Real-world RTK test: 65.7% token savings on a Rust project

[LAIguapi]

I tested RTK on a real Rust project to see how much it actually saves on Claude Code token costs.

Setup

• Environment: Linux x86_64
• RTK version: v0.42.3
• Test project: RTK's own codebase (64 modules, ~17K lines of Rust)

Key findings

Command	Raw output	RTK output	Savings
rtk ls src	563 bytes	80 bytes	85.8%
rtk read filter.rs -l aggressive	17,935 bytes	3,932 bytes	78%
rtk env	~50K bytes	~150 bytes	99.7%
rtk git status	~100 bytes	~50 bytes	48.1%
rtk git log -n 5	56 bytes	86 bytes	-53% (negative!)

Overall: 14 commands, 65.7% average savings

Honest take

The 60-90% savings claim on the README is aspirational. In practice:

• High-output commands (read, env, test runners): 80-99% savings, very real
• Low-output commands (git log with few commits): can actually add overhead
• Blended real-world savings: closer to 60-70%

The aggressive read filter is the killer feature. It strips function bodies and keeps only signatures/types, perfect for AI architecture exploration.

The gain dashboard is great for making savings tangible. Seeing "10.5K tokens saved" after 14 commands makes the value concrete.

Minor issues

• rtk grep -t rs fails. Need to use -t rust (different from ripgrep convention)
• rtk json can't parse TOML files (expected, but confusing)
• Hook-not-installed warning on every command is noisy during evaluation

---

Tested as part of a weekly deep-dive series on open-source developer tools.