PDF MCP Server — 14 PDF Tools for Claude, Cursor and AI Agents

Most AI assistants can read a PDF you paste into the chat, and that is where it ends. They cannot merge a stack of invoices, strip the blank pages out of a scan, watermark a draft contract, shrink a 60 MB deck, or rebuild a file that will not open. The PDFWix MCP server closes that gap: one endpoint, 14 production tools, and structured results your assistant can chain together inside a single conversation.

An MCP server is a service that hands an AI assistant a set of real tools over the Model Context Protocol — the open standard Anthropic published in late 2024 and now supported by Claude, Cursor, Windsurf, ChatGPT developer connectors, Gemini CLI, Copilot Studio and most agent runners. The model no longer guesses at an answer: it calls a named function with typed arguments and gets a structured result back, exactly like a developer calling an API.

A PDF MCP server applies that to documents. PDFWix exposes 14 PDF operations — metadata inspection, text extraction, merge, split, page removal, reordering, rotation, page numbering, watermarking, compression, repair, unlocking, images-to-PDF and a browser handoff — over a single HTTPS endpoint with OAuth sign-in. One paste of a URL, and your assistant can actually finish document work instead of describing it.

Chaining PDF tools by hand means knowing every argument name. Describe the job instead and the builder returns an ordered pipeline with valid arguments and copyable JSON-RPC calls for this server — useful when you are writing an agent loop rather than chatting.

Every major MCP client accepts a remote server URL. Pick your client below. The first tool call opens a browser window where you approve access with a free PDFWix account — after that the connection stays live.

Claude (web and desktop, remote connector)

In Claude, open Settings → Connectors → Add custom connector, paste the URL below and approve the sign-in prompt. If you prefer the one-click route, use the public listing on the PDFWix Claude connector page.

https://www.pdfwix.com/api/mcp

Claude Desktop config file

{
  "mcpServers": {
    "pdfwix": {
      "url": "https://www.pdfwix.com/api/mcp"
    }
  }
}

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "pdfwix": {
      "url": "https://www.pdfwix.com/api/mcp",
      "transport": "http"
    }
  }
}

Windsurf, Open WebUI, LM Studio, AnythingLLM

All of these speak MCP natively. Add a remote/HTTP server, paste https://www.pdfwix.com/api/mcp, leave every credential field empty — there is nothing to copy, because authorisation happens in the browser rather than through a pasted key.

Claude Code (command line)

One command registers the server, then /mcp starts the sign-in:

claude mcp add --transport http pdfwix https://www.pdfwix.com/api/mcp
# then, inside Claude Code:
/mcp   →  select pdfwix  →  Authenticate

Before you debug a client, confirm the server side end to end. The panel below runs the same four checks an MCP client performs — the RFC 9728 discovery document, the unauthenticated authorization challenge, your PDFWix account and token format, and the published tool catalog — and gives you a report you can paste into a support ticket or a registry submission.

If you open the endpoint in the Glama MCP Inspector, the official MCP Inspector, Postman or curl with no credentials and press Connect, you will see {"error":"unauthorized"} with HTTP 401. That is the server working correctly: PDFWix requires an account so your files land in private storage nobody else can reach.

The fix: choose OAuth 2.0, not a pasted token

In the inspector's authentication settings pick OAuth 2.0 (leave client ID and secret empty — the server registers the client for you) and press Connect. The inspector reads our discovery document, opens the PDFWix sign-in window, and the status turns green with all 14 tools listed under the Tools tab.

An access token copied out of a browser session will not work here, and no manual token is needed. The MCP endpoint only accepts tokens issued through the OAuth flow, which is what keeps one account's documents unreachable from another's.

Verify the server is live from your terminal

These two calls confirm the endpoint and its discovery metadata without any credentials:

# 1. discovery document (expect 200 with JSON)
curl -s https://www.pdfwix.com/.well-known/oauth-protected-resource/api/mcp

# 2. the endpoint itself (expect 401 + a WWW-Authenticate header)
curl -i -X POST https://www.pdfwix.com/api/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

The first returns the resource metadata naming https://www.pdfwix.com/api/mcp and its authorisation server. The second returns 401 plus a WWW-Authenticate header pointing at that same document — which is precisely the signal MCP clients follow to start the sign-in. Both responses mean the server is healthy.

MCP is becoming the shared standard, but each platform is at a different stage. Here is exactly where PDFWix plugs in today and what to do on platforms that speak a different protocol.

OpenClaw and other autonomous agent runners

OpenClaw, OpenHands, AutoGPT-style runners and Claude Code all read a JSON server list. Add PDFWix as a remote HTTP server and approve the browser sign-in the first time a tool runs; the runner stores the authorisation and refreshes it on its own afterwards.

{
  "mcpServers": {
    "pdfwix": {
      "type": "http",
      "url": "https://www.pdfwix.com/api/mcp"
    }
  }
}

For a fully headless machine, run the first authorisation once on a desktop with the same config directory, then copy the stored credentials across — the runner needs a browser only for that initial approval.

ChatGPT — Developer mode connector or a custom GPT action

ChatGPT now supports remote MCP servers directly. In ChatGPT, open Settings → Connectors → Advanced → Developer mode, choose Create, paste https://www.pdfwix.com/api/mcp, set authentication to OAuth and approve the PDFWix sign-in. The 14 tools then appear in chat, and in Deep Research and agent runs.

If your plan does not expose developer mode, build a custom GPT instead: Explore GPTs → Create → Configure → Create new action, and point the action at the PDFWix REST endpoints documented on the developer page, with https://www.pdfwix.com/privacy-policy as the privacy policy URL. Custom GPT actions speak OpenAPI, not MCP, which is why the endpoint differs.

Google Gemini — Gemini CLI, Antigravity and Vertex AI

Gemini CLI and Google's agent tooling read MCP servers from a settings file. Create or edit ~/.gemini/settings.json:

{
  "mcpServers": {
    "pdfwix": {
      "httpUrl": "https://www.pdfwix.com/api/mcp"
    }
  }
}

Restart the CLI and run /mcp to confirm the 14 PDFWix tools are listed. In the Gemini app itself there is no custom connector slot yet; for Vertex AI agents, register the same endpoint as an MCP toolset in your agent definition.

Microsoft Copilot Studio and Copilot agents

In Copilot Studio open your agent, go to Tools → Add a tool → New tool → Model Context Protocol, paste https://www.pdfwix.com/api/mcp, pick OAuth 2.0 authentication and publish. Makers then approve PDFWix once per user.

Local and self-hosted assistants

Open WebUI, LM Studio, AnythingLLM, Ollama front-ends, Continue and Cline all accept a remote MCP URL in their tool settings. Paste the endpoint, choose HTTP/streamable transport, leave the credential fields empty and complete the browser sign-in when it appears.

Meta AI and other closed assistants

Meta AI's consumer apps have no public plugin surface today, so there is nothing to install there. If you run a Llama model yourself through any of the front-ends above, PDFWix works immediately over the same endpoint.

Anything else — the raw JSON-RPC route

Any system that can send an HTTP POST can call PDFWix: hit the endpoint with the tools/list and tools/call methods shown earlier, carrying the access token your OAuth client obtained. That covers n8n, Make, Zapier webhooks, LangChain, CrewAI and your own scripts — use a standard OAuth 2.1 authorisation-code flow with PKCE against the authorisation server named in our discovery document.

Inspection and extraction

• get_pdf_info — page count, file size, page dimensions, title, author, creation and modification info, and whether the file is encrypted.
• extract_pdf_text — clean text with optional page ranges, parsed server-side without any browser rendering, capped at 120,000 characters per response with an explicit truncation notice so your assistant knows when to ask for the next range.

Document assembly

• merge_pdfs — combine several PDFs in the order you name them.
• split_pdf — break one file into ranges or single pages.
• remove_pages — delete blank scans, cover sheets or duplicated pages.
• reorder_pdf_pages — rearrange pages into any sequence.
• rotate_pdf — fix sideways or upside-down scans, per page or whole file.

Styling, optimisation and security

• add_page_numbers — position, start number and formatting under your control.
• add_watermark — text watermarks with opacity, rotation and placement, including non-Latin scripts.
• compress_pdf — reduce file size for email and upload limits.
• repair_pdf — rebuild damaged or partially downloaded files that refuse to open.
• unlock_pdf — remove a password you already know from a protected file.

Conversion and browser handoff

• images_to_pdf — turn a batch of JPG, PNG or WebP images into one PDF.
• open_in_pdfwix — for anything that genuinely needs a canvas (placing a signature, drawing redaction boxes, filling form fields, OCR on image-only scans, page-level visual editing) the server returns the exact tool link, for example pdfwix.com/sign-pdf, so you finish in two clicks instead of hitting a dead end.

• "Using PDFWix, inspect this PDF and tell me the page count, dimensions and whether it is encrypted."
• "Extract the text from pages 4 to 9 with PDFWix and summarise the payment terms."
• "Merge these three invoices with PDFWix, drop the blank last page, then compress the result."
• "Watermark this contract with DRAFT at 30% opacity using PDFWix."
• "This PDF will not open — repair it with PDFWix and confirm the page count."
• "I need to sign this file myself. Give me the PDFWix link for that."

Authorisation uses OAuth 2.1 with dynamic client registration and per-user, token-scoped access, discovered over RFC 9728 from our own origin. Documents you send arrive as an HTTPS URL or a base64 payload, are processed in isolated server functions, and land in private per-user storage that nobody else can list or read. Download links are signed and expire after 24 hours.

We never train AI models on your documents and never store document content in our database — only operational metrics such as which tool ran, whether it succeeded and how long it took. There is no sponsored or promoted content in any tool response, no payment handling, and no AI-generated media. Each document is capped at 100 MB; beyond that, compress or split first. Full detail lives on the security page, alongside our privacy policy and terms.

What is the difference between an MCP server and a plugin?

A plugin is built for one product and lives inside it. An MCP server is a single service that any MCP-capable client can consume, so the same PDFWix endpoint serves Claude, Claude Code, Cursor, Windsurf, ChatGPT developer connectors, Copilot Studio and local assistants without a separate build for each.

How do I add an MCP server to Claude Code?

Run claude mcp add --transport http pdfwix https://www.pdfwix.com/api/mcp, then type /mcp in Claude Code and pick authenticate. Yes, Claude Code can use MCP servers on every plan that includes Claude Code, and the PDFWix side costs nothing.

Is the PDFWix MCP server free on the Claude free plan?

Yes. Remote connectors are usable on paid Claude plans and through Claude Code; the PDFWix server itself never charges and never asks for a card, whichever plan you are on.

Can Claude or ChatGPT edit a PDF with this?

They can change a PDF's structure and content in the ways listed above — pages, rotation, numbering, watermarks, size, repair, unlocking. Drawing on a page (signatures, redaction boxes, form fields) stays in your browser, and the assistant gives you the exact link.

Is the PDFWix MCP server really free?

Yes. Every tool listed above is free to use with a free account, and no credit card is requested at any point. Read the reasoning on how we make money.

Why does it ask me to sign in?

Sign-in is what keeps your files in a private space that only your token can reach, and it stops anonymous abuse of the shared processing capacity. It takes one click with Google.

Which tools are not available over MCP?

Anything that needs to draw on a page — signing, manual redaction, form filling, visual page editing, PDF-to-image rendering, cropping and visual comparison. Those run in your browser, and open_in_pdfwix hands you the direct link.

Can I use it in ChatGPT or Gemini?

Yes. ChatGPT connects through Developer mode connectors (or a custom GPT action for the REST endpoints), and Gemini CLI reads the same endpoint from its settings file — both are covered step by step under "Using PDFWix in other AI platforms" above. MCP-native clients such as Claude, Cursor, Windsurf, OpenClaw and local Llama front-ends work straight away.

Where do I report a problem?

Email support@pdfwix.com or use the contact form. Live service health is on the status page.

• PDFWix Claude connector — one-click setup and prompt recipes for Claude.
• Developer API — the REST side of PDFWix for your own backend.
• Security architecture — where files are processed and for how long.
• All 24 PDF tools — the browser versions of everything above.