Turn Files Into Markdown Before Using Claude

Everything you need

• Microsoft MarkItDown The official repository, supported formats, installation instructions, command-line examples, and security notes.Open
• Official MarkItDown MCP guide Microsoft's installation, Docker, Claude Desktop configuration, usage, and security guidance for markitdown-mcp.Open
• Claude file-upload guide Anthropic's current supported file types, limits, PDF processing behaviour, and advice for large documents.Open

What MarkItDown can convert

• PDF Text and document structure
• Word DOCX document content
• Excel Spreadsheet values and tables
• PowerPoint Slide text and structure
• Images Metadata and OCR when configured
• Audio Metadata and transcription when configured
• Web formats HTML, CSV, JSON, XML, and more
• Other sources ZIP files, EPUBs, and YouTube URLs

How to use it

1. 1

   Decide whether Markdown is the right input

   Use MarkItDown for text-heavy reports, articles, policies, transcripts, Word files, slide text, and straightforward tables. Upload the original PDF when you need Claude to inspect diagrams, charts, photos, handwriting, or the exact page layout. Claude analyses both text and visual elements in PDFs of 100 pages or fewer, so converting everything is not always better.

2. 2

   Install MarkItDown locally

   MarkItDown requires Python 3.10 or newer. Microsoft's repository recommends using a virtual environment. Install the complete converter with: pip install 'markitdown[all]'. Download and install only from the official Microsoft repository or Python package instructions linked below.

3. 3

   Convert your file

   Run: markitdown input.pdf -o output.md. Replace the filenames with your real input and output paths. Open the resulting Markdown file and quickly check headings, tables, page order, and missing text before relying on it.

4. 4

   Upload the Markdown file to Claude

   Attach the checked .md file and tell Claude what outcome you need. Markdown is close to plain text while retaining headings, lists, tables, and links, which makes it well suited to text analysis. Token savings vary, so compare the context or usage shown by your AI tool rather than assuming a fixed percentage.

   Use the attached Markdown document as the source of truth. First map its headings and main sections, then answer my question with concise citations to the relevant heading or quoted passage. Flag any broken tables, missing context, image references, or formatting that may have been lost during conversion. MY QUESTION: [Write your question here]

   ClaudeCopy
5. 5

   Optional: connect the official MCP server

   Microsoft also maintains markitdown-mcp for local MCP clients such as Claude Desktop. The official instructions recommend Docker and show how to add the server to claude_desktop_config.json. The server exposes a convert_to_markdown tool; it does not automatically intercept every ordinary file upload. You or Claude must call the tool, and its returned Markdown still enters the conversation context.

6. 6

   Restrict what the MCP server can read

   The MCP server runs with your local user permissions and can read files that process can access. Use Microsoft's Docker setup, mount only the folder you intend to convert, keep the server bound to localhost, and never expose it to the public internet without proper security controls.

Tips

• The tool is called MarkItDown, not Markdown. Markdown is the output format.
• There is no official universal 70% token-saving guarantee. The result depends on the original file and what the conversion keeps.
• Do not discard the original document. Markdown conversion can lose visual context, layout, annotations, and complex table structure.
• Scanned PDFs may need OCR and a manual accuracy check before the Markdown is trustworthy.
• The MCP tool returns converted content to the client, so very large Markdown output can still consume context.
• Only convert documents you are allowed to process, and avoid giving a local MCP server broader file access than it needs.