pdf2mcp

██████╗ ██████╗ ███████╗██████╗ ███╗   ███╗ ██████╗██████╗
██╔══██╗██╔══██╗██╔════╝╚════██╗████╗ ████║██╔════╝██╔══██╗
██████╔╝██║  ██║█████╗   █████╔╝██╔████╔██║██║     ██████╔╝
██╔═══╝ ██║  ██║██╔══╝  ██╔═══╝ ██║╚██╔╝██║██║     ██╔═══╝
██║     ██████╔╝██║     ███████╗██║ ╚═╝ ██║╚██████╗██║
╚═╝     ╚═════╝ ╚═╝     ╚══════╝╚═╝     ╚═╝ ╚═════╝╚═╝

Turn any PDF folder into a searchable MCP server with semantic, hybrid, or keyword search.

Installation

From PyPI (recommended)

pip install pdf2mcp

Or with uv:

uv tool install pdf2mcp

From source

git clone https://github.com/iSamBa/pdf2mcp.git
uv tool install ./pdf2mcp

To update after pulling new changes:

uv tool install --force ./pdf2mcp

Optional: Tesseract OCR

Tesseract is only needed if you want to extract text from scanned or image-only PDFs. Without it, pdf2mcp works fine for text-based PDFs — image-only pages are simply skipped with a warning.

macOS:

brew install tesseract

Ubuntu / Debian:

sudo apt-get install tesseract-ocr

Windows:

Download the installer from UB-Mannheim/tesseract.

Additional languages: install language packs for non-English PDFs:

# Example: French and German
sudo apt-get install tesseract-ocr-fra tesseract-ocr-deu
# or on macOS
brew install tesseract-lang

Then set PDF2MCP_OCR_LANGUAGE to the appropriate language code (e.g., fra, deu).

Verify

pdf2mcp --version

Related MCP server: MCP Server Knowledge Engine

Quick Start

Interactive Setup (recommended)

pdf2mcp init -i ./my-project

The interactive wizard walks you through all configuration in 6 steps:

1. Project directory — confirm or change the target path

2. OpenAI API key — securely enter your key (masked input) and optional base URL

3. Documents directory — where your PDFs live (default: docs)

4. Embedding settings — choose model, chunk size, and overlap

5. Server settings — name, transport, host, and port

6. OCR settings — enable/disable OCR for scanned PDFs

After setup, the wizard optionally offers to ingest any PDFs found in your docs directory and generate ready-to-paste MCP client config snippets.

Manual Setup

# 1. Scaffold a project (creates docs/ and .env template)
pdf2mcp init ./my-project
cd my-project

# 2. Add your PDFs to docs/ and set OPENAI_API_KEY in .env

# 3. Ingest
pdf2mcp ingest

# 4. Start the server
pdf2mcp serve

# 5. Get config snippets for your MCP client
pdf2mcp config

Architecture

pdf2mcp separates server and client concerns:

• Server (pdf2mcp serve) — runs independently, handles PDF ingestion, embedding, and search. Configured via PDF2MCP_* environment variables.

• Client (Claude Code, Cursor, VS Code, etc.) — connects to a running server over HTTP. Only needs the server URL.

The default transport is streamable-http. The server listens on http://127.0.0.1:8000/mcp and shuts down gracefully on SIGINT/SIGTERM.

OCR / Scanned PDF Support

pdf2mcp automatically detects image-only pages in PDFs and falls back to Tesseract OCR when available:

• Per-page strategy: text pages are extracted via pymupdf4llm; image-only pages are OCR'd via Tesseract.

• Automatic detection: each page is checked for extractable text (via _page_has_text) and image dominance (via _is_image_dominant). Pages without sufficient text are classified as image-only.

• Graceful degradation: if Tesseract is not installed or OCR is disabled, image-only pages are skipped with a warning — text-based pages are still extracted normally.

• Configuration: use PDF2MCP_OCR_ENABLED, PDF2MCP_OCR_LANGUAGE, and PDF2MCP_OCR_DPI environment variables (see Environment Variables).

Commands

Common Flags

# Override docs directory
pdf2mcp ingest --docs-dir ./my-pdfs
pdf2mcp serve --docs-dir ./my-pdfs

# Force re-ingestion (clears DB and re-ingests all documents)
pdf2mcp ingest --force

# Enable debug logging
pdf2mcp ingest -v
pdf2mcp serve --verbose

# Use stdio transport (for clients that spawn the server)
pdf2mcp serve --transport stdio

# Custom host/port
pdf2mcp serve --host 0.0.0.0 --port 9000

# Custom server name
pdf2mcp serve --name my-docs

# Config for a specific client
pdf2mcp config --client cursor
pdf2mcp config --client claude-desktop --transport stdio

# Interactive setup wizard
pdf2mcp init -i ./my-project
pdf2mcp init --interactive

# View index statistics
pdf2mcp stats

# Search the index from CLI
pdf2mcp search "safety requirements"
pdf2mcp search "torque settings" --filename manual.pdf
pdf2mcp search "installation" -n 10

# Delete a document from the index
pdf2mcp delete old-manual.pdf
pdf2mcp delete old-manual.pdf -y   # skip confirmation

Client Configuration

pdf2mcp config generates ready-to-paste JSON for all supported clients. The default is HTTP — clients just need the server URL:

{
  "mcpServers": {
    "pdf-docs": {
      "type": "http",
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}

Use --transport stdio for clients that need to spawn the server process (e.g., Claude Desktop):

{
  "mcpServers": {
    "pdf-docs": {
      "command": "uv",
      "args": ["run", "pdf2mcp", "serve"]
    }
  }
}

Environment Variables

Server settings (PDF2MCP_*)

These configure the server process. MCP clients never need these.

Search Modes

pdf2mcp supports three search modes, controlled by the PDF2MCP_SEARCH_MODE environment variable:

To switch modes, set PDF2MCP_SEARCH_MODE in your .env and re-ingest:

# In .env
PDF2MCP_SEARCH_MODE=hybrid

# Re-ingest to build the FTS index
pdf2mcp ingest --force

Hybrid and keyword modes automatically create a full-text search index. If you switch modes without re-ingesting, the FTS index is created lazily on the first query.

MCP Tools

The server exposes six tools:

Typical workflow

1. list_docs — discover available documents

2. get_sections — browse a document's structure

3. read_section or read_page — read specific content

4. search_docs or search_in_doc — find information by query

MCP Prompts

The server provides five prompts that guide LLMs through multi-tool workflows:

Prompts return step-by-step instructions that reference the existing tools, enabling LLMs to perform complex multi-step document analysis automatically.

MCP Resources

Development

git clone https://github.com/iSamBa/pdf2mcp.git
cd pdf2mcp
uv sync --all-extras
uv run pytest
uv run ruff check src/
uv run mypy src/

License

MIT