pdf-toolkit-mcp

<div align="center">

pdf-toolkit-mcp

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

       

Install · Tools · Recipes · Architecture · Security · FAQ

</div>

---

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

• Why pdf-toolkit-mcp
• Quick tour
• Install
• Tool catalog
• Workflow recipes
• Architecture
• Configuration reference
• Security model
• Development
• Testing & CI
• Releasing
• Roadmap
• FAQ
• Contributing
• License & trademarks
• Credits

---

🎯 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)

<details> <summary>Click to expand</summary>

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. </details>

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.

<details> <summary><b>1. Assemble</b> — merge + compress + (optional) protect</summary>

merge_pdf([a.pdf, b.pdf, c.pdf])
   → compress_pdf(level="recommended")
   → protect_pdf(password="…")

</details>

<details> <summary><b>2. Convert & secure</b> — Office doc to locked PDF</summary>

office_to_pdf(report.docx)
   → add_watermark(mode="text", text="CONFIDENTIAL", opacity=30)
   → protect_pdf(password="…")

</details>

<details> <summary><b>3. Extract & OCR</b> — pull pages out of a scan and make them editable</summary>

extract_pdf_pages(scan.pdf, pages="3-9")
   → ocr_pdf(languages=["eng"])
   → pdf_to_word

</details>

<details> <summary><b>4. Prep for email</b> — shrink + lock</summary>

compress_pdf(big.pdf, level="extreme")
   → protect_pdf(password="…")

</details>

<details> <summary><b>5. Scan cleanup</b> — repair + OCR + number</summary>

repair_pdf(scan_broken.pdf)
   → ocr_pdf(languages=["eng"])
   → add_page_numbers(position="bottom-center")

</details>

---

🏛 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"

1. Set PDF_TOOLKIT_DEBUG=1 in your shell before launching Claude Code for verbose stderr.
2. 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

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

• iLovePDF / iLoveAPI — the cloud PDF processing API that powers every tool call.
• Anthropic — for Claude, Claude Code, and the Model Context Protocol specification.
• @modelcontextprotocol/sdk, Zod, Vitest, ESLint — the toolchain this project stands on.
• Every contributor who reports issues, suggests tools, and sends pull requests. 🙌