Merge, split and fill PDF forms from Claude

Say "stamp PAID on that invoice and save a copy" or "pull pages 2 to 6 out of this scan and merge them with the contract" and it happens, on your machine, in a second. This MCP server does the small PDF jobs that otherwise send you to a web uploader: merge, split by page range, extract or reorder pages, rotate a sideways scan, stamp PAID or DRAFT or any text you like, put your business name and VAT id in the footer, count pages across a folder of files, and read a PDF's text back as text. No upload, no account, no native dependency, no office install.

In the official MCP Registry (io.github.theluckystrike/pdf-merge-split-stamp-extract-pages).

Listed on the AI Product Index - live remote endpoint at mcp.zovo.one/s/pdf, free tier, no signup.

The PDF chores of a freelance business, done from chat instead of from a browser tab you do not trust with an invoice.

60-second install

npm publish for @theluckystrike/mcp-pdf is pending. Until then, the .mcpb one-click bundle or a clone+build is the working path, both are verified below.

One-click (.mcpb): download pdf.mcpb from the latest release and double-click it in Claude Desktop: https://github.com/theluckystrike/mcp-servers/releases/latest

(claude_desktop_config.json):

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

Claude Code:

claude mcp add pdf -- npx -y @theluckystrike/mcp-pdf

(.cursor/mcp.json):

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

The npx form above starts working the moment the package is published. Until then, use the .mcpb bundle above, or build from source with exactly these three commands:

git clone https://github.com/theluckystrike/mcp-servers.git && cd mcp-servers
npm install
npm run build -w packages/mcp-license -w servers/pdf

Then point your client's command at node with one arg: the absolute path to servers/pdf/dist/index.js.

To run in Pro mode set MCP_LICENSE_KEY in the same config block, or call license_activate once with your key.

Tools

Resource: pdf://recent returns the last 25 operations, newest first, with the inputs, the files written and when. Prompt: mark_invoice_paid chains mcp-invoice's invoice_get into pdf_stamp PAID and saves the stamped copy beside the original.

What you can say

Worked example

You: Stamp PAID on ~/invoices/INV-2026-0007.pdf and save a copy.

  pdf_stamp {
    path: "~/invoices/INV-2026-0007.pdf",
    text: "PAID",
    position: "center",
    out_path: "~/invoices/INV-2026-0007-paid.pdf"
  }
  -> Stamped "PAID" on 1 page
  -> ~/invoices/INV-2026-0007-paid.pdf, 0.9 MB

The stamp is drawn text, not a flattened image: it can be selected and searched, it is drawn on the 45-degree diagonal in the preset green, and the original file is byte-for-byte unchanged.

Free vs Pro

A tier limit is an answer, not an error, and nothing is written when one refuses a call.

Pro is a one-time $19, or $39 for every server in the collection, lifetime.

pdf_text is best effort, and here is exactly what that means

A PDF does not store text. It stores drawing operators, and the text you see is bytes handed to a font. This server decompresses each page's FlateDecode content stream with node:zlib and reads the four text-showing operators -- Tj, TJ, ' and ", plus the positioning operators that end a line. That is a parser written here, in about two hundred lines, with no pdfjs and no native module, and it has three real limits the tool states in its own answer every time:

• A scan has no text. An image-only page carries no text operators, so nothing comes back. There is no OCR here and there will not be one; the answer says the page is probably a scan rather than returning an empty string with no explanation.
• A custom or CID encoding returns glyph indices, not letters. Subset-embedded and CJK fonts map bytes to glyph numbers through a table this parser does not read. When the output has no readable characters, the answer says the font is the reason.
• Reading order is drawing order. There is no column detection and no layout reconstruction. Word spaces are recovered from the large negative kerns a TJ array uses for them, so a PDF that positions every word separately can come back with words run together.

PDFs written by word processors, invoicing tools and report generators, everything with a standard or fully embedded font, read back cleanly. Anything else tells you why it did not.

Encrypted PDFs are refused, not guessed

pdf-lib cannot decrypt, and loading an encrypted file while ignoring the encryption produces pages of garbage that would be written straight into the output. So a file whose trailer names an /Encrypt dictionary is refused by every tool that writes, with the reason and the fix in the message: open it in a reader with the password, export or print to a new PDF, and use that file. pdf_info still answers, with encrypted: true, because reporting is its whole job, and pdf_count reports it as one unreadable file while still counting the rest.

Existing files are never overwritten, and inputs are never modified

Every tool writes a new file and leaves its inputs byte-for-byte alone, there is no in-place mode, on purpose. An out_path that already exists is refused:

Error: /path/merged.pdf already exists and nothing was written.
Pass overwrite: true to replace it, or give a different out_path.

The path is reserved with an exclusive create, not an existence check, so two processes writing the same out_path at the same time cannot clobber each other: one wins, the other is refused and writes nothing. pdf_split reserves every one of its output paths before it writes any of them, so a collision on part 3 does not leave parts 1 and 2 behind as a half-done split, and the reservations are released, so nothing empty is left on disk either.

Pass overwrite: true when replacing the file is what you want.

overwrite: true still does not let an output be an input. Writing the result of an operation back over one of its own sources destroys that source, the pages are already in memory and get written over the file they came from, which takes a three-page file to the one page you extracted, and every later read of that path is then quietly wrong. So an out_path that resolves to (or shares an inode with) any input of the same call is refused before any work happens:

Error: out_path /path/scan.pdf is also an input of this operation, so writing it would destroy the source.
Nothing was written. Write beside it instead - /path/scan-out.pdf - and, if the result really is meant to
take the original's place, rename it yourself once you have checked it.

Limits

• Inputs over 100 MB are refused: rewriting a PDF needs several times its size in memory.
• A file that does not start with %PDF- is refused before anything is read.
• Rotation is recorded as page metadata, in multiples of 90 degrees, which is all the format has. Nothing is redrawn.
• Stamp text goes through a built-in PDF font, which carries WinAnsi and its 256 code points. A character outside it is transliterated where it has an obvious Latin body, Polish OPŁACONE stamps as OPLACONE, and the answer says so and prints what was actually drawn, and removed and counted where it does not, so there is no CJK stamp. A newline is a word separator, never a deletion.
• font_size must be between 1 and 1600 points. A stamp too long to fit even at the smallest size this server will use is still drawn, and the answer says how far past the edge it runs, so you can shorten it.
• pdf_text caps one answer at 200,000 characters and names the pages argument that continues from where it stopped. Nothing is missing from the file; the cut is in the answer, not the read.
• No page numbering, no signatures, no redaction, no compression. This server does the page-level jobs; it is not a PDF editor. It does not write PDF/A: pdf_info reports a file's own pdfa_claim, and every tool that writes says plainly that its output no longer holds that claim.
• Forms are read, not filled: pdf_text prints the AcroForm field values, which live in the fields rather than in the page content stream and are therefore not part of the page text.
• No OCR. See the pdf_text section.

How it stores data

Only a register of what it did: ${XDG_DATA_HOME:-~/.local/share}/mcp-servers/pdf/operations.json, the last 500 operations, as plain JSON. Your PDFs stay where you put them. Every write to the register runs inside an advisory lock on .../pdf/.lock, so two clients on one data directory cannot lose a record, and saves go to a temporary file and are renamed into place.

If the register is unreadable or not valid JSON it is never treated as "empty": it is moved aside byte-for-byte as operations.json.corrupt-<timestamp> with a marker beside it. The PDF you asked for is still written, the file is on disk before the register is touched, and the answer tells you the history could not be updated.

The business footer reads the one shared profile the whole suite uses, ${XDG_DATA_HOME:-~/.local/share}/mcp-servers/profile/business.json, written by business_set in mcp-invoice or mcp-docx. Nothing is invented: with no name stored, pdf_watermark_business refuses and says which tool to run.

Privacy

All data stays local. The server reads and writes files on your machine and makes no network request of any kind -- not for licensing (keys are verified offline), not for fonts, not for telemetry. A PDF you stamp is never uploaded anywhere, which is the entire reason this exists.

Pairs with

• mcp-invoice, the invoice you sent, stamped PAID: the mark_invoice_paid prompt chains invoice_get into pdf_stamp.
• mcp-docx, write the proposal as .docx, print it to PDF, then stamp, merge or split it here.
• mcp-resume, join a CV, a cover letter and a portfolio into the one file an application form accepts.
• office-suite, several servers behind one install, one config entry.

Troubleshooting

• npx hangs or fails to find the package: npm publish for this package is pending. Use the .mcpb bundle or the clone-and-build path above until it lands.
• "is encrypted and was refused": the file is password or permissions protected. Open it in a PDF reader, export or print to a new PDF, and run the tool on that copy. This server never guesses or strips a password.
• pdf_text returned nothing: the page is a scan, or its font uses a custom encoding. The answer says which. There is no OCR here.
• The stamp is too small or too big: by default it is sized to fit the page width. Pass font_size for an exact point size, and opacity between 0 and 1.
• "already exists and nothing was written": pass overwrite: true, or a different out_path.
• Node version: requires Node >= 18. Check with node -v.

MIT licensed. Support: support@zovo.one

Built by theluckystrike.

Frequently asked questions

Is there an MCP server to merge and split PDFs?

Yes. The pdf server at mcp.zovo.one merges, splits, rotates and fills PDF forms while preserving form fields. Free tier of 30 pages per file on split/pages/rotate operations, hosted remote endpoint, no install.

Can Claude fill PDF forms with an MCP server?

Yes - connect https://mcp.zovo.one/mcp/pdf and ask Claude to fill a form's fields or merge several PDFs. Output stays form-preserving, so filled fields remain editable.

Use these docs as an MCP server

Any MCP client (Claude, Cursor, Windsurf, VS Code) can read this repository's documentation directly via GitMCP - no install:

• Docs MCP URL: https://gitmcp.io/theluckystrike/mcp-pdf