Editorial illustration for Token Saver Cuts Claude PDF Costs 90-99% with Local Hybrid RAG
Claude PDF Costs Drop 90-99% With Local Hybrid RAG
Upload a 200-page PDF into Claude and the price tag doesn't stop at that first exchange. Every follow-up question resends the whole document to the model again, because the conversation history gets re-transmitted on each turn. That's the mechanic behind a cost problem most users never notice until the bill arrives.
Marktechpost AI has released Token Saver, an open-source MCP extension for Claude Desktop, MIT licensed and now at version 1.0. Arnav Rai, a computer science student at Rochester Institute of Technology, built it during an internship at Marktechpost AI Media Inc, with Jean-marc Mommessin and Asif Razzaq supervising. The tool runs a Local Hybrid RAG system directly on a user's machine, letting Claude answer questions about large PDFs without the file ever getting uploaded to the model itself. Marktechpost reports token consumption drops by 92% to 99%, no Python environment or terminal setup is needed, and the document never leaves the local machine.
The bigger issue is what Claude actually does with a PDF once it's dropped into a chat, and why that default behavior is what makes the token bill balloon in the first place.
Marktechpost’s Token Saver acts as a local MCP server : a lightweight background program on your machine that Claude can call as a tool. The PDF never leaves your hard drive.
Why this matters For anyone building on top of Claude, the math in this release is the real story. Re-sending a 200-page PDF on every turn was never sustainable for teams running document-heavy workflows, and Token Saver's local hybrid RAG approach attacks that cost directly rather than asking Anthropic to fix pricing. That it's open source and runs as an MCP server on your own machine matters too: developers can inspect how retrieval is happening instead of trusting a black box, and founders burning through API budgets on repeated context get a concrete lever to pull.
We'd want to see the 90-99% figure tested against real workloads before treating it as gospel. Marktechpost built this, and self-reported savings from a project's own creators deserve a second look, especially around how "hybrid" retrieval handles edge cases like tables or scanned pages. Still, the underlying problem, paying for the same document over and over in a chat history, is real and widely felt. A local, inspectable fix aimed at that specific pain point is worth watching, and worth trying on your own PDFs before you trust the numbers.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives Editor Urges Writers to 'Be More Human' as Anti-AI Literary Push Grows
From Archives Anthropic AI Finds Potential Weaknesses in NIST-Approved Cryptographic Algorithms
From Archives DeepMind AlphaFold team disbands as researchers depart for Anthropic, Isomorphic
Research & Benchmarks
From Archives Anthropic's Mythos Tool Meets Its Hype in Internal Testing
Business & Startups
Common Questions Answered
How does Token Saver reduce Claude PDF costs by 90-99%?
Token Saver uses a local hybrid RAG (Retrieval-Augmented Generation) approach that processes PDFs locally on your machine instead of resending the entire document with every follow-up question. By eliminating the need to retransmit conversation history and full PDF content on each turn, it dramatically reduces the number of tokens consumed during multi-turn interactions with Claude.
What is the main cost problem with uploading PDFs to Claude that Token Saver solves?
When users upload a PDF to Claude, every follow-up question causes the entire document and conversation history to be retransmitted to the model again, creating exponential token usage costs that most users don't notice until their bill arrives. Token Saver eliminates this problem by keeping the PDF on your local machine and only sending relevant retrieved content to Claude.
How does Token Saver work as an MCP extension for Claude Desktop?
Token Saver acts as a local MCP (Model Context Protocol) server—a lightweight background program running on your machine that Claude can call as a tool. The PDF never leaves your hard drive, and Claude communicates with the local server to retrieve only the relevant information needed for each query, rather than processing the entire document repeatedly.
Why is Token Saver's open-source approach important for developers?
As an open-source MIT-licensed tool, Token Saver allows developers to inspect exactly how the retrieval process works instead of trusting a proprietary black box solution. This transparency enables developers building document-heavy workflows to understand and customize the retrieval mechanism while maintaining full control over their data and costs.
Who created Token Saver and what is its current version status?
Token Saver was released by Marktechpost and is now at version 1.0, having been created by Arnav Rai, a computer science student at Rochester Institute of Technology. The tool is available as an open-source MCP extension for Claude Desktop under the MIT license.
Further Reading
- Introducing Free RAG for Claude Code — Save Tokens & Time - DEV Community
- Cut Claude Code Token Costs - BuildThisNow
- RAG for Claude on a Budget: Retrieval Strategies That Reduce Context Costs - Blockchain Council
- How to save millions in Claude tokens (code included) - BoringBot Substack
- token-reducer - GitHub
