All three say they merge, split and stamp a PDF from chat. pdf-mcp
(npm package mcp-pdf) goes much further than page-level jobs, adding form filling, annotations
and signature fields. DocWand runs the opposite way: no install at all, a
hosted MCP endpoint whose signing and form-filling actually execute in your browser tab rather than on the
server. Ours sits between them: page-level jobs only, done locally, with an honest answer about the one
thing none of the three can promise -- reading text out of a PDF whose font was not built for it.
| Fact | Ours | pdf-mcp | DocWand |
|---|---|---|---|
| Tools | 12 (10 pdf tools plus 2 license tools) | ~24 across documents, forms, pages, signatures, security, export | 4 documented flows: sign, fill, merge, split |
| Transport | stdio, no hosted route yet | stdio (npx or global install) | streamable HTTP only, hosted at mcp.docwand.app |
| Runs where | Your machine, no network call of any kind | Your machine | Your browser, on the DocWand site, files stay client-side |
| Text extraction | pdf_text: names the reason when it fails (scan, or subset/CID font returning glyph numbers) | Extract Text listed as a feature; the README marks it "requires additional implementation" | Not offered; it is a sign-and-fill tool, not an extraction tool |
| Encrypted PDFs | Refused with the fix named in the message; pdf_info still reports encrypted: true | Encrypt/decrypt are listed features with limited support noted | Not documented |
| Account or key needed | No. No account, no API key | No. No account, no API key | No account for the free tools; nothing to sign up for |
| Licence | MIT | Not stated in the npm record | Not published; closed-source hosted service |
| Cost | Free tier, Pro $19 once | Free, per its npm listing | Free tier per its pricing page |
Pick it for form work: it lists form field discovery, filling single or multiple fields, exporting form
data, flattening forms and adding annotations, none of which we do at all. Its own README also marks
several export features, including text extraction, as "requires additional implementation," so check
that a feature you need is actually wired up before relying on it rather than trusting the feature list on
its own. It installs the same way ours does, with npx in a Claude Desktop or Cursor config.
Pick it when you want zero install and the file never has to touch a server you do not control at all, including ours: DocWand's own description is that signing, filling, merging and splitting run entirely in the browser tab, so the PDF never leaves the device even over its hosted MCP endpoint. It does not do page-level jobs like rotate, reorder or per-page stamping, and it has no page-count or metadata reporting tool; it is a signing and light-editing tool, not a page-manipulation one.
Pick ours for the specific page-level jobs a freelancer actually repeats: merge, split by range, extract
or reorder pages, rotate a scan, stamp PAID or DRAFT or custom text, and a business watermark that reads the
same shared profile MCP Invoice and MCP Docx write. It runs
fully local with no network code at all, not even for licensing, and it is specific about the one thing it
cannot do well: pdf_text is a two-hundred-line parser with no pdfjs dependency, so
a subset-embedded or CID-encoded font returns glyph index numbers instead of letters, and the answer names
the font as the reason rather than returning that garbage as if it were real text.
The pdf package carries 39 automated tests, all passing, covering merge, split, stamping, rotation, encrypted-file refusal, the operations register and the shared business profile. Neither competitor publishes a test count in its README or npm record.
Ours, Claude Code:
claude mcp add pdf -- npx -y @theluckystrike/mcp-pdfpdf-mcp:
claude mcp add pdf-mcp -- npx -y mcp-pdfDocWand has no install line: it is a remote connector URL, https://mcp.docwand.app/mcp,
pasted into a client that accepts a remote MCP server the way connect-by-URL describes.
Exact config file paths per client are on the setup pages, and the longer walkthrough, including the glyph-index caveat in detail, is in merging, splitting and stamping PDFs from chat.
Yes. MCP clients load several servers at once and the tool names do not collide: ours are pdf_merge, pdf_stamp and so on; pdf-mcp uses its own tool names for the same idea. Nothing is shared between them.
Ours does, with named limits: no OCR for scans, and glyph numbers instead of letters when a subset or CID font is in the way, both stated in the answer. pdf-mcp lists text extraction as a feature but marks it "requires additional implementation" in its own README. DocWand does not offer extraction at all.
Ours never does: no network call of any kind. pdf-mcp runs locally over stdio the same way. DocWand's processing runs in your browser tab even though the connection is a hosted MCP endpoint, per its own description.
pdf-mcp is free per its npm listing. DocWand's core tools are free per its pricing page. Ours has a free tier and a one-time $19 Pro key, or $39 for the whole server collection, with no subscription.
pdf-mcp's tool list and text-extraction caveat come from its README at github.com/nitaiaharoni1/pdf-mcp and its npm record for mcp-pdf. DocWand's tool set and hosted endpoint come from its official MCP registry entry, app.docwand/pdf, and docwand.app. All read on 2026-09-03.