MDReduce

MDReduce — Strip the noise. Keep the meaning.

Every file you send to an LLM carries weight it doesn't need. A screenshot includes the status bar, the battery icon, the browser chrome. A PDF repeats the same legal footer on every page. None of that helps the model answer your question, but you pay for every token of it.

MDReduce is built to clean files before they reach an AI: PDFs, Word, Excel, PowerPoint, images, and screenshots. Strip the headers, the repeated boilerplate, the HTML tags, the irrelevant pixels, and hand back a clean, LLM-ready file with the meaning intact.

Measured, not estimated

The numbers below aren't projections. They come from testing on real files:

−23% tokens on a standard 4-page PDF, just from removing repeated boilerplate Up to 3.7× reduction on PNG screenshots, from stripping UI chrome and switching from a vision model to a text model for extraction

Where things actually stand

MDReduce isn't live yet. Before building out the full product, we want to know which use case matters most to real users:

You use an LLM API and pay per token You process high volumes of documents You regularly send screenshots or scans to an AI

Whichever one resonates most with the people who show up is what gets built and shipped first. This page exists to collect that signal, not to sell a finished product.

Privacy by design, from day one

No account, no email, no cookies — that's the plan for the eventual product, and it's already true of this page.

Get involved

If any of this sounds like your workflow, tell us which use case fits you. That's what decides what we build next.

mdreduce-production.up.railway.app