What this is
Every file you hand an AI model gets read in full. Not the words you care about, the whole thing: the layout, the tables, the image data, the headers and footers, the invisible structure holding the page together. You pay for all of it, on every message that file stays in the conversation.
MarkItDown is Microsoft's answer. It is free, open source under MIT, and it does one job: it takes almost any file and hands back clean Markdown text. Read on the GitHub API on 9 September 2026 it has 182,033 stars and 13,371 forks. Star counts move, so check the repository page before you quote that figure to anyone.
This guide is the part that would not fit in the video: the install, the exact commands, the MCP server config for Claude Desktop word for word, the formats table, what it handles badly, and how to measure the saving on your own documents instead of taking mine.
- Part 1, why a PDF is expensive, and what is actually costing you.
- Part 2, install it, in one line.
- Part 3, the three commands you will actually use.
- Part 4, the MCP server, so you never think about this again.
- Part 5, what it does badly, and the two ways round it.
- Part 6, measure your own saving in about two minutes.
1. Why a PDF is expensive
A PDF is not a document. It is a set of drawing instructions for putting marks in positions on a page. There is no paragraph, no heading, no table, only glyphs with coordinates. Every tool that reads one has to reconstruct the meaning from the geometry, and everything it cannot reconstruct it passes along anyway.
- Layout and positioning data, which carries no meaning once the text is extracted.
- Tables flattened into runs of loose cells, which is why a model so often gets a number from the wrong row.
- Images, which are expensive per unit and frequently decorative.
- Repeated headers, footers, page numbers and watermarks, once per page, for every page.
Markdown is the opposite. A heading is a hash, a list is a dash, a table is pipes. It is close to plain text, so it is cheap, and mainstream models were trained on enormous amounts of it, so the structure actually survives into the answer.
That is the real reason the answers get better, and it is worth being precise about it. Conversion is not making the model smarter. It is removing the part of the job where the model was guessing at a layout instead of reading content.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (7 more sections, 10 copy-paste prompts) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: 2. Install it · 3. The three commands you will actually use · 4. The MCP server, so you never think about it again · 5. What it does badly, and the two ways round it · 6. Measure your own saving · +2 more
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.