# lowfat (CLI) lowfat is an open-source CLI tool written in [[Rust]] that strips noise from shell command output before it reaches an [[How Coding Agents Work|AI coding agent]], cutting the [[AI Tokenization|tokens]] spent on verbose tool output. Same category as [[RTK]]: it reshapes the *inputs* an agent reads instead of touching the prompt or the model. > Slim your command output. Strips noise, saves tokens. The premise is the same insight [[RTK]] is built on; most tokens burned in an AI coding session come from boilerplate command output, not reasoning. A `git log`, `git status`, or `docker ps` floods the [[Context Window]] with formatting and redundancy the model never needs. lowfat filters that at the source. ## How it works A small single binary with a UNIX-style, composable pipeline architecture; built-in filters chain together like pipes, and you can drop your own in. Three compression levels let you dial how aggressive it gets: - `lite` ; light trimming - `full` ; default - `ultra` ; maximum reduction It's **local-first** with no telemetry ; data stays on your machine. ## Reported savings Reduction depends on the command and level: - `git log`: 53–91% - `git status`: 62–74% - `docker ps`: 38–85% - `ls -la`: 2–87% As with [[RTK]], the payoff scales with how verbose the underlying tool is ; near-useless on already-short output, large on noisy commands inside long agent loops. ## Extensibility Custom filters can be written three ways: - a dedicated `.lf` DSL - plain shell scripts - [[Python]] (PEP 723 / `uv` inline-dependency scripts) ## Integrations Hooks into [[Claude Code]] via `.claude/settings.json` (transparent output filtering before the agent sees anything), plus [[OpenCode]], generic shell environments, and the Pi agent. ## Observability - `lowfat history` ; frequently-run commands and their potential token savings - `lowfat level <lite|full|ultra>` ; set the active compression level ## Install ```bash cargo install lowfat brew install zdk/tools/lowfat ``` Pre-built binaries are also published via GitHub Releases. Apache-2.0 licensed. ## lowfat vs rtk Both are Rust CLIs that compress command output before it hits an agent's context, both hook into [[Claude Code]] transparently, and both report similar savings on verbose commands. Differences in emphasis: - **lowfat** leans on a composable UNIX-pipe model and an explicit plugin story (`.lf` DSL, shell, [[Python]]), with three named levels - **[[RTK]]** ships per-command optimization logic and richer adoption observability (`rtk gain`, `rtk discover`, `rtk session`) across 13+ clients and 100+ commands They occupy the same niche; the choice comes down to whether you prefer lowfat's composable-filter philosophy or rtk's broader out-of-the-box command coverage. ## References - GitHub: https://github.com/zdk/lowfat ## Related - [[RTK]] - [[Claude Code]] - [[Claude Code Hooks]] - [[Context Window]] - [[AI Tokenization]] - [[How Coding Agents Work]] - [[OpenCode]] - [[Rust]] - [[Python]]