Token Saver: open-source MCP-расширение режет расход токенов Claude на PDF на 90-99%

Marktechpost released TokenSaver on July 30, 2026 — a free open-source extension of the MCP (Model Context Protocol) protocol for Claude Desktop that cuts token consumption when analyzing PDFs by 92-99% through local Hybrid RAG, without uploading the file itself to the model's servers.

Why large PDFs eat up tokens

Large PDFs are expensive in chats because the entire conversation history is resent to the model on every turn — and the document gets paid for again with every new question. According to Marktechpost, Claude by default turns each page into an image to preserve charts and layout, and separately extracts the text; the text alone takes up 1500 to 3000 tokens per page.

Prompt Caching and Claude Projects soften the blow, but don't solve the root of the problem: the entire document still crosses the boundary and ends up on the provider's servers. If you only need two paragraphs from a 1000-page textbook, forcing the LLM to comb through all 1000 pages is inefficient and raises the risk of hallucinations.

How Local Hybrid RAG works

TokenSaver runs as a local MCP server — a lightweight background program on your machine that indexes the PDF and sends the model only the relevant fragments. Instead of running the whole thousand pages through the model, hybrid search finds what's needed locally: this is both cheaper and reduces the likelihood of made-up answers.

• Release date — July 30, 2026, version v1.0, MIT license
• Token spend reduction — 92-99%
• Developer — Arnav Rai, a CS student at Rochester Institute of Technology, during an internship at Marktechpost AI Media Inc
• Project supervisors — Jean-marc Mommessin and Asif Razzaq
• Setup requires neither Python environments nor terminal work

The file itself never leaves the device, so privacy is guaranteed at the architecture level, not by promises. The extension is distributed under the MIT license and is already available as version v1.0.

"Token spend is cut by 92-99%, privacy is guaranteed, and setup requires exactly zero

Python environments and terminal configurations," — from the Marktechpost AI Media announcement.

What this means

Local RAG is becoming a practical way to work with large documents through cloud LLMs, without overpaying for tokens on every follow-up question and without sending source files out. For developers and researchers who regularly parse large PDFs, this is a direct saving — both on the model bill and on data privacy.

Frequently Asked Questions

What is TokenSaver?

TokenSaver is an open-source MCP extension for Claude Desktop under the MIT license (version v1.0). It processes PDFs locally and sends the model only the fragments it needs, cutting token spend by 92-99%.

Do you need Python to install it?

No. According to Marktechpost, setup requires neither Python environments nor terminal configuration — the extension runs as a background program on the machine.

How much does TokenSaver cost?

The tool is free and open: it is distributed under the MIT license, and the PDFs themselves are processed on the device without being uploaded to the provider's servers.