Token Saver: Local MCP Extension Cuts Claude PDF Reading Costs to 1/13

Forum topic · 小凯 · 2026-08-02

Summary

Token Saver is an open-source (MIT) local MCP extension that dramatically reduces the token cost of reading large PDFs in Claude Desktop. Instead of sending entire documents to the model, it indexes PDFs locally and serves only relevant excerpts (capped at 8,000 characters) with page citations via the MCP protocol. It uses hybrid RAG combining BM25 keyword search (weight 0.4, on SQLite FTS5) and local all-MiniLM-L6-v2 embeddings (weight 0.6), with a quality gate requiring 0.25 similarity threshold for semantic-only hits. Real-world tests show savings of 95.7-99.4%: a 233-page legal document dropped from 133,349 tokens to just 740 per query. Beyond cost savings, the tool offers a compliance advantage - PDFs never leave the user's machine (folder allowlist, zero uploads). Notable limitations include needing OCR for scanned PDFs, an ~80MB initial model download, and occasional Windows sentence-splitter errors. The project signals a broader trend of MCP tools moving from generic utilities to domain-specific solutions that make expensive workflows affordable.

Category: ai-products · AI coding infrastructure Released: v1.0 on 2026-07-30 Sources: Official Marktechpost GitHub repository (MIT license), kiadev.net hands-on testing, techsparking secondary coverage

What It Does

Before feeding a large PDF into Claude Desktop, Token Saver indexes it locally. When you ask a question, Claude calls this local tool via the MCP protocol and receives only "the most relevant passages + page numbers" for the current query. The entire PDF never leaves your hard drive - the model only sees precise snippets trimmed to 8,000 characters.

The underlying approach is hybrid RAG: BM25 (keyword, weight 0.4, running on SQLite FTS5) + local all-MiniLM-L6-v2 embeddings (semantic, weight 0.6). After both scoring paths, results pass a "quality gate" - snippets hit by semantics alone must exceed a 0.25 similarity threshold to be included.

The Numbers

Measured across three documents (estimated with cl100k_base):

| Document | Pages | Full-doc tokens | Tool-returned tokens | Savings | |---|---:|---:|---:|---:| | FDA drug label | 33 | 23,959 | 1,021 | 95.7% | | GDPR full text | 88 | 70,260 | 996 | 98.6% | | SFFA v. Harvard | 233 | 133,349 | 740 | 99.4% |

Feed a 1,000-page textbook into Claude, and you pay for the whole thing on every conversational turn. With Token Saver, you pay only a few hundred tokens of precise excerpts per turn. The bigger the document, the greater the savings - the ratio is not linear.

Why This Is More Than a Token-Saving Utility

1. The cost model changed. Claude bills by "context tokens + context resending per turn." Token Saver doesn't cut the cost of a single query - it cuts the cost structure of "long documents + multi-turn conversations." A reading session over a 200-page PDF that might cost $5-10 could run about $0.50 with Token Saver.

2. The MCP ecosystem is sinking from "generic tools" to "domain tools." Token Saver isn't general-purpose; it does one thing only: PDF reading. MIT-licensed, 17 commits, built by Arnav Rai, an undergraduate intern at Rochester Institute of Technology, during his internship at Marktechpost. This shows individual developers can already write MCP tools and ship them to production.

3. Local-first compliance value. Legal, medical, and financial PDFs are nearly impossible to upload to third-party services. Token Saver's design - folder allowlist + zero uploads + only metadata sent to Claude - is a "compliance first, features second" fit for these workflows.

4. Retrieval evaluation is open-sourced. The repo includes eval/retrieval_eval.py and RESULTS.md, with measured data on recall@5 and false-abstain rates. Marktechpost treats this RAG tool as "an evaluable product," not a demo.

Limitations

Embedding model choice: all-MiniLM-L6-v2 requires an ~80MB download on first use. If the download fails, it degrades gracefully to pure BM25 - retrieval quality drops but it still works.
Not all PDFs qualify: scanned or image-only PDFs need OCR first; if pypdfium2 extraction fails, pypdf serves as a fallback.
On Windows, the Stanford NLP sentence splitter occasionally errors - documented in known-issues by the author.
The performance ceiling is Claude itself: no matter how good the returned snippets are, the model's reasoning is bounded by the 8,000-character input constraint.

How to Think About This Category of Tools

Token Saver's real significance isn't "saving 99% on costs" - it's turning large-document Q&A from a luxury into a daily utility. Teams doing enterprise knowledge bases, document review, or compliance analysis used to spend $200 having a newcomer "read and understand GDPR" before work could begin; now the same job can be done for $15.

This is the same story AI coding tools have been telling, in a different setting: wrap a single-point capability into a pluggable MCP tool, let the model decide when to invoke it. The model weights themselves don't upgrade, but a hole gets punched through the workflow.

Original links:

https://github.com/Marktechpost/Token-Saver
https://kiadev.net/news/2026-07-30-token-saver-local-hybrid-rag
https://techsparking.com/meet-the-token-saver-an-open-source-mcp-extension-using-local-hybrid-rag-to-reduce-token-costs-claude-pdf-90-99