it free?
The MCP integration is part of PDF Content Search and included in every license — both App Store and Web Store versions. For Claude Desktop or Claude Code, you need an account with Anthropic.