PDF Toolkit MCP

A write-capable PDF toolkit for any MCP client. It provides 22 tools for reading, creating, rendering, transforming, and securing PDFs. That includes rendering pages to images so vision models can read scanned documents, building PDFs from Markdown or structured data, AES-256 encryption, and merge and split operations that keep form fields intact. There are no native dependencies, so it runs locally from a single npx command.

npx -y @aryanbv/pdf-toolkit-mcp

It needs no config files, API keys, Docker, or compiler, and it works offline.

Overview

Most PDF servers for MCP only read. This one also writes: it creates documents from Markdown or structured data, fills and flattens forms, rearranges page structure, and applies AES-256 encryption, all without a native build toolchain.

A few things worth knowing:

• It reads scans. pdf_render_pages rasterizes pages to images, so a vision-capable model can read scanned or image-only PDFs that have no text layer.
• Merge, split, reorder, and delete preserve AcroForm fields rather than dropping them. Names that collide between inputs are namespaced per source, and every call reports what it preserved, renamed, or dropped.
• Encryption is AES-256 through qpdf, not the legacy RC4 scheme.
• Every engine is WASM or plain JavaScript, so npx works on Node 20 and later across Windows, macOS, and Linux with no node-gyp, canvas binding, or prebuilt binary.
• Errors carry stable codes, stack traces stay internal, off-page placements are rejected instead of silently clipped, and large responses are truncated without breaking JSON.

Client setup

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "pdf-toolkit": {
      "command": "npx",
      "args": ["-y", "@aryanbv/pdf-toolkit-mcp"]
    }
  }
}

claude mcp add pdf-toolkit -- npx -y @aryanbv/pdf-toolkit-mcp

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "pdf-toolkit": {
      "command": "npx",
      "args": ["-y", "@aryanbv/pdf-toolkit-mcp"]
    }
  }
}

VS Code uses "servers", not "mcpServers". Copying another client's config will fail silently. This also requires the GitHub Copilot extension with Agent mode.

Add to .vscode/mcp.json:

{
  "servers": {
    "pdf-toolkit": {
      "command": "npx",
      "args": ["-y", "@aryanbv/pdf-toolkit-mcp"]
    }
  }
}

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "pdf-toolkit": {
      "command": "npx",
      "args": ["-y", "@aryanbv/pdf-toolkit-mcp"]
    }
  }
}

Once connected, ask for what you want in plain language and the client selects the tool and fills in the arguments. The JSON blocks below show the arguments each tool accepts, for reference.

Tools

Create PDFs from Markdown

Turn Markdown into a multi-page PDF in a single call. It supports CommonMark and GFM: headings, bold and italic, tables, ordered and bullet lists, fenced code, and blockquotes, rendered with @react-pdf/renderer.

"Create a PDF from this Markdown report."

pdf_create_from_markdown arguments:

{
  "markdown": "# Quarterly Report\n\nRevenue grew **23% YoY**.\n\n| Region | Q1 2025 | Q1 2026 |\n|--------|---------|--------|\n| Americas | $1.2M | $1.5M |\n| EMEA | $800K | $960K |\n\n## Key Wins\n\n1. 12 new enterprise contracts\n2. Churn down to 3.1%",
  "outputPath": "/path/to/report.pdf",
  "pageSize": "Letter"
}

Tables size their columns to content and honor alignment, nested lists indent, and long code lines wrap. Add page numbers afterward with pdf_add_page_numbers.

Templates

Generate documents from structured data using the invoice, report, and letter templates.

"Create an invoice for Riverbend Outfitters."

pdf_create_from_template arguments:

{
  "templateName": "invoice",
  "data": {
    "companyName": "Northpoint Design",
    "clientName": "Riverbend Outfitters",
    "invoiceNumber": "2026-0042",
    "invoiceDate": "2026-04-01",
    "items": [
      { "description": "Website redesign", "quantity": 40, "unitPrice": 150 },
      { "description": "Annual hosting", "quantity": 1, "unitPrice": 299 }
    ],
    "taxRate": 18,
    "currency": "USD",
    "paymentTerms": "Net 30"
  },
  "outputPath": "/path/to/invoice.pdf"
}

The invoice template's optional currency accepts an ISO code or a symbol. WinAnsi-safe symbols ($ € £ ¥) render as glyphs; a code that Helvetica cannot draw, such as INR, KRW, or TRY, falls back to its ISO code label (INR 20.00), so any currency works without error. The pdf-toolkit://templates resource lists every template and the fields it accepts.

Read scanned and image-only PDFs (vision)

Many PDFs are scans with no text layer. pdf_render_pages rasterizes pages so a vision-capable client can read them.

"Read this scanned contract."

Inline mode returns pages as images the model reads directly (up to 5 pages; DPI is auto-capped to protect the context window):

{ "filePath": "/path/to/scanned.pdf", "inline": true }

Or write image files to disk (default 150 DPI, first 50 pages, PNG):

{
  "filePath": "/path/to/scanned.pdf",
  "pages": "1-3",
  "dpi": 200,
  "format": "jpeg",
  "outputDir": "/path/to/output"
}

Convert a PDF to Markdown

"Convert report.pdf to Markdown so I can summarize it."

pdf_to_markdown reconstructs reading order from text positions. It clusters up to two content columns (plus full-width title and footer bands), infers headings from font size, and detects lists. It works best on clean digital PDFs; use pdf_render_pages for scans. Returns the first 10 pages by default.

{ "filePath": "/path/to/report.pdf", "pages": "1-5" }

Search and compare

"Find every mention of 'indemnification' in contract.pdf."

pdf_search arguments:

{
  "filePath": "/path/to/contract.pdf",
  "query": "indemnification",
  "caseSensitive": false
}

Each match comes back with its page number and a surrounding snippet. Matching is a literal, case-insensitive substring by default; set caseSensitive: true for exact case. Regex search is intentionally left out, because an attacker-supplied pattern can trigger catastrophic backtracking (ReDoS) that single-threaded JavaScript cannot reliably interrupt. Safe regex is planned for a later release.

"What changed between v1.pdf and v2.pdf?"

pdf_compare arguments:

{ "filePathA": "/path/to/v1.pdf", "filePathB": "/path/to/v2.pdf" }

It reports a page-by-page text diff (added and removed) and sets identical: true when the text matches. The diff is text only, so purely visual changes are not detected.

Form-preserving merge, split, delete, and flatten

Merging, splitting, reordering, and deleting pages preserve AcroForm fields. Names that collide across inputs are namespaced per source, and each tool returns { preserved, renamed, dropped }, where renamed is a list of { from, to } pairs (address a renamed field by its to name afterward). These tools and pdf_flatten also return a flattened boolean.

"Merge these three forms and flatten the result."

pdf_merge arguments:

{
  "filePaths": ["/path/a.pdf", "/path/b.pdf", "/path/c.pdf"],
  "outputPath": "/path/merged.pdf",
  "flatten": true
}

"Remove pages 2 and 5 from report.pdf."

pdf_delete_pages arguments:

{
  "filePath": "/path/report.pdf",
  "pages": "2,5",
  "outputPath": "/path/trimmed.pdf"
}

Use pdf_flatten on its own to bake an existing form's values into static content. The output path must differ from the input.

Encryption

"Encrypt report.pdf with the password 'secure123'."

Encryption is AES-256. Set separate user (open) and owner (edit) passwords for granular access; the owner password defaults to the user password when omitted.

pdf_encrypt arguments:

{
  "filePath": "/path/report.pdf",
  "outputPath": "/path/report-encrypted.pdf",
  "userPassword": "secure123",
  "ownerPassword": "admin456"
}

QR codes and barcodes

"Add a QR code linking to our website on page 1."

pdf_embed_qr_code supports QR Code, Code128, DataMatrix, EAN-13, PDF417, and Aztec. Position and size are configurable, the symbology's aspect ratio is preserved, placement is rotation-aware, and off-page placements are rejected instead of clipped.

Guided prompts

The server ships five MCP prompts that script multi-step workflows for the client:

Resources

pdf-toolkit://templates is a JSON resource that lists the templates available to pdf_create_from_template and the fields each one accepts.

Try it in plain language

• "Create a PDF from this Markdown report"
• "Generate an invoice for Riverbend Outfitters, 10 hours of consulting at $150/hr"
• "Merge january.pdf and february.pdf into q1-combined.pdf"
• "Convert this PDF to Markdown so I can summarize it"
• "Render this scanned PDF so you can read it"
• "Search contract.pdf for 'termination'"
• "Compare draft-v1.pdf and draft-v2.pdf"
• "Fill the Name field with 'John Doe' in application.pdf"
• "Add a CONFIDENTIAL watermark to draft.pdf"
• "Encrypt financials.pdf with the AES-256 password 'budget2026'"
• "Embed a QR code with our URL on the cover page"
• "Reorder pages as 3,1,2 in report.pdf"

Errors and output semantics

• Coded errors. Validation and load failures throw a PdfError with a stable code, surfaced as Error [CODE]: message (for example FILE_NOT_FOUND, NOT_A_PDF, PAGE_OUT_OF_RANGE, ENCRYPTED_PDF, RESOURCE_LIMIT). Clients can branch on the code instead of parsing message text, and stack traces are never leaked.
• Write-tool output. Write tools create a file at outputPath and return that path plus its size as text, since MCP has no file-content type. outputPath can name an existing file and will overwrite it, so choose a path that does not collide with something you want to keep.
• JSON-safe truncation. Responses are capped at 25,000 characters. Object payloads return a valid { truncated, note, preview } envelope rather than a string cut mid-token, so a client's JSON.parse never breaks.

Known limitations

• Merge, split, reorder, delete. Form fields are preserved, and colliding names are namespaced and reported in renamed as { from, to } pairs. Unusual forms that cannot be safely reconstructed are reported under dropped rather than failing the operation.
• Text extraction. Returns PDF stream order, not visual reading order. Use pdf_to_markdown when reading order matters; raw pdf_extract_text can interleave multi-column layouts.
• PDF to Markdown. Reconstructs up to two content columns (plus full-width title and footer bands); pages with three or more columns fall back to single-column reading order. It works best on clean digital PDFs. Tabular content is emitted as positioned text in reading order, not rebuilt as Markdown tables.
• Markdown to PDF. Supports CommonMark and GFM (headings, bold and italic, links, lists, tables, fenced code, blockquotes, and horizontal rules). Raw HTML, task-list checkbox state, footnotes, and code syntax highlighting are not supported.
• Compare. Text-only diff; visual or layout changes that do not alter text are not detected.
• Image embedding. JPEG and PNG only. Off-page placements are rejected with a coded error instead of being silently clipped.
• Fonts. Built-in fonts are Latin-only (WinAnsi). For non-Latin scripts such as Arabic, CJK, or Devanagari, pass a .ttf or .otf file through fontPath to pdf_fill_form or pdf_create. Markdown and template PDFs use Helvetica by default.

Tech stack

A multi-engine design. Every engine is pure WASM or JavaScript:

Requirements

• Node.js 20 or later. Node 18 and the 20.x line are end-of-life, so Node 22 or 24 LTS is recommended.

Development

npm install        # install dependencies
npm run build      # compile TypeScript
npm test           # run the vitest suite (160 tests)
npm run test:cov   # tests with coverage
npm run lint       # ESLint
npm run format     # Prettier
npm run inspect    # MCP Inspector (requires Node >= 22.7.5)

See CLAUDE.md for architecture and contribution notes.

License

MIT