pdf-toolbox-mcp

中文文档 | Local-first PDF processing for AI agents.

Built for people already using Claude Desktop, Claude Code, Cursor, or another MCP client who want local PDF OCR, unlock, split/merge, render, and compress without uploading files.

Others help AI read PDFs. This one helps AI process them — OCR a scan into a truly searchable file, unlock encrypted PDFs, split/merge/rotate, re-encrypt for sharing. 100% on your machine: no cloud calls, no file uploads, no per-page fees.

Quick start

Add to any MCP client:

{
  "mcpServers": {
    "pdf-toolbox": {
      "command": "uvx",
      "args": ["--from", "pdf-toolbox-mcp", "pdftoolbox"]
    }
  }
}

PyPI project page: pdf-toolbox-mcp

Need a paste-ready setup for a specific client?

• List clients: uv run pdftoolbox client list
• Claude Desktop: uv run pdftoolbox client show claude-desktop
• Cursor: uv run pdftoolbox client show cursor
• Universal project .mcp.json: uv run pdftoolbox client show universal
• Export all client files: uv run pdftoolbox client export
• Detect the current client surface: uv run pdftoolbox client detect
• Semi-auto install: uv run pdftoolbox client install or uv run pdftoolbox client install --scope auto
• Import an existing Claude Desktop setup into Claude Code: uv run pdftoolbox client import-claude-desktop
• Add --all only if you want every supported client bundle

Need a one-shot diagnosis and dependency snapshot before your first task? Run uv run pdftoolbox doctor or uv run pdftoolbox doctor --json. It prints available_now, starter_action, starter_cli, and starter_tool so you can jump straight to the first supported move.

First task:

• OCR a scan: uv run pdftoolbox ocr scan.pdf --lang chi_sim+eng
• Unlock a file: uv run pdftoolbox unlock locked.pdf --password 'xxx'

MCP first task:

1. Ask tool_doctor
2. Then call tool_ocr_pdf

Python dependencies resolve automatically. System tools are capability-leveled — missing ones never crash the server; the tool returns a structured error with the exact install command:

Need the full stack in one shot?

• macOS: brew install qpdf poppler tesseract tesseract-lang ghostscript
• Debian/Ubuntu: sudo apt install qpdf poppler-utils tesseract-ocr tesseract-ocr-chi-sim ghostscript
• Windows: use the per-package commands in the table below

Windows note: Ghostscript's binary is gswin64c.exe there — the probe detects it automatically, so compress_pdf works out of the box. Tesseract language packs (e.g. chi_sim) must be downloaded to its tessdata folder separately.

Upstream / reference:

• qpdf: website · repo
• poppler: website
• tesseract: repo
• ghostscript: website · releases

Every successful response carries a _deps summary ({"level": 2, "missing": ["gs"]}) so the agent always knows what's available.

In an MCP session, use tool_doctor.

Why another PDF MCP?

The PDF MCP space is crowded — but only on the reading side. Based on a hands-on survey of the ecosystem (2026-09):

Pain points this addresses directly:

• Claude natively refuses encrypted PDFs; ChatGPT reports "No text could be extracted" on scans — here, OCR writes a real text layer back into the file, and unlock_pdf decrypts with just the user password.
• Claude Code burns ~30× more tokens reading a PDF page-as-image than extracting text locally.

Tools (25)

Error contract (agents self-route): failures return {"ok": false, "error": "<code>"} — missing_dependency (with install per platform), encrypted_pdf (hint: call unlock_pdf first), wrong_password, output_exists (explicit overwrite required), invalid_page_range, …

Examples

In an MCP client, just describe the outcome — the agent chains the tools itself, and the error contract makes it self-routing (an encrypted_pdf error tells it to call unlock_pdf first, and so on). For headless use, define once:

PTX="uvx --from pdf-toolbox-mcp pdftoolbox"
# PyPI form: uvx --from pdf-toolbox-mcp pdftoolbox

1 · Scan → searchable PDF (the flagship)

“contract-scan.pdf is a scanned contract I can't search. Make it searchable — mostly Chinese with some English.”

Agent: pdf_info → is_searchable reports low text density → ocr_pdf(path, lang="chi_sim+eng") writes contract-scan_ocr.pdf. Text extraction and Ctrl+F now work on the output.

$PTX ocr contract-scan.pdf --lang chi_sim+eng
$PTX text contract-scan_ocr.pdf --pages 1-3

2 · Encrypted PDF → readable

“locked.pdf is password-protected; the password is hunter2. Unlock it and summarize page 3.”

Agent: unlock_pdf(path, password="hunter2") → locked_unlocked.pdf → extract_text(pages="3").

$PTX unlock locked.pdf --password 'hunter2'
$PTX text locked_unlocked.pdf --pages 3

3 · Redact secrets before sharing

“Black out every occurrence of 张三 and HT-2026-088 in draft.pdf — it must be physically unrecoverable.”

Agent: redact_text(queries=["张三", "HT-2026-088"]) → draft_redacted.pdf. Pages containing hits are rasterized, so the strings vanish from the pixels and the text layer; other pages keep their selectable text. Verify by running extract_text on the output: zero hits expected.

$PTX redact-text draft.pdf --query 张三 --query HT-2026-088

More recipes — merge & protect, compress-to-target, batch OCR, the publish-hygiene chain (sanitize → edit_metadata → linearize), vision rendering, locate-and-redact, form filling, damaged-file rescue — in the cookbook.

Configuration

CLI

Everything is also available headless (great for scripts and CI):

uvx --from pdf-toolbox-mcp pdftoolbox ocr scan.pdf --lang chi_sim+eng
uvx --from pdf-toolbox-mcp pdftoolbox unlock locked.pdf --password 'xxx'
uvx --from pdf-toolbox-mcp pdftoolbox split big.pdf --every-n 10
uvx --from pdf-toolbox-mcp pdftoolbox probe all

(Use uvx --from pdf-toolbox-mcp … when installing from PyPI.)

Security & privacy

• No network calls. Files never leave the machine.
• All subprocess calls use argument lists (no shell interpolation); page-range parsing is shared and validated.
• Outputs never silently overwrite: overwrite=true must be passed explicitly.
• Passwords are never logged in error payloads.
• Untrusted PDF content is flagged in tool descriptions (prompt-injection awareness).

License compliance

MIT. System tools are invoked as independent processes (aggregation): poppler (GPL-2.0), qpdf (Apache-2.0), tesseract (Apache-2.0), ghostscript (AGPL, optional); Python deps ocrmypdf/pikepdf are MPL-2.0. See PLAN.md §7 for the full table.

Development

uv sync --dev          # install
uv run pytest -m "not realworld"  # fast path
uv run pytest -m realworld        # noisy / slower regression pack
uv run pytest                     # full suite
uv run pdftoolbox probe all
uv run pdftoolbox probe all --json   # structured dependency snapshot
uv run pdftoolbox doctor
uv run python tools/onboarding_check.py
uv run python tools/onboarding_check.py --json

Cross-platform check without leaving macOS:

docker run --rm -v "$PWD":/src:ro python:3.12-slim bash -c \
  'apt-get update -qq >/dev/null && apt-get install -y -qq poppler-utils tesseract-ocr qpdf ghostscript >/dev/null &&
   pip install -q uv && cp -r /src /work && cd /work && uv sync --dev --quiet && uv run pytest -q'

Roadmap: v0.1.0 ships all 25 tools above. Next up: hardening against real-world scanned documents. Explicit non-goals: editing existing text, password cracking — see PLAN.md.

License

MIT