Give Your AI Assistant the Power to Convert Documents Itself (MCP)

There's a small, repetitive friction in working with AI on documents. You have a PDF. You want to ask Claude about it. So you leave your conversation, go find a converter, turn the PDF into Markdown, come back, paste it in, and then ask your question. Every single time. The assistant is right there, capable of reasoning about the content — but it can't reach the file, so you become the glue between the document and the model.

MCP removes that step. It lets your assistant convert the document itself — no leaving the chat, no manual conversion, no paste. Here's what that means and how it works.

What MCP actually is

MCP — the Model Context Protocol — is a standard way to give an AI assistant new tools. An MCP "server" runs on your machine and exposes capabilities the assistant can call directly. Clients like Claude Code and Claude Desktop speak the protocol, so any MCP server you add becomes a set of actions the assistant can perform on its own.

A document-conversion MCP server exposes exactly one useful capability: "convert this file to Markdown." Once it's connected, your assistant can take a file path and get back clean Markdown without you doing anything. The conversion step disappears from your workflow entirely.

The workflow, before and after

Before: open converter → upload/convert file → copy Markdown → return to chat → paste → ask.

After: "Convert the report on my desktop and summarise the risks section." The assistant calls the tool, gets the Markdown, and answers. One message.

That's the whole pitch, and it's a bigger quality-of-life change than it sounds, because it turns "document work" from a chore you do around the assistant into something the assistant just does.

The part that makes it viable: it's local

Here's why this matters more for documents than for most MCP tools. Handing an assistant the power to convert files is only comfortable if those files aren't being shipped somewhere. MarkPrep's MCP server runs the conversion on your machine — the same in-browser engine, packaged for the command line. The file the assistant converts never leaves your device; the only network call is a quick licence check. So you can point it at a contract or an internal report without that document travelling to anyone's server.

That local-first design is what makes "let the AI touch my files" a reasonable thing to do rather than a security headache.

Setting it up

It's a small config addition. You add the MarkPrep server to your MCP client's config with your licence key:

{
  "mcpServers": {
    "markprep": {
      "command": "npx",
      "args": ["-y", "markprep", "mcp"],
      "env": { "MARKPREP_LICENSE": "your-licence-key" }
    }
  }
}

Restart the client and your assistant has a convert_document tool. The same package also gives you a plain CLI — npx markprep convert report.pdf — for scripts and terminals, and a watch mode that converts anything dropped into a folder. Full setup lives on the MCP page.

Who this is for

If you paste documents into an AI a few times a month, the manual way is fine. If you do it constantly — feeding contracts, reports, research, or a whole knowledge base into an assistant — removing the conversion step compounds fast, and doing it locally means you can include the sensitive files you'd never upload to a web converter.

The takeaway

The manual convert-then-paste dance exists only because your assistant couldn't reach your files. MCP closes that gap: connect the server and the assistant converts documents to Markdown itself, on your machine, keeping the file private. The conversion step doesn't get faster — it disappears.