How Rust Token Killer Cuts AI Coding Token Costs by 90% in Seconds

CLI Noise

When a command like git status is run in Claude Code, the full output—showing branch status, modified files, and usage hints—consumes about 2,000 tokens, even though the AI only needs roughly 200 tokens to understand the essential information.

RTK: A CLI Proxy Designed for AI Programming

RTK (Rust Token Killer) intercepts command output before it reaches the AI, removes noisy parts, and returns a concise summary.

The tool is written in Rust, compiled to a single binary with zero dependencies, adds less than 10 ms overhead, and has earned 11.9k GitHub stars under an MIT license.

Core metrics: average token savings of 89%, session length increased by about three times, and support for more than 30 common commands.

Example after RTK processing:

Branch: master...origin/master
Modified: 3 files
   index.html
   src/main.rs
   src/config.rs
Untracked: 2 files
   .fastembed_cache/
   tests/

The concise output uses roughly 200 tokens instead of 2,000.

Real‑world data shows that after 15 days of use, a developer ran 7,061 commands, consuming 29.3 M input tokens and 4.8 M output tokens, saving 24.6 M tokens (83.7%). Another case saved 138 M tokens across 15,720 commands.

Token‑saving statistics can be viewed with the rtk gain command, which generates detailed reports.

30‑Second Setup

Installation requires two steps:

# macOS and Linux (recommended)
brew install rtk
# or use the quick‑install script
curl -fsSL https://raw.githubusercontent.com/rtk-ai/rtk/refs/heads/master/install.sh | sh

Activate the global hook: rtk init --global The hook installs a PreToolUse interceptor in Claude Code’s configuration, automatically rewriting all Bash commands to their RTK equivalents.

For example, when Claude Code runs git status, the hook rewrites it to rtk git status, a transformation that is completely transparent to the AI.

After installation, restart Claude Code and continue normal “vibe coding.”

An additional useful command, rtk discover, scans Claude Code’s history to identify commands that could have been optimized with RTK, revealing missed token‑saving opportunities.

Long‑Term Value of Cost Optimization

With RTK, a typical session consumes only about 23 k tokens, allowing sessions to be roughly three times longer. Longer sessions let the AI retain more context, resulting in more accurate and helpful suggestions.