An MCP middleware that Markdown-ifies everything

What it does

Markdownify is a Model Context Protocol server that exposes a toolbox of converters, letting an MCP client ask for a PDF, image, audio file, Office document, YouTube video, or web page and get back clean Markdown. It is essentially a TypeScript/Bun wrapper that orchestrates underlying utilities—primarily the markitdown Python package—to normalize the world into a format LLMs actually like to ingest.

The interesting bit

The project doesn’t try to reinvent document parsing; it glues a Python virtual environment and a Bun runtime together into an MCP-compatible interface. That slightly awkward polyglot setup is the whole point: it gives AI agents a single, standardized socket to read almost anything without caring which library does the heavy lifting.

Key highlights

• Converts PDFs, images (with OCR), audio (with transcription), DOCX, XLSX, and PPTX into Markdown via MCP tools.
• Also scrapes web pages, Bing search results, and YouTube transcripts into the same format.
• Enforces read boundaries with MD_ALLOWED_PATHS, so file-input tools can be locked to specific directories.
• Ships a Docker image, though the published image is slimmed down to PDF support only; audio and image OCR require a local install with the full markitdown[all] extras.

Caveats

• The published Docker image lacks audio and image OCR; the full feature set requires a local install with the markitdown[all] extras.
• A git-repo-to-markdown tool is referenced in environment variables but omitted from the documented tool list, so its status is unclear.

Verdict

Worth wiring up if you want your MCP client to ingest arbitrary documents without opening five different apps. Skip it if you already have a local Markdown conversion pipeline that your LLM can access directly.