rtk: The Essential CLI Tool That Reduces LLM Tokens by 90%

The Hidden Cost of AI Coding Agents: CLI Output Logs

When using AI coding assistants like Claude Code or Gemini CLI, we often let the agent execute terminal commands. However, when the raw outputs of commands like git status, npm test, or next build enter the LLM's context window directly, it results in a massive waste of tokens.

Agent8's dev partner Kai analyzed an open-source tool that drastically solves this problem: rtk (Rust Token Killer).

What is rtk?

rtk-ai/rtk is a high-performance CLI proxy written in Rust. It intercepts command outputs before they reach the LLM, filtering out noise and compressing only the essential information.

4 Core Compression Strategies of rtk

1. Smart Filtering: Removes noise unnecessary for LLM understanding, such as comments, meaningless whitespace, and boilerplate text.
2. Grouping: Condenses output by grouping files by directory or aggregating logs by error type.
3. Truncation: Cuts off unnecessarily repeating or redundant context, leaving only the core details.
4. Deduplication: Detects identical log lines repeating hundreds of times and abbreviates them by only showing the "repetition count".

How Much Does It Save? (Use Cases)

It demonstrates a 60-90% token reduction effect across common development commands.

• cargo test, npm test: Hides passing test logs and delivers only failed cases (approx. 90% savings)
• git status, git diff: Removes unnecessary Git guides, provides condensed diffs (approx. 80-90% savings)
• ls, cat, grep: Optimized directory trees and context cleanups (approx. 80% savings)

Consequently, the Claude --git status--> shell structure shifts to Claude --git status--> RTK --> shell, dramatically reducing what was previously a 2,000 token response to a mere 200 tokens.

Applicability in the Agent8 System

Agent8's Agent 8 Agent Architecture already excellently controls token lengths via built-in tools in our VS Code extension (Agent 8), using parameters like OutputCharacterCount.

However, if you parallelly run workflows operating Claude Code or Gemini CLI directly from the terminal, we strongly recommend adopting rtk via brew install rtk-ai/tap/rtk. It will not only improve the AI's response time but also dramatically lower your API costs.