Marktechpost has released Token Saver , an open-source Claude Desktop extension designed to reduce the number of tokens used when working with large PDF documents.
The MIT-licensed tool, currently at version 1.0, uses local retrieval to search PDFs on a user’s computer and sends Claude only the passages relevant to a question instead of the entire document. Marktechpost reports token reductions of 92% to 99%, depending on document size and usage.
Token Saver was developed at Marktechpost AI Media Inc. by Arnav Rai, a computer science student at Rochester Institute of Technology, during his internship under the supervision of Jean-Marc Mommessin and Asif Razzaq.
The tool targets a problem that appears when users repeatedly question AI models about large documents.
According to Marktechpost, adding a 200-page PDF to a conversation does not create only a one-time context cost. The document remains part of the conversation context and can be processed again during later turns, increasing token usage as the discussion continues.
Marktechpost says PDF processing can also involve both extracted text and page images to preserve elements such as charts and layouts. The text alone can reach roughly 1,500 to 3,000 tokens per page, according to its report.
Prompt caching and Claude Projects can reduce some of the overhead, but Marktechpost argues that they do not remove the need to send the document to the provider. It also says asking an AI model to search through a 1,000-page document when only a few paragraphs are relevant can waste context and increase the chance of incorrect answers.
Token Saver runs as a local Model Context Protocol (MCP) server, allowing Claude Desktop to use it as a document-search tool.
The PDF itself stays on the user’s computer. When a question is asked, Token Saver searches the document locally and sends Claude only the relevant sections.
Its Hybrid RAG system combines two retrieval methods.
BM25 keyword search uses SQLite FTS5 to find passages containing relevant terms and receives a weighting of 0.4.
Semantic search uses the local all-MiniLM-L6-v2 embedding model and cosine similarity to find passages based on meaning, with a weighting of 0.6.
The embedding component is optional. If the model cannot load, Token Saver falls back to keyword-based retrieval.
Marktechpost describes an eight-stage process that runs locally before information reaches Claude.
The approach means Claude receives a smaller collection of page-cited passages rather than the complete PDF.
Marktechpost benchmarked Token Saver on three documents and measured token counts using tiktoken .
Marktechpost notes that the figures use cl100k_base as a proxy tokenizer and assume the entire document would otherwise be charged during every search.
The project documentation also makes clear that savings depend on document size and the comparison baseline. Its current GitHub documentation says the benefit grows as documents become larger because the amount of retrieved text remains relatively small.
Token Saver is also designed around local document privacy.
Marktechpost says the original PDF is never uploaded through Token Saver. Extraction and retrieval happen on the user’s machine, with only selected passages reaching Claude.
Users must also choose an allowed folder. Token Saver refuses to access documents outside that directory.
Communication with Claude Desktop takes place over standard input/output, or stdio, rather than through a listening network port.
The project recommends using a small dedicated folder containing only documents the user intends Claude to access.
Marktechpost says Token Saver works with Claude models but found differences during testing.
Claude 3.5 Sonnet was recommended as the general option. It handled less precise file references, asked users to clarify when filenames were similar, and correctly used retrieved page citations.
Claude 3 Opus performed better with complicated documents containing different voices, such as legal judgments with majority and dissenting opinions. It could also indicate when an attribution was inferred from context rather than directly stated.
Token Saver is distributed as a single .mcpb bundle, avoiding the Python environments and terminal setup often required by local AI tools.
Users download token-saver-ccr.mcpb , open Claude Desktop → Settings → Extensions → Install extension, enable it, and select the folder containing their PDFs.
Claude asks for permission when the tool is first used, after which users can allow continued access.
The extension does not operate until an allowed folder has been selected. That folder defines what Token Saver can read, lets users refer to documents by filename instead of full paths, and provides Claude with the available document list.
Marktechpost positions the tool for people repeatedly working with large material such as legal judgments, technical standards, financial reports, and textbooks.
Get the latest tech news, telecom insights, and product launches wherever you prefer.
Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.
Technology and Automotive Specialist covering the latest cars, smartphones, AI breakthroughs, and...
Shares
