pdf-search-mcp

MCP server for full-text search across PDF document collections. Built for AI agents — index once, search instantly from any MCP client.

• Search entire collections — pre-indexes all PDFs for instant ranked results with snippets, not one file at a time
• Fully offline — no API keys, no cloud services, just SQLite FTS5 and PyMuPDF
• Page rendering — render pages as PNG for formulas, diagrams, and tables; crop to a region with auto-DPI scaling for detail shots
• Dual renderer — CoreGraphics on macOS (sharper math fonts), PyMuPDF on Linux/Windows
• German-aware — automatic expansion of ß↔ss, ä↔ae, ö↔oe, ü↔ue so both spellings match
• stdio or HTTP — per-client subprocess by default, or a standalone shared server for trusted networks (Dockerfile included)

Installation

From PyPI

pip install pdf-search-mcp

From source

git clone https://github.com/renvk/pdf-search-mcp.git
cd pdf-search-mcp
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Requires Python 3.10+. On macOS, pyobjc-framework-Quartz is installed automatically for native CoreGraphics PDF rendering (sharper formula and math font output). On Linux/Windows, PyMuPDF is used as the renderer.

Quick Start

1. Index your PDFs

PDF_SEARCH_DIR=/path/to/your/pdfs python -m pdf_search_mcp.pdf_search index

2. Register with your MCP client

The server runs over stdio. Example for Claude Code:

# project-scoped (only available in the current directory)
claude mcp add pdf-search -- pdf-search-mcp

# or global (available in all projects)
claude mcp add --scope global pdf-search -- pdf-search-mcp

For other MCP clients, add to your MCP config:

{
  "mcpServers": {
    "pdf-search": {
      "command": "pdf-search-mcp"
    }
  }
}

3. Search

Ask your AI agent to search your PDFs — it will use the search, read_page, and read_page_image tools automatically.

Configuration

Self-Hosted Server (HTTP)

The server has two transports:

• stdio (default, no flags): each MCP client launches its own server subprocess on the same machine. Use this for single-machine setups.
• http: one standalone server shares one indexed PDF collection with multiple clients over a trusted network:

pdf-search-mcp --transport http --host 0.0.0.0 --port 8000

The MCP endpoint is http://<server>:8000/mcp. Connecting from Claude Code:

claude mcp add --transport http --scope user pdf-search http://<server>:8000/mcp

Security: the HTTP transport has no authentication or TLS. Run it only on trusted networks (LAN, VPN) and never expose it to the internet. The default --host 127.0.0.1 keeps it local to the machine; binding 0.0.0.0 is an explicit opt-in.

Note: over HTTP, read_page_image returns the rendered PNG as inline MCP image content, so page rendering works for clients on other machines (including clients without filesystem access, such as Claude Desktop). Over stdio it returns a file path for the client to open, as before.

Docker

The repository includes a Dockerfile for container hosts (home servers, NAS devices). On startup the container runs an incremental index sync against the mounted PDF directory (only new, changed, and deleted files are processed) and then serves on port 8000:

git clone https://github.com/renvk/pdf-search-mcp.git
cd pdf-search-mcp
docker build -t pdf-search-mcp .
docker run -d --name pdf-search \
  -p 8000:8000 \
  -v /path/to/pdfs:/pdfs:ro \
  -v pdf-index:/data \
  pdf-search-mcp

Or with Docker Compose:

services:
  pdf-search:
    build: .
    ports:
      - "8000:8000"
    volumes:
      - /path/to/pdfs:/pdfs:ro
      - pdf-index:/data
    restart: unless-stopped

volumes:
  pdf-index:

New PDFs in the mounted directory are picked up on container restart, or immediately with:

docker exec pdf-search python -m pdf_search_mcp.pdf_search index

Set PDF_SEARCH_INDEX_ON_START=0 in the container environment to skip the index sync (e.g. when the index is maintained by an external job).

CLI Usage

The pdf_search.py module doubles as a CLI for indexing and direct search:

# Build index (first time — PDF_SEARCH_DIR required)
PDF_SEARCH_DIR=/path/to/pdfs python -m pdf_search_mcp.pdf_search index

# Subsequent syncs (path remembered from first index)
python -m pdf_search_mcp.pdf_search index

# Search from command line
python -m pdf_search_mcp.pdf_search search "query terms"

# Read a specific page
python -m pdf_search_mcp.pdf_search read filename.pdf 5

# Show index statistics
python -m pdf_search_mcp.pdf_search stats

# Report extraction-quality problems (scanned files needing OCR,
# stale normalization, broken font mappings, near-empty pages)
python -m pdf_search_mcp.pdf_search quality

# Rebuild index from scratch (path remembered)
python -m pdf_search_mcp.pdf_search reindex

Search Syntax

Uses SQLite FTS5 query syntax:

Auto-quoting: Terms containing any special character (dots, hyphens, commas, slashes, colons, ...) are automatically quoted (e.g., ISO-27001 becomes "ISO-27001", 1:100 becomes "1:100") because FTS5 treats these as token separators or operators. Query preparation guarantees valid FTS5 syntax — stray quotes are dropped, unbalanced parentheses are repaired, and dangling AND/OR operators are trimmed. The one exception is NOT without a left operand (FTS5's NOT is binary): it is passed through and returns a clear error, because silently searching the excluded term would invert the query's meaning.

German expansion: Umlauts and eszett are automatically expanded to their digraph equivalents and vice versa (ß↔ss, ä↔ae, ö↔oe, ü↔ue). Searching for Größe also finds Groesse, and Weißbuch also finds Weissbuch. Reverse expansion (ss→ß) replaces one position at a time. Expansion also applies inside NEAR() expressions.

Auto-relaxation: When a multi-term query returns no results (all terms must appear on the same page), the search automatically relaxes: first by dropping the term least represented in the corpus (chosen by uncapped match counts), then by OR-ing all terms. A note in the output explains what was actually searched. Structured queries (explicit AND, OR, NOT, NEAR, parentheses) are not relaxed.

MCP Tools

When the same filename exists in several subfolders, read_page and read_page_image require the subfolder parameter ("" selects the root folder); an unspecified subfolder returns an error listing the candidates instead of picking one arbitrarily.

Python API

from pdf_search_mcp import (
    search_with_relaxation, search_pdfs, prepare_query,
    read_pdf_page, render_pdf_page, index_pdfs,
)

# Index PDFs
index_pdfs("/path/to/pdfs")

# Search with the full pipeline (auto-quoting, German expansion,
# relaxation) — same behavior as the MCP search tool and the CLI
results, note = search_with_relaxation("ISO-27001 Anhang", limit=5)
for r in results:
    print(f"{r['subfolder']}/{r['file']} p.{r['page']}: {r['snippet']}")

# Low-level: search_pdfs takes a RAW FTS5 MATCH string (no preparation).
# Run user input through prepare_query first.
results = search_pdfs(prepare_query("garbage collection"), limit=5)

# Read full page text
text = read_pdf_page("document.pdf", 42)

# Render full page as PNG
png_path = render_pdf_page("document.pdf", 42)

# Render cropped region (DPI auto-scales to maximize detail)
png_path = render_pdf_page("document.pdf", 42, region=[0.0, 0.5, 1.0, 0.8])

How It Works

1. Indexing incrementally syncs your PDF directory into a SQLite FTS5 virtual table — on first run all PDFs are indexed; afterwards only new, changed (by mtime/size), and deleted files are processed, each committed individually so an interrupted run resumes where it stopped. Only page content is searchable; filenames, subfolders, and page numbers are stored as unindexed metadata so query terms cannot match them. Directories starting with _ are skipped. Extracted text is normalized before indexing — typographic ligatures are decomposed (efficiency → efficiency), words split by line-break hyphenation are rejoined, and invisible soft hyphens are removed — so words that would otherwise never match a query do. read_page applies the same normalization, so the text you read is exactly the text that was searched.

Upgrading: a pre-0.3 index (searchable metadata columns) is refused with a clear error; an index built before text normalization keeps working but stores unnormalized text. In either case, run python -m pdf_search_mcp.pdf_search reindex once to rebuild.

2. Searching runs FTS5 MATCH queries and re-ranks results by combining BM25 relevance with match density — pages where search terms cluster together score higher than pages with the same terms scattered throughout. The density signal blends term concentration (matches per character) and spatial clustering (how tightly grouped the matches are).

3. Reading re-opens the original PDF file on disk (path resolved via the stored pdf_dir metadata) for full page text or image rendering. Region crops auto-scale DPI to fill a 1568 px long-edge budget, maximizing detail without producing oversized images.

The database stores the text content only — original PDFs are accessed on disk for read_page and read_page_image. Rendering uses CoreGraphics on macOS and PyMuPDF elsewhere.

License

MIT