Token Saver: Local Hybrid RAG cuts Claude PDF token costs by up to 99% without uploading documents
Published by AINave Editorial • Reviewed by Ramit
If you regularly feed large PDFs into Claude, you know the hidden cost: every follow-up question re-sends the entire document, compounding token charges. Token Saver, an open-source Model Context Protocol (MCP) extension for Claude Desktop, tackles this by running a Local Hybrid RAG pipeline on your machine. It processes PDFs locally and returns only the relevant passages, cutting token consumption by 92% to 99% while keeping your files private.
What happened
Marktechpost AI Media Inc. released Token Saver v1.0 under an MIT license. Developed by Arnav Rai (Rochester Institute of Technology) during his internship, supervised by Jean-marc Mommessin and Asif Razzaq, the extension acts as a standalone local MCP server that Claude can invoke as a tool.
When you ask a question about a PDF, the server runs an eight-step Hybrid RAG pipeline locally. It combines semantic search (via an optional embedding model) with keyword matching to find the most relevant passages. Claude then synthesizes the answer, citing exact page numbers and showing a running tally of tokens saved. If the embedding model fails to load, the system gracefully falls back to keyword-only matching.
Token Saver requires selecting a folder as the interaction boundary. This folder defines which documents can be read, lets you ask for files by name instead of typing paths, and populates Claude with the list of accessible files at conversation start.
Why AI builders should care
For teams building document-heavy workflows, the cost of context windows is a recurring pain. Claude's default behavior converts each PDF page into an image (to preserve charts and layouts) while separately extracting text. Before image tokens are even counted, the text alone can run 1,500 to 3,000 tokens per page. Prompt Caching and Claude Projects help, but the entire document still crosses to the provider's servers.
Token Saver keeps processing local. The PDF never leaves your hard drive. For enterprise users, legal professionals, or anyone handling sensitive documents, this eliminates the data exposure risk of uploading full files to a cloud API.
The extension is packaged as a single .mcpb bundle, avoiding complex Python environments or JSON configurations. That lowers setup friction for developers who want to test or integrate it quickly.
Practical implications
Token Saver's savings grow with document size. The source article reports benchmarks (measured via tiktoken, cl100k_base) showing that for a 1,000-page document, a single question that would normally cost 2,500,000 tokens drops to roughly 5,000 tokens with the extension. That is a 99.8% reduction.
| Document size | Baseline tokens (per question) | Token Saver tokens | Savings |
|---|---|---|---|
| 100 pages | 250,000 | 5,000 | 98% |
| 500 pages | 1,250,000 | 5,000 | 99.6% |
| 1,000 pages | 2,500,000 | 5,000 | 99.8% |
Note: Baseline assumes the whole document is charged on every search. Actual savings depend on document type, embedding model availability, and local hardware.
Token Saver works across Claude 3.5 tiers, making it compatible with different deployment options. The token-savings tally gives you a real-time view of costs avoided during a session.
Caveats
This article is based on a single source. Independent benchmarks, repository status, and community validation are not yet available. Token savings will vary by document type, embedding model availability, and local hardware. The MIT license is described in the source; verify license terms and repository status before using in production.
