If you've ever dropped a big PDF into Claude Desktop and then kept chatting about it, you've probably paid for that PDF more than once. A new open-source tool called Token Saver fixes that — and the numbers behind it are worth paying attention to.
The problem: your PDF never actually leaves the conversation
Most people assume uploading a PDF is a one-time cost. It isn't. Claude resends your entire conversation history — including the full document — with every single follow-up question. On top of that, Claude often converts PDF pages into images to preserve formatting, which multiplies the token cost even further. A single page can burn through 1,500 to 3,000 tokens just from text extraction, before you've asked a single real question.
Ask ten follow-ups about a 200-page document, and you've effectively paid to re-read that document ten times.
What Token Saver does differently
Token Saver is a free, open-source MCP (Model Context Protocol) extension for Claude Desktop. Instead of sending the whole PDF to Claude, it processes the document locally on your machine and only sends the specific passages relevant to your question.
It does this with a hybrid retrieval approach:
- Keyword matching (BM25) — finds exact terminology using SQLite's full-text search, weighted 40%
- Semantic search — uses a local embedding model to understand meaning, weighted 60%
The two signals get combined, then run through an 8-stage pipeline: extraction, chunking, hybrid scoring, quality filtering, deduplication, trimming, and a hard character budget before anything reaches Claude.
The numbers are the real story
The team benchmarked this on real, long documents, not toy examples:
- An FDA drug label (33 pages): 95.7% token savings
- A GDPR regulation document (88 pages): 98.6% token savings
- A 233-page Supreme Court case file (SFFA v. Harvard): 99.4% token savings
Across the board, the tool claims a 92–99% reduction in token consumption compared to just pasting the file in. That's not a marginal optimization — for anyone working with long PDFs regularly, it changes the economics of using Claude for document work entirely.
Privacy is part of the pitch, not an afterthought
Because processing happens locally, your documents never touch an external server. The extension also uses folder allowlisting, so it can only see files in directories you explicitly grant it access to, and it talks to Claude Desktop over standard I/O with no open network ports. For anyone handling legal, medical, or otherwise sensitive PDFs, that's arguably as valuable as the token savings.
Setup is deliberately low-friction: download the .mcpb bundle from GitHub, install it through Claude Desktop's settings, point it at a folder, and grant permission when prompted. No Python environment, no terminal.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.