AI Pulse · Under the Radar

A four-day-old GitHub project called Token-Saver is getting quiet traction for doing something most PDF tools skip: it cuts Claude Desktop’s token use by 92 to 98% on large documents without sending anything to the cloud. The repo sits at 87 stars and is climbing at 22 a day, which is modest but steady for a utility that solves a real cost problem.

How it works

Token-Saver runs hybrid search, meaning it combines keyword matching with semantic similarity to pull only the relevant chunks from a PDF before handing them to Claude. Instead of feeding an entire 200-page contract or proposal into the context window, you get a few paragraphs and exact page citations. The tool stays local, so sensitive files never leave your machine. That matters if you’re dealing with client proposals, financial reports, or anything under NDA.

The author is Marktechpost, a known AI news publisher, not an anonymous account, which adds a layer of credibility. The code is open, the setup is straightforward, and the use case overlaps directly with workflows we see every day: querying dense PDFs without burning tokens or waiting on cloud uploads. If your team spends time summarizing RFPs, digging through technical specs, or extracting clauses from contracts, this is the kind of pattern that belongs in an AI command centre built for real work, not demos.

Token costs on large documents add up fast, especially if you’re running dozens of queries a week. A 90%+ reduction is not theoretical. It’s the difference between a $50 monthly bill and a $500 one. The project is young, so expect rough edges, but the core idea is sound and the timing is right. Worth a look if PDF-heavy work is part of your operation.