michael-berardi

Ultracompress — Development skill for Claude Code

Development community

Deterministic local compaction for Pi, Claude Code and Hermes Agent: VCC briefs, snap frames and lossless recall.

How to install Ultracompress

This entry records only its repository, not the path inside it, so there is no exact command to give. Open michael-berardi/ultracompress and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Ultracompress does

Deterministic local compaction for Pi, Claude Code and Hermes Agent: VCC briefs, snap frames and lossless recall.

Alternatives in Development

  • Deck — Generate slide deck presentations from briefs or research 2.8k ★
  • Setup Video Vision — Interactive setup wizard for claude-video-vision — configure backend, whisper, frames, and verify dependencies 1.3k ★
  • Lacp — Local-first policy, evidence, memory, and recovery controls for Claude, Codex, Hermes, and other CLI coding ag 300 ★

README

UltraCompress

Deterministic context compaction. Local, searchable recall. No LLM in the loop.

**Structured briefs instead of model-written summaries. Every condensed byte still searchable. $0 per compaction.**

*The compaction layer for people who read their agent's bills.*

[![Latest release](https://img.shields.io/github/v/release/michael-berardi/ultracompress?label=release)](https://github.com/michael-berardi/ultracompress/releases/latest) [![MIT License](https://img.shields.io/github/license/michael-berardi/ultracompress)](LICENSE) ![Rust](https://img.shields.io/badge/built%20with-Rust-orange) ![Claude Code plugin](https://img.shields.io/badge/Claude%20Code-plugin-D97757)

[Install](#install) · [Claude Code](#claude-code) · [Hermes Agent](#hermes-agent) · [How it works](#representations-and-savings) · [Recall](#history-and-evidence) · [Measurements](#measurements) · [Security](#related-work-and-security)


UltraCompress is content-aware context compaction for LLM coding agents. It builds a structured conversation brief without an LLM request, replaces bulky tool output with image frames when supported, and keeps the raw session on disk so a recall tool can still find anything that was condensed. It ships as a Rust CLI plus a [Pi coding agent](https://pi.dev) adapter, a standalone [Claude Code](#claude-code) plugin and a [Hermes Agent](#hermes-agent) context engine, all served by the same engine.

Summary generation by another model call is a strange way to save model calls. UltraCompress skips the middleman: compaction runs locally in roughly 10–300 ms, deterministically, at zero API cost. In Claude Code that means about 65 ms where the built-in summarizer took 23–52 seconds ([measured](#against-claude-codes-own-compaction)). What it drops stays recoverable, ranked recall measured at **94.4% hit@5** on its published fixture ([docs/BENCHMARKS.md](docs/BENCHMARKS.md)). Context management should not require another context-