PDF to Markdown Converters I've Actually Tried in 2026

I've been converting PDFs to Markdown almost daily for about a year, mostly for two reasons: feeding API docs into Claude Code while building small web tools, and turning research material into notes for articles I write. So I've had reason to try most of the converters that come up when you search for one. Here's what I actually use and what I'd suggest depending on what you're doing.

This isn't a paid review. PDF2MD is mine, so I'll be upfront about where it wins and where it falls short.

What I look for

After a year of this, three things matter to me and one thing doesn't:

What matters: privacy (I don't want client contracts or unpublished drafts on someone's server), batch handling (I rarely have one file, I usually have twenty), and table fidelity (most of my source PDFs have tables and most converters destroy them).

What I stopped caring about: fancy AI summarisation. I have Claude for that. I just want clean Markdown.

The tools I use

PDF2MD — this site. Browser-only, no upload, batch with zip download. Works for native PDFs. No OCR, which is the main limitation. I use this for 80% of my conversions because it's the fastest path: open tab, drop files, done.

Marker (github.com/datalab-to/marker) — a Python library. The output quality on complex layouts (academic papers, two-column reports) is the best I've seen, better than mine, honestly. It also handles math reasonably. The downside is setup: you need Python, you need to install it, and the first run downloads a model. I keep it on my desktop and reach for it when PDF2MD chokes on a tricky paper.

MinerU (github.com/opendatalab/MinerU) — similar territory to Marker, also self-hosted. I've found Marker easier to install on a Mac, but MinerU has stronger Chinese-language handling, which matters for some of my research.

Pandoc — old reliable. Not great for PDFs specifically (Pandoc's PDF input is weak), but if your source is a Word doc or HTML, Pandoc to Markdown is two seconds and the output is solid. I use it for the non-PDF half of my research collection.

Nanonets PDF to Markdown — server-side, AI-powered, has a free tier. Output is clean and they specifically pitch it for RAG pipelines. I don't use it because of the upload, but if privacy isn't a concern for you and you want one nice API, it's a reasonable pick.

What I'd pick if I were you

If you have native PDFs and want it done in 30 seconds, use PDF2MD or any browser-based tool. The convenience-to-quality ratio at this end is excellent.

If you have scanned PDFs (photographs of pages, text not selectable), you need OCR, and that means Marker, MinerU, or a paid service. The browser-only tools, mine included, can't help you here.

If you have one weird PDF that nothing else handles, throw it at all three of PDF2MD, Marker, and Nanonets and use whichever output looks least broken. After a year I still do this maybe once a week. PDF parsing is a hard problem and no tool wins on every document.

If you're doing this in a script or pipeline, skip the web tools entirely. Marker as a Python library is the right answer, and MinerU is a fine alternative.

The honest catches

A few things I had to learn the hard way:

Tables in PDFs are not really tables. They're often just text positioned in columns with no semantic structure. Every converter has to guess where the cell boundaries are. When you see "this converter destroyed my table," it's usually a PDF where the table cells have wrapped text and the converter merged the wrapped lines into different rows. There's no perfect fix, but I've found the post-conversion fix is faster than trying ten different tools.

Headings are also a guess. PDFs don't have a "this is an H2" attribute — converters infer it from font size and weight. If your PDF uses unusual typography, expect to fix the heading levels manually.

Page breaks in the original become invisible in Markdown, which is usually what you want, but sometimes you lose the "this section ends here" cue. I've taken to running a quick scroll-through after conversion to make sure no two sections got fused.

What's next

I'll keep updating this page as the tools change. The space moves quickly: a year ago none of the AI-powered options were viable, and now two of them are good enough for production. Worth checking back if you're shopping around.

If you have a PDF that PDF2MD handles badly, send it to support@pdf2md.net — I keep a folder of pathological cases and use them to test improvements.

Last updated: March 4, 2026