Token-Saver
A local Claude Desktop extension to query large PDFs with 92–98% fewer tokens. Performs local hybrid search, cites exact page numbers, and keeps your documents private on your machine
What is it?
What it is
Token Saver is a one‑click Claude Desktop extension that performs local hybrid search on PDFs, retrieves only the relevant passages with page citations, and sends those to the model, keeping documents private on the user’s machine.
Why it exists
It was created to cut token usage (and cost) by 92–98% and improve answer accuracy by providing the model with a concise, page‑cited slice of the document instead of the entire file, while ensuring privacy and ease of use.
Who should use it
Claude Desktop users who want to query large PDF documents efficiently, reduce LLM token costs, and maintain document privacy through local hybrid RAG.
Who should avoid it
Users with very small documents (under 15-20 pages) where the overhead of retrieval might exceed the cost of pasting, or users who cannot install third-party Claude Desktop extensions.
How it works
A quick walkthrough in plain English
How Token-Saver works
Step 1 of 3
You interact with it
Open Token-Saver, send a request, or connect it to your stack.
Features
Advantages
- Massive reduction in API costs by minimizing context window usage
- Higher accuracy due to reduced irrelevant context in model reasoning
- Full document privacy — no data leaves your machine
- Simple installation: just download and enable the .mcpb file
- Citations include exact page numbers for verifiable answers
- Works well even with ambiguous queries via self-correction (especially Sonnet/Opus)
- Retrieval quality remains consistent across document sizes
- Transparent savings tracking after each interaction
- Fallback to keyword-only search ensures basic functionality without internet
- Well-documented with install guides, tests, and evaluation scripts
Disadvantages
- Less effective for documents under ~20 pages (may cost more tokens than naive paste)
- File selection can fail in folders with multiple similar documents
- Generic queries (e.g., 'the textbook') may resolve to wrong file without context
- Page-only citations lack section/provenance info (e.g., majority vs. dissent)
- Abstain gate relies on keyword presence, risking false accepts on off-topic passages
- First run requires 2–5 minutes to download and initialize the search engine
- Haiku model may need extra correction turns for ambiguous requests
- Extension shows 'not verified by Anthropic' warning (expected for local install)
- Requires manual folder configuration to scope document access
- No built-in support for document formats beyond PDF
Installation
native
Download token-saver-ccr.mcpb from Releases, install via Claude Desktop extensions, configure folder, enable extension.
FAQ
How do I install Token Saver?
Download the `token-saver-ccr.mcpb` file from the Releases page, then install it via Claude Desktop's Extensions settings by selecting 'Install extension' and choosing the file.
How does Token Saver reduce token usage?
It performs local hybrid search to extract only relevant passages from your PDFs, reducing the context window size by 92–98% compared to pasting entire documents.
Is my data kept private?
Yes. All document processing happens locally on your machine. No data is sent to external servers after the initial setup.
Which Claude models are supported?
Token Saver works with Opus, Sonnet, and Haiku. Sonnet or Opus is recommended for best results.
Can I query multiple PDFs?
Yes. During setup, configure a dedicated folder containing all your PDFs. The extension will search across all files in that folder.
What if Token Saver can't find the right document?
Ensure you've selected the correct folder in the Configure step. If it misses, try rephrasing your question or use 'list my documents' to verify available files.
Featured in Videos
YouTube tutorials and walkthroughs for Token-Saver
Alternatives
Similar projects ranked by category, topics, and text overlap.
