A Rust proxy that trims 80% of your LLM token bloat

What it does

rtk sits between your AI coding agent and the shell, filtering and compressing command output before it lands in the LLM context window. It is a single Rust binary that rewrites verbose terminal noise—directory listings, test logs, and version-control diffs—into compact summaries. The README claims this cuts token consumption by 60–90% across more than a hundred supported commands with under 10 ms of overhead.

The interesting bit

The hook-based integration is the main event: it transparently intercepts shell tool calls from agents like Claude Code and rewrites them so the LLM never sees the raw output. For everything else, it applies four strategies—smart filtering, grouping, truncation, and deduplication—tailored per command type rather than using a generic text compressor.

Key highlights

• Claims 60–90% token reduction on common dev workflows; a sample 30-minute session drops from roughly 118,000 tokens to about 23,900.
• Supports 100+ commands spanning version control, test runners, AWS CLI, containers, and package managers.
• Single static Rust binary with zero runtime dependencies.
• Includes built-in analytics to track token savings over time.
• Offers an ultra-compact mode for even more aggressive output shrinking.

Caveats

• The shell hook does not catch built-in agent tools like Claude Code’s Read, Grep, or Glob; you must fall back to shell equivalents or explicit rtk calls.
• On Windows, the full hook system is described as working best under WSL; the native experience is more limited.

Verdict

Worth a look if you pay per token or constantly hit context limits with agentic coding tools. If your LLM usage is already cheap or you rely heavily on built-in agent tools that bypass the shell, the friction may outweigh the savings.