RTK โ Rust Token Killer
A single Rust binary that cuts LLM token usage by 60โ90% by filtering CLI output before it hits the context window.
Created Jun 28, 2026 - Last updated: Jun 28, 2026
RTK โ Rust Token Killer
RTK is a high-performance CLI proxy written in Rust that sits between your shell and your AI coding assistant. Instead of dumping raw command output into the LLM context, RTK compresses it first โ stripping boilerplate, deduplicating repeated lines, and keeping only what matters.
The Problem It Solves
AI coding assistants read everything in the terminal. A git status on a large repo, a full cargo test run, or a verbose docker ps output can burn thousands of tokens on noise that the model doesn’t need. Over a 30-minute session this adds up fast โ one benchmark showed 118,000 tokens reduced to 23,900 for typical dev operations.
How It Works
RTK intercepts commands and applies four strategies per tool:
- Filtering โ removes known boilerplate and progress lines
- Grouping โ aggregates similar items (e.g. multiple changed files of the same type)
- Truncation โ keeps the head and tail of long outputs, cuts the middle
- Deduplication โ collapses repeated log lines with a count
It ships with handlers for 100+ tools: git, GitHub CLI, cargo, pytest, docker, AWS CLI, kubectl, and most common linters and package managers.
Integration
RTK works as a transparent hook in Claude Code, Cursor, Copilot, Cline, Gemini CLI, and others. The hook rewrites bash commands automatically โ git status becomes rtk git status without any change to your workflow.
# Install via Homebrew
brew install rtk
# Check savings
rtk gain
# Use directly or let the hook handle it
rtk cargo test
Why Rust?
The binary needs to be fast and dependency-free โ it runs on every command invocation. Rust gives both, and the resulting binary is cross-platform (macOS, Linux, Windows via WSL).
Bottom Line
RTK is a thin, invisible layer that makes AI-assisted development cheaper and faster by doing one thing well: reducing the token tax of standard dev tooling. The savings compound quickly across a full coding session.
Links: GitHub ยท Apache 2.0