Markdownify MCP

What the zcaceres/markdownify-mcp MCP server does

The zcaceres/markdownify-mcp MCP server turns several document, media, and web inputs into Markdown. Its MCP tools cover PDF, DOCX, XLSX, and PPTX files; images with metadata; audio with transcription; YouTube videos; Bing search results; and general web pages. A separate tool retrieves an existing Markdown file when its extension is .md or .markdown.

The output is intended to give an AI client a readable text representation of source material without requiring each client workflow to implement its own file or web conversion logic. The repository uses the markitdown executable for conversion-related work and includes support for a broader set of formats in the local installation.

How it works

The server runs as a local MCP process and exposes conversion operations as callable tools. A client sends an input such as a file path, web page, search request, or YouTube video to the corresponding tool, and the server returns Markdown-formatted content. Audio conversion includes transcription, while image conversion includes image metadata.

For file-based operations, the server can enforce a read boundary. Set MD_ALLOWED_PATHS to one or more permitted directories; path separators are : on POSIX systems and ; on Windows. MD_SHARE_DIR remains available as a deprecated single-directory alias. Without either setting, the README describes file access as unrestricted.

Setup and configuration

The documented local setup is:

1. Clone the repository.
2. Run bun install to install dependencies and create the Python virtual environment containing markitdown[all].
3. Run bun run build.
4. Start the server with bun start.

The server can also run in Docker. The documented image mounts host directories into the container and sets MD_ALLOWED_PATHS to the corresponding container paths. File tools must receive container paths rather than the original host paths.

The supported configuration variables are:

• MARKITDOWN_PATH: selects the markitdown executable, overriding the bundled virtual-environment path.
• REPOMIX_PATH: selects the repomix executable used by git-repo-to-markdown.
• MD_ALLOWED_PATHS: restricts file-input tools to specified directories.
• MD_SHARE_DIR: deprecated compatibility alias for a single allowed directory.

Tools and capabilities

The zcaceres/markdownify-mcp MCP server provides these tools:

• youtube-to-markdown for YouTube videos
• pdf-to-markdown for PDF files
• bing-search-to-markdown for Bing search results
• webpage-to-markdown for web pages
• image-to-markdown for images and metadata
• audio-to-markdown for audio files and transcription
• docx-to-markdown for DOCX files
• xlsx-to-markdown for XLSX files
• pptx-to-markdown for PPTX files
• get-markdown-file for existing .md and .markdown files

Limitations and notes

The published Docker image installs markitdown[pdf] rather than the full extras set. As a result, the README notes that audio transcription and image OCR require the local installation with markitdown[all] and may fail in the slim Docker image.

The desktop-client configuration uses the built dist/index.js file and requires an absolute path to that file. Docker users must mount every host directory the server needs to read and pass the mounted container location to the tools. The repository is licensed under MIT.