Token Saver is an open-source MCP extension that uses local hybrid RAG to reduce Claude PDF token consumption by up to 99%. It runs locally for privacy and uses an 8-stage pipeline for efficient retrieval.
Processing large PDFs with Claude Desktop has always carried a steep token cost. When a user feeds a 200-page legal brief or a dense scientific journal into the model, Claude must convert each page into an image and extract the text, burning through 1,500 to 3,000 tokens per page in the process. For heavy users relying on Claude for deep analysis, this translates directly into rapid rate limit hits, frustrating interruptions, and skyrocketing operational expenses.
Enter Token Saver, an open-source MCP extension that fundamentally changes this equation. By implementing Local Hybrid RAG—combining the precision of BM25 keyword search with the contextual understanding of semantic embeddings—entirely on the user’s machine, Token Saver slashes token consumption by 92% to 99% without compromising response quality. This is not just an incremental improvement; it is a fundamental redesign of how Claude interacts with long-form documents.
Token Saver is a Model Context Protocol (MCP) server that runs entirely on your local machine with no external dependencies. It acts as an intelligent middle layer between Claude Desktop and your PDF files, essentially giving Claude the ability to “search” and “retrieve” from a document intelligently rather than forcing it to “read” the entire file page by page. This fundamental architectural shift is what enables the massive token savings.
.mcpb bundle file directly through Claude Desktop’s extension settings, taking just seconds to deploy from zero.The secret to Token Saver’s performance lies in its sophisticated 8-stage processing pipeline. Instead of sending the entire document to Claude, the system intelligently finds and passes only the most relevant passages, stripping away the vast majority of irrelevant text.
The performance claims are backed by rigorous benchmarking across a variety of document types and sizes. The results clearly demonstrate that the system scales beautifully with document size, delivering the most impressive savings on the largest and most complex files.
These numbers are not just abstract metrics. They translate directly into real dollar savings and drastically reduced latency. Querying a 200-page document becomes as cheap and fast as querying a single page. The return on investment is essentially immediate for any professional or team regularly working with large document sets.
Token Saver directly addresses three major pain points that define the current experience for professionals using Claude in demanding environments.
API costs are often the single largest barrier to deploying AI agents at scale in the enterprise. By reducing token consumption by up to 99%, Token Saver effectively decouples your operational costs from the size of your document base. You can query massive legal archives or technical libraries repeatedly without blowing your monthly budget.
Many industries—legal, healthcare, and finance—operate under strict data residency and compliance requirements. Because Token Saver processes everything locally, sensitive information never has to leave your network perimeter. The stdio-only communication model combined with folder allowlisting provides a security posture that cloud-based RAG systems simply cannot match.
As an open-source MCP extension, Token Saver is fully transparent. Developers can audit the code, contribute improvements, and adapt the system to their specific workflows. The .mcpb bundle format makes it trivial to deploy across an entire team without individual configuration headaches, promoting consistency in setup and behavior.
Ready to start saving immediately? The setup process is remarkably simple and requires no terminal commands:
.mcpb bundle file from the project’s official release page on GitHub.The entire process takes less than a minute, and the savings begin with your very first query.
Token Saver is a powerful indicator of where the broader AI ecosystem is heading. As models become more powerful and widely adopted, the cost of serving them becomes the critical bottleneck for practical deployment. Intelligent, local retrieval pipelines like this one are not just nice-to-haves—they are becoming essential infrastructure for making AI economically viable and secure at the enterprise level. We can expect to see more projects embracing this hybrid approach, combining the precision of keyword search with the nuance of semantic understanding, all while maintaining strict local control over data.
Token Saver is more than just a clever utility; it represents a practical and necessary paradigm shift in how we interact with large language models during document analysis. By achieving 92% to 99% token reduction through a local, multi-stage Hybrid RAG pipeline, it solves the most pressing issues facing Claude power users today: cost, speed, and privacy.
Whether you are a legal researcher analyzing case law, a financial analyst reviewing regulatory filings, or a developer building the next generation of AI-powered tools, Token Saver provides the efficiency and security you need to work at scale. The tool is available now, it is completely open source, and it takes less than a minute to set up. Stop paying for tokens you do not need and start retrieving only what matters.
Token Saver is an open-source MCP extension that acts as an intelligent middle layer between Claude Desktop and your PDFs. Instead of sending the entire document page by page, it uses a local hybrid RAG pipeline—combining BM25 keyword search and semantic embeddings—to retrieve only the most relevant text sections, slashing token consumption by 92–99%.
Installation is zero-friction: you simply download a single .mcpb bundle file and import it through Claude Desktop's Extensions menu. No Python environment setup, no dependency hell—it takes just seconds to deploy and start saving tokens immediately.
Claude's native PDF support converts each page into an image and extracts text, consuming 1,500–3,000 tokens per page. Token Saver's hybrid RAG approach only sends the retrieved relevant sections, using up to 99% fewer tokens while maintaining high response quality. The key difference is moving the cost of 'reading' from token consumption to local compute.
Yes. Token Saver runs entirely on your local machine with no external API calls or data transfers. Communication with Claude happens over stdio, and folder allowlisting ensures Claude can only access directories you explicitly authorize—so sensitive legal, financial, or medical documents never leave your computer.
Token reduction typically ranges from 92% to 99% depending on document size and content relevance. The trade-off is a slight delay while the local retrieval pipeline processes the PDF—typically a few seconds—but the massive reduction in token usage means far fewer rate limit hits and much lower operational costs.
