DensePack

DensePack saves on ENTIRE Claude Code conversations by handing your agent images of text which cost ~50% less to tokenize, compared to the actual raw text. New upgrades allow more things to be packed which makes the savings grow with longer read-heavy sessions. The benchmarks are easy to reproduce and the numbers do not lie! DensePack 1.3.2 upgrades: • Word files (doc and docx) and Bash output over ~400 characters (~170 tokens) are now converted into DensePack images and are read by the agent in a SINGLE turn when the text fits into a single image. • All text that DensePack packs into an image keeps a text copy as a fallback. • DensePack's image renderer received a major facelift. A 404KB file went from 70 sec to 17.5 sec, and small 11KB files from ~4 sec to ~1 sec • Limit for Bash output converted into images raised from 250KB to 1MB, and for files from 500KB to 1MB. • Icon added (Plugin submitted to Anthropic and currently waiting on a human review of v1.3.2) Benchmarks were also remade with Anthropic's new Evaluation tool for plugins and skills. DensePack saved 31.8% to 38.3% on average against the same tasks without DensePack. These are realistic single tasks, so they show real savings but don't give DensePack time to REALLY shine. The savings grow in heavy read conversations and longer sessions, because every packed read is cache read at the image price and keeps saving tokens on each turn after it. I have tested DensePack's accuracy by asking Claude to rebuild a text file byte identically from its DensePack image. When I started the project earlier this month, Claude recreated one of DensePack's own .md files byte identically ~1 out of 10 times, and a python file 1 out of 20 times. I asked Claude to recreate that same .md file 100 times and after days of fine tuning, it recreated that same .md file byte identical to the text, 100 times in a row. The Python file I benched before was run the same way and Claude recreated it ~73 times byte identically, and the other 27 were only off by a few spaces. There have been many improvements since 1.1. I added safeguards to allow similar functionality with and without DensePack with the main difference being that DensePack gives your agent images. You still have the text when you need to verify. That being said, this is far from perfect! Random characters like IDs with a capital I or small l can still get misread sometimes with Opus 5.5. This is why the safeguards were built and they have been working great but I still encourage verification rounds with your agent to verify that things went smoothly. While it's been super cool to see my bash output as a DensePack image and I have been able to build more, I HIGHLY recommend verifying things that are crucial to your project/session. You can simply ask Claude to read the portion of the image that you need to be 100% accurate as text from its saved text file. To my 7 stargazers, thank you for believing in this early! Please keep following the project. I'm sure DensePack will continue to improve as OCR improves. Here's how to update. Run these in your terminal: claude plugin marketplace update densepack-marketplace claude plugin update densepack@densepack-marketplace Then restart Claude Code. New to DensePack? • Never trust code from the internet blindly, and that includes mine. • Point a frontier model like Opus 5.5 at the repo first to scan it before installing. • Keep yourself safe. GitHub: https://github.com/Fabian-Galvez/DensePack Run these commands one at a time inside a Claude Code session to install DensePack! /plugin marketplace add Fabian-Galvez/DensePack /plugin install densepack@densepack-marketplace