Token Saver: open-source MCP cuts Claude PDF token costs 92-99%
A local hybrid-RAG server keeps your PDFs off the cloud and off your token bill — plus Google's Lyria 3.5 section-edits songs and Pangram 4 hunts AI text.
Copy markdown
How it claws back 92-99% of your tokens
Token Saver runs as a local MCP server that indexes your PDFs with hybrid retrieval (BM25 keyword search plus a local MiniLM embedding model) and hands Claude only the ~8K-character relevant slice instead of the whole document. The maker's own benchmarks: 95.7% fewer tokens on a 33-page FDA label, 98.6% on the 88-page GDPR text, and 99.4% on a 233-page legal ruling.
MIT-licensed, no Python, nothing leaves your machine
Install is a single .mcpb bundle dropped into Claude Desktop settings: point it at a folder, authorize access, done. It's recommended on Claude 3.5 Sonnet. Weekend build: aim it at your contract, spec, or research-paper pile and stop re-uploading the same PDFs every session.
Google's Lyria 3.5 adds 'Selective Section Painting'
Google's new music model lets you rewrite one part of a track — a chorus, a verse — without regenerating the whole song, and turn a short melody into a full arrangement, with per-element tempo and duration control across 30-second-to-3-minute tracks. It's live in the Flow Music app now; no developer API or pricing announced yet.
Pangram 4 claims 99.66% AI-text detection
Pangram's new detector says it separates AI-polished writing from fully-generated text, resists 'humanizer' evasion, and now covers images too, priced around $0.05 per 100 words. If you gate user submissions, moderate content, or screen freelance work, it's a usable API today — treat the headline accuracy figure as the vendor's own.