Skip to main content
Glama
pdf-toolkit-mcp

pdf-toolkit-mcp

by EigenCharlie

pdf-toolkit-mcp

18 PDF operations for Claude Code, Claude Desktop, and any MCP client — via iLoveAPI.

npm version npm downloads License: MIT Node.js TypeScript MCP Claude Code CI

Install · Tools · Recipes · Architecture · Security · FAQ


Disclaimer. pdf-toolkit-mcp is an independent, community-built open-source client. It is not affiliated with, endorsed, sponsored, or certified by iLovePDF SL. The project talks to the public iLoveAPI REST service using your project credentials — nothing is proxied, resold, or relicensed. "iLovePDF" and "iLoveAPI" are trademarks of iLovePDF SL; this project does not claim ownership of those marks.


📖 Table of contents


Related MCP server: gurupdf-mcp

🎯 Why pdf-toolkit-mcp

Manual iLoveAPI

Raw @ilovepdf/ilovepdf-nodejs

pdf-toolkit-mcp

Callable from Claude Code / Claude Desktop / any MCP client

❌

❌

✅

JWT signing + 5-step lifecycle handled for you

❌

✅

✅

Validated Zod schemas (no bad inputs reach the API)

❌

❌

✅

Path sandboxing against traversal

❌

❌

✅

Structured error codes (RATE_LIMITED, PLAN_LIMIT, …)

❌

partial

✅

Progress notifications during long OCR tasks

❌

❌

✅

Bundled Claude Code skill with multi-step pipelines

❌

❌

✅

Zero-setup install via npx -y

❌

❌

✅

Free for personal use. iLoveAPI's free tier grants ~2,500 credits / month; this MCP itself is MIT-licensed and costs nothing.


🚀 Quick tour

After installation, just talk to Claude. The model figures out which tools to call:

You: Merge invoice-jan.pdf and invoice-feb.pdf into Q1.pdf, then compress it hard
     and lock it with the password "2026q1".

Claude (planning):
  1. merge_pdf     → invoice-jan.pdf + invoice-feb.pdf  →  Q1.pdf
  2. compress_pdf  → Q1.pdf                             →  Q1-compressed.pdf  (level: extreme)
  3. protect_pdf   → Q1-compressed.pdf                  →  Q1-compressed-protected.pdf

Claude (result):
  ✅ Created ~/docs/Q1-compressed-protected.pdf (312 KB, password-protected).

📦 Install

Prerequisites

  • Node.js ≥ 18 (LTS recommended).

  • iLoveAPI project keys — free tier at developer.ilovepdf.com. Takes ~2 minutes:

    1. Sign up → create a project → copy Project public key and Project secret key.

    2. Export them (or drop them into your client's MCP config — examples below).

Option A — Claude Code (CLI one-liner)

export ILOVEAPI_PROJECT_PUBLIC_KEY="project_public_xxx"
export ILOVEAPI_PROJECT_SECRET_KEY="secret_key_xxx"

claude mcp add pdf-toolkit -- npx -y pdf-toolkit-mcp

Restart Claude Code and all 18 tools appear in the picker. Verify with /mcp → you should see pdf-toolkit: connected (18 tools).

Option B — Claude Code Plugin (includes the pdf-workflow skill)

/plugin marketplace add EigenCharlie/pdf-toolkit-mcp
/plugin install pdf-toolkit@EigenCharlie/pdf-toolkit-mcp

The plugin auto-configures the MCP server and installs a skill that teaches Claude five canonical multi-step PDF pipelines (see Workflow recipes).

Option C — Claude Desktop (claude_desktop_config.json)

Edit claude_desktop_config.json (%APPDATA%\Claude\ on Windows, ~/Library/Application Support/Claude/ on macOS):

{
  "mcpServers": {
    "pdf-toolkit": {
      "command": "npx",
      "args": ["-y", "pdf-toolkit-mcp"],
      "env": {
        "ILOVEAPI_PROJECT_PUBLIC_KEY": "project_public_xxx",
        "ILOVEAPI_PROJECT_SECRET_KEY": "secret_key_xxx"
      }
    }
  }
}

Restart Claude Desktop — the 🔌 icon should show pdf-toolkit connected.

Option D — any MCP client via stdio

ILOVEAPI_PROJECT_PUBLIC_KEY=… ILOVEAPI_PROJECT_SECRET_KEY=… npx -y pdf-toolkit-mcp

The server speaks the standard MCP JSON-RPC 2.0 framing over stdio. Wire it into Cursor, Windsurf, mcphub, mcp-inspector, or anything else that speaks MCP.

Option E — Claude Desktop .mcpb bundle (drag-and-drop)

Grab pdf-toolkit-mcp-<version>.mcpb from the latest GitHub Release and drag it into Claude Desktop → Settings → Extensions. Claude will prompt for your iLoveAPI keys via the user_config section of the bundled manifest. No npx, no Node, no terminal.

Bundles ship unsigned (no publisher certificate yet). macOS Gatekeeper / Windows SmartScreen may warn on first install; verify the SHA-256 in the release notes before accepting.

Option F — MCP Registry (auto-discovered by compatible clients)

The server is indexed on the official MCP Registry under io.github.EigenCharlie/pdf-toolkit-mcp. Clients that browse the registry (Claude Desktop extension pane, VS Code MCP picker, mcphub, etc.) can install it without any manual config — they'll point at the npm package and prompt for the two iLoveAPI env vars automatically.


🧰 Tool catalog

All 18 tools are exposed with strict Zod schemas, additionalProperties: false, and return both a text summary and a file:// resource URI so the client can surface the output.

📚 Organize

Tool

What it does

Key inputs

merge_pdf

Combine ≥ 2 PDFs in order

input_files[], output_path?

split_pdf

Split by page ranges ("1-3,5-7") or fixed chunk size

input_file, ranges?, fixed_range?

extract_pdf_pages

Keep a specific subset of pages

input_file, pages ("1,3,5-9")

🔄 Convert — from PDF

Tool

Output

Notes

pdf_to_word

.docx

Scanned PDFs → run ocr_pdf first for best results

pdf_to_excel

.xlsx

Works best on tabular source PDFs

pdf_to_powerpoint

.pptx

One slide per PDF page

pdf_to_jpg

.zip of .jpg

mode: "pages" renders pages; "extract" pulls embedded images

🔄 Convert — to PDF

Tool

Accepts

Notes

office_to_pdf

.doc, .docx, .xls, .xlsx, .ppt, .pptx

Server-side rendering

html_to_pdf

.html, .htm

Local HTML only; external assets may not resolve

image_to_pdf

.jpg, .jpeg, .png

One image per page, preserves order

✏️ Edit

Tool

What it does

Key inputs

rotate_pdf

Rotate pages 90 / 180 / 270° clockwise

rotation, pages? (default: all)

add_page_numbers

Stamp numbered footer/header

starting_number, vertical_position, horizontal_position

add_watermark

Text or image watermark

mode: "text" + text, OR mode: "image" + image_file

🔐 Security

Tool

What it does

Key inputs

unlock_pdf

Remove known password

input_file, password

protect_pdf

Add password

input_file, password

🩹 Repair / OCR

Tool

What it does

Notes

repair_pdf

Attempt structural repair on damaged PDFs

Useful before further processing

ocr_pdf

Run OCR to make scans searchable

languages[] (e.g. ["eng"], ["spa"], ["eng","spa"]). ⏱ Can exceed 60s on image-heavy PDFs

All tools accept absolute or CWD-relative paths for input_file(s) and an optional output_path (file or directory). Defaults place the result next to the first input with a timestamped name.


🍳 Workflow recipes

The bundled pdf-workflow skill (skills/pdf-workflow/SKILL.md) teaches Claude five canonical multi-step pipelines. You can also run these manually — just describe the end state and Claude chains the tools for you.

merge_pdf([a.pdf, b.pdf, c.pdf])
   → compress_pdf(level="recommended")
   → protect_pdf(password="…")
office_to_pdf(report.docx)
   → add_watermark(mode="text", text="CONFIDENTIAL", opacity=30)
   → protect_pdf(password="…")
extract_pdf_pages(scan.pdf, pages="3-9")
   → ocr_pdf(languages=["eng"])
   → pdf_to_word
compress_pdf(big.pdf, level="extreme")
   → protect_pdf(password="…")
repair_pdf(scan_broken.pdf)
   → ocr_pdf(languages=["eng"])
   → add_page_numbers(position="bottom-center")

🏛 Architecture

┌────────────────────────────┐    stdio    ┌──────────────────────────┐
│     Claude Code / Desktop  │◀───────────▶│   pdf-toolkit-mcp server │
│   (or any MCP client)      │  JSON-RPC   │   (this repo)            │
└────────────────────────────┘             └────────────┬─────────────┘
                                                        │
                                                        ▼
                                             ┌──────────────────────┐
                                             │  @ilovepdf/          │
                                             │  ilovepdf-nodejs     │
                                             │  (JWT + HTTP)        │
                                             └──────────┬───────────┘
                                                        │
                                                        ▼  HTTPS
                                             ┌──────────────────────┐
                                             │   iLoveAPI servers   │
                                             │   api.ilovepdf.com   │
                                             └──────────────────────┘

iLoveAPI 5-phase lifecycle (orchestrated in src/api/tasks.ts)

  [ 10% ]   start     →  POST /v1/start/{tool}     (server assignment + task id)
  [ 40% ]   upload    →  POST /v1/upload           (one call per file, progress scales)
  [ 50% ]   process   →  POST /v1/process          (run the tool with params)
  [ 90% ]   download  →  GET  /v1/download/{task}  (bytes → Buffer)
  [100% ]   done      →  write to disk, emit file:// resource URI

The server emits MCP notifications/progress at each boundary so your client can render a live progress bar for slow operations (OCR, large merges).

Project layout

pdf-toolkit-mcp/
├── .claude-plugin/
│   ├── plugin.json           # Plugin manifest
│   └── .mcp.json             # MCP server config (npx -y pdf-toolkit-mcp)
├── skills/
│   └── pdf-workflow/
│       └── SKILL.md          # 5-recipe skill for multi-step pipelines
├── src/
│   ├── index.ts              # #!/usr/bin/env node shebang
│   ├── server.ts             # MCP stdio bootstrap
│   ├── api/
│   │   ├── client.ts         # iLoveAPI client singleton (CJS interop via createRequire)
│   │   ├── tasks.ts          # 5-phase lifecycle orchestrator
│   │   ├── errors.ts         # HTTP → structured PdfToolkitError mapping
│   │   └── types.ts          # Types + error class
│   ├── tools/                # 18 tools, one file per concern
│   │   ├── _shared.ts        # Zod fragments + writeOutputAndReport helper
│   │   ├── merge.ts  split.ts  compress.ts
│   │   ├── convertFromPdf.ts convertToPdf.ts
│   │   ├── pageOps.ts  security.ts  watermark.ts
│   │   ├── repair.ts  ocr.ts
│   │   └── index.ts          # allTools[] barrel
│   └── util/
│       ├── paths.ts          # resolveInputs/resolveOutput + sandbox enforcement
│       ├── progress.ts       # Progress adapter
│       └── logger.ts         # stderr-only logger (stdio-safe)
├── tests/
│   ├── unit/                 # client, paths, errors, tools.merge (25 tests)
│   └── integration/          # smoke.test.ts — gated on iLoveAPI creds
├── scripts/
│   └── inspector.sh          # npm run inspect → MCP Inspector UI
└── .github/workflows/
    ├── ci.yml                # ubuntu+windows × node 18/20/22
    └── publish.yml           # Publishes to npm on v* tags with --provenance

⚙️ Configuration reference

Environment variables

Variable

Required

Default

Description

ILOVEAPI_PROJECT_PUBLIC_KEY

✅

—

Project public key from developer.ilovepdf.com

ILOVEAPI_PROJECT_SECRET_KEY

✅

—

Project secret key. Never logged. Used for local JWT signing.

ILOVEAPI_SANDBOX_ROOT

❌

—

Absolute path. When set, all input/output paths must resolve inside this directory — traversal attempts throw PATH_TRAVERSAL.

PDF_TOOLKIT_DEBUG

❌

—

Set to 1 to emit verbose stderr logs (request shape, phase timings). Secrets never logged.

Structured error codes

Every failure surfaces a PdfToolkitError with a stable machine-readable code:

Code

Trigger

What to tell the user

MISSING_CREDENTIALS

Env vars not set

Point them at developer.ilovepdf.com

INVALID_INPUT

HTTP 400 / Zod parse failure

Fix the arguments

AUTH_FAILED

HTTP 401

Regenerate the project keys

PLAN_LIMIT

HTTP 402

Free tier exhausted or tool not in plan

NOT_FOUND

HTTP 404

Input file or task missing

RATE_LIMITED

HTTP 429

Back off and retry

TASK_LIMIT

SDK TaskLimit error

Too many concurrent tasks

PATH_TRAVERSAL

Path escapes ILOVEAPI_SANDBOX_ROOT

Reject the request

API_ERROR

Unmapped iLoveAPI error

Check data.http_status + data.iloveapi_code


🔒 Security model

pdf-toolkit-mcp is designed to be safe to install on a developer machine:

  • Stdio only. The server never opens a network socket; it only makes outbound HTTPS calls to api.ilovepdf.com via the official SDK.

  • No console.log. JSON-RPC over stdio would corrupt on any stray stdout write, so every log line goes through process.stderr.

  • Secrets never persisted. Credentials live in env vars for the lifetime of the process and are never written to disk, log files, or tool responses.

  • JWT generated locally. The @ilovepdf/ilovepdf-nodejs SDK self-signs JWTs with your secret key — no secret leaves the machine.

  • Path sandbox (opt-in). Set ILOVEAPI_SANDBOX_ROOT to constrain the server to a single directory tree; any path resolving outside throws before the API is ever called.

  • Extension whitelists per tool. html_to_pdf rejects .exe; image_to_pdf only accepts common raster formats, etc.

  • MIT licensed, audit-friendly. ~2k lines of TypeScript. No obfuscation, no minification, no postinstall scripts.


👩‍💻 Development

git clone https://github.com/EigenCharlie/pdf-toolkit-mcp.git
cd pdf-toolkit-mcp
npm install

Common tasks

Command

What it does

npm run build

Compile TypeScript → dist/

npm run dev

Watch-mode build

npm run typecheck

tsc --noEmit — fast error surface

npm run lint

ESLint on src/ + tests/

npm test

Vitest unit suite (no network)

npm run test:integration

Real iLoveAPI calls (requires creds)

npm run inspect

Launch MCP Inspector against the local build

Debugging inside Claude Code

  1. Build locally: npm run build.

  2. Register the local dist instead of npm:

    claude mcp remove pdf-toolkit    # if previously registered
    claude mcp add pdf-toolkit -- node "$(pwd)/dist/index.js"
  3. Set PDF_TOOLKIT_DEBUG=1 in your shell before launching Claude Code for verbose stderr.

  4. Stderr is visible with claude mcp logs pdf-toolkit.


✅ Testing & CI

  • Unit tests — 25 tests across client, paths, errors, and tools.merge. Fully mocked, no network. Run with npm test.

  • Integration smoke test — tests/integration/smoke.test.ts spins up two minimal valid PDFs and exercises a real merge_pdf end-to-end. Gated on ILOVEAPI_*_KEY env vars; automatically skipped in PR CI for security.

  • CI matrix — .github/workflows/ci.yml runs lint → typecheck → build → test on:

    • OS: ubuntu-latest, windows-latest

    • Node: 18, 20, 22


🚢 Releasing

Publishing is automated on tag push:

npm version patch           # or minor / major
git push --follow-tags

The publish workflow then:

  1. Re-runs lint + typecheck + build + tests.

  2. Publishes to npm with --access public --provenance (supply-chain attestation).

  3. Creates a GitHub Release with auto-generated notes.

Repo secret required: NPM_TOKEN (npm automation token).


🗺 Roadmap

  • v0.1 — 18 PDF tools, stdio transport, plugin + skill, CI, npm publish

  • v0.1.1 — MCPB bundle (.mcpb) for Claude Desktop, marketplace manifest, MCP Registry submission (mcpName)

  • v0.2 — Submit the Claude plugin to anthropics/claude-plugins-official, publish an image-tools companion pack

  • v0.3 — 8 image tools (resize, convert, compress, crop, rotate, watermark, upscale, remove background)

  • v0.4 — Signature tools (signing flows require paid tier — gated behind env flag)

  • v1.0 — Optional Streamable-HTTP transport for hosted/multi-user setups

Have a feature request? Open an issue.


🙋 FAQ

Is this legal / safe / allowed by iLovePDF? Yes. It's a thin open-source client that uses your credentials — nothing is resold, relicensed, or proxied. The iLoveAPI Terms of Service explicitly permit open-source libraries using user-provided keys (several community SDKs have existed for years). This project does not ship any keys, does not use "iLovePDF"/"iLoveAPI" in its package name, and carries the required disclaimer.

Does it cost money? iLoveAPI offers ~2,500 credits per month free (enough for thousands of small ops). Heavier workloads need a paid iLoveAPI plan. This MCP itself is MIT and free forever.

Does it work offline? No. iLoveAPI is cloud-only, so every tool call requires internet access.

Can I pin a specific version? Yes: claude mcp add pdf-toolkit -- npx -y pdf-toolkit-mcp@0.1.0.

Can I self-host? Yes. Clone the repo, npm run build, and point your client at node ./dist/index.js. The server is entirely stateless.

What happens if iLoveAPI is down? The tool returns a PdfToolkitError with code: "API_ERROR" and data.http_status. Claude can retry with exponential backoff on your behalf.

Can I use this without Claude? Yes — any MCP-compatible client works: Cursor, Windsurf, Zed, MCP Inspector, custom clients. The server is 100% spec-compliant.


🤝 Contributing

Contributions welcome! Small checklist before opening a PR:

  1. npm run lint && npm run typecheck && npm test → all green.

  2. New tools follow the pattern in src/tools/merge.ts (Zod schema → runTask → writeOutputAndReport).

  3. Add at least one unit test.

  4. Update the Tool catalog in this README if you add/rename a tool.

Bug reports are even more welcome — include the tool name, the arguments, and (if safe to share) the stderr output.


📜 License & trademarks

Released under the MIT License © 2026 Carlos.

"iLovePDF" and "iLoveAPI" are trademarks of iLovePDF SL. This project is an independent, community-built client and is not affiliated with, endorsed by, sponsored by, or certified by iLovePDF SL. The package name intentionally avoids both marks.


🙏 Credits

Available Tools

18 tools
add_page_numbersA

Stamp page numbers onto a PDF. Configure starting number, which pages to number, and placement (top/bottom + left/center/right).

Inputs: input_file (required), pages, output_path, starting_number, +2 more.

add_watermarkA

Stamp a text or image watermark onto every page of a PDF. For mode="text", provide text. For mode="image", provide image_file (.jpg/.jpeg/.png). Tune placement, rotation, and opacity as needed.

Inputs: mode (required), input_file (required), text, layer, +6 more.

compress_pdfA

Reduce the file size of a PDF. Choose compression_level "low", "recommended" (default), or "extreme". Outputs a .pdf file on disk and returns its path.

Inputs: input_file (required), output_path, compression_level.

extract_pdf_pagesA

Extract a subset of pages from a PDF into a new PDF. Specify pages as a list or range, e.g. "1,3,5-7".

Inputs: pages (required), input_file (required), output_path.

html_to_pdfA

Convert a local HTML file (.html or .htm) to a PDF. The HTML is rendered server-side; external assets referenced by the file may not be fetched.

Inputs: input_file (required), output_path.

image_to_pdfA

Combine one or more images (.jpg, .jpeg, .png) into a single PDF, one image per page, in the order provided.

Inputs: input_files (required), output_path.

merge_pdfA

Merge two or more PDF files into a single PDF, preserving order. Outputs a .pdf file on disk and returns its path.

Inputs: input_files (required), output_path.

ocr_pdfA

Run OCR on a PDF to make scanned text selectable/searchable. Accepts one or more language codes (e.g. "eng", "spa"). Large or image-heavy PDFs can take more than 60 seconds.

Inputs: input_file (required), languages, output_path.

office_to_pdfA

Convert a Microsoft Office document (Word, Excel, or PowerPoint) to a PDF. Accepts .doc, .docx, .xls, .xlsx, .ppt, and .pptx files.

Inputs: input_file (required), output_path.

pdf_to_excelA

Convert a PDF to a Microsoft Excel .xlsx spreadsheet. Works best when the source PDF contains clear tabular data; non-tabular PDFs may produce messy results.

Inputs: input_file (required), output_path.

pdf_to_jpgA

Convert a PDF to JPG images. In "pages" mode (default) each page becomes a JPG; in "extract" mode only embedded images are pulled out. Returns a .zip archive of the images.

Inputs: input_file (required), mode, output_path.

pdf_to_powerpointA

Convert a PDF to a Microsoft PowerPoint .pptx presentation. Each PDF page becomes an editable slide; conversion quality depends on the source.

Inputs: input_file (required), output_path.

pdf_to_wordA

Convert a PDF to an editable Microsoft Word .docx document. Conversion quality depends on the source (scanned PDFs produce poor results; use ocr_pdf first if needed).

Inputs: input_file (required), output_path.

protect_pdfA

Add password protection to a PDF. Anyone opening the output will need the supplied password.

Inputs: password (required), input_file (required), output_path.

repair_pdfA

Attempt to repair a damaged or malformed PDF. Useful for scanned PDFs that fail to open or render correctly in some readers. Produces a .pdf file on disk.

Inputs: input_file (required), output_path.

rotate_pdfA

Rotate pages in a PDF by 90, 180, or 270 degrees clockwise. Target specific pages with "pages" (e.g. "1,3,5" or "1-5"), or rotate all pages by default.

Inputs: rotation (required), input_file (required), pages, output_path.

split_pdfA

Split a PDF into multiple PDFs by page ranges (e.g. "1-3,5-7") or into fixed-size chunks. Returns a .zip archive containing the resulting PDFs.

Inputs: input_file (required), ranges, fixed_range, output_path.

unlock_pdfA

Remove password protection from a PDF. Requires the correct password. Produces an unprotected .pdf file on disk.

Inputs: password (required), input_file (required), output_path.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 18 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose, from conversions (pdf_to_excel, pdf_to_word) to manipulations (merge_pdf, split_pdf) and utilities (add_watermark, ocr_pdf). No two tools have overlapping functionality, ensuring an agent can reliably select the correct one.

    Naming Consistency3/5

    Naming follows two main patterns: 'source_to_target' for conversions (e.g., pdf_to_excel, office_to_pdf) and 'verb_pdf' for operations (e.g., merge_pdf, rotate_pdf). However, 'add_page_numbers' and 'add_watermark' break the verb_pdf pattern, creating minor inconsistency.

    Tool Count5/5

    With 18 tools, the server covers a comprehensive range of PDF operations without being overwhelming. Each tool serves a specific need, and the count is well-scoped for a PDF utility toolkit.

    Completeness4/5

    The toolkit covers most essential PDF operations: conversion to/from various formats, merge/split/extract, compression, security, OCR, watermarking, and repair. Minor gaps exist (e.g., lack of direct text extraction or annotation), but the set is highly functional.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Enables AI-powered PDF manipulation through natural language commands, integrating with Claude to perform operations like merging, splitting, encrypting, optimizing, and analyzing PDFs.
      16
      -
    • A
      license
      A
      quality
      B
      maintenance
      Convert, compress, merge, split and OCR PDFs plus 100+ file formats (Word, Excel, images, ebooks, video) right inside your AI agent. Exposes 126 GuruPDF tools over MCP — works with Claude, Cursor, VS Code, Windsurf, or any MCP client.
      4
      224 npm
      2
      MIT
    • F
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to perform 13 PDF operations (merge, split, compress, watermark, encrypt, and more) on local files via MCP.
      3
      -
    • A
      license
      C
      quality
      D
      maintenance
      Enables AI assistants to perform PDF operations like merge, split, compress, convert, encrypt, OCR, and summarize using natural language.
      98
      11 npm
      MIT