MarkItDown MCP Server

MarkItDown MCP is the official MCP server for MarkItDown, Microsoft’s open-source converter that turns documents into Markdown for LLMs. It exposes one tool, convert_to_markdown(uri), which takes an http:, https:, file: or data: URI and returns Markdown with headings, lists and tables kept. It is handy when you want your AI to read a PDF report, a Word file, a spreadsheet or a web page without pasting it in.

We tested it on Windows with Python 3.12: listing tools, converting an HTML page (the table came back as a proper Markdown table) and converting a PDF.

Setup, two ways:
1. Docker (recommended in the README for Claude Desktop): clone the repo, run docker build -t markitdown-mcp:latest . in packages/markitdown-mcp, then use the config in the asset. To read local files, mount a folder with -v /your/folder:/workdir and pass paths under /workdir.
2. pip: pip install markitdown-mcp, then use “command”: “markitdown-mcp”. Tip from our test: the current PyPI release installs only the base converter, so also run pip install “markitdown[all]” to get PDF and Office support. If Windows blocks the markitdown-mcp.exe launcher, use “command”: “python” with “args”: [“-m”, “markitdown_mcp”] instead.

It can also run over Streamable HTTP or SSE with markitdown-mcp –http –host 127.0.0.1 –port 3001.

Good to know: there is no authentication, and the tool can read any file the server’s user can read and fetch any URL. Keep it on localhost (the default) and run it in Docker if you want it limited to one folder.

Licence: MIT. Credit to Microsoft and the AutoGen team.