Marktechpost AI’s Token Saver delivers page-cited document snippets directly to Claude Desktop, eliminating repetitive token charges and keeping sensitive files entirely on the machine.
On 30 July 2026, Marktechpost AI Media Inc. released Token Saver, an open source Model Context Protocol (MCP) extension designed for Claude Desktop that reduces PDF token consumption by 90% to 99% when analysing large documents. The project was developed by Arnav Rai, under the supervision of Jean-marc Mommessin and Asif Razzaq, and is distributed under the permissive MIT License v1.0.
Instead of uploading entire PDF files into Claude Desktop’s context window, which repeatedly recharges users for full document tokens on every conversation turn, Marktechpost AI’s Token Saver operates as a local MCP server running on the user’s machine. It utilises a Local Hybrid Retrieval-Augmented Generation (RAG) pipeline combining keyword search (BM25) with semantic search via a local all-MiniLM-L6-v2 embedding model (weighted at 0.6).
The retrieved passage chunks pass through a quality gate that enforces a 0.25 semantic similarity threshold for non-keyword matches, followed by a deduplication step to eliminate redundant snippets. Token Saver attaches exact page numbers to every retrieved text chunk delivered to Claude, enabling users to verify generated claims against the local source document.
The security and privacy model features Zero Uploads, ensuring document text remains stored locally on the user’s hard drive and is never uploaded to external AI servers. Communication between the extension and Claude Desktop occurs strictly via local standard input/output (stdio). Users configure a dedicated reference folder, creating a security boundary that controls which files the extension can access and list to Claude.
While compatible across Claude 3.5 models like Sonnet and Opus, tests showed optimal defaults with Claude 3.5 Sonnet, whereas Claude 3 Opus excelled at complex legal or multi-author documents with dissenting opinions.
