What rtk's estimate table says, vs. what Atelier measured.
rtk is a Rust CLI proxy that hooks Bash tool calls in Claude Code (and 14 other
agents) and rewrites commands like git status, cat, and
pytest to compact equivalents before the output reaches the model.
It doesn't search or retrieve code at all -- a different category from the 10
tools in the MRR table -- so this page compares it on the one pillar it actually
overlaps with Atelier: compact command output.
“CLI proxy that reduces LLM token consumption by 60-90% on common dev commands.”
“Estimates based on medium-sized TypeScript/Rust projects. Actual savings vary by project size.”
| Operation | Standard (est.) | rtk (est.) | Savings |
|---|---|---|---|
| ls / tree | 2,000 | 400 | -80% |
| cat / read | 40,000 | 12,000 | -70% |
| grep / rg | 16,000 | 3,200 | -80% |
| git status | 3,000 | 600 | -80% |
| cargo test / npm test | 25,000 | 2,500 | -90% |
| Total (12 operation types) | ~118,000 | ~23,900 | -80% |
5 of 12 published rows shown -- full table at the source link above.
| Method | Accuracy control | Result |
|---|---|---|
| rtk: per-command estimate, one stated project profile | None -- no task run, no correctness checked | -80% avg (self-reported) |
| ★ Atelier: Terminal-Bench 2.1, 89 real agentic tasks | 78.7% vs. 78.9% expected (-0.2pp, held flat) | -28.1% cost, real tasks |
rtk's table is real and useful, but it's an estimate on one stated project
shape, not a run against an actual coding task -- rtk has never published
whether compressing output this much changes whether the agent solves
anything. That's the exact question Terminal-Bench answers for Atelier:
cost drops by a similar order of magnitude (28.1%), but with accuracy
checked against a public leaderboard, not assumed. Worth noting from rtk's
own README too: the rewrite hook only fires on Bash tool calls -- Claude
Code's built-in Read/Grep/Glob bypass it entirely unless a user calls
rtk read/rtk grep by hand.
The true story
rtk is the most-starred tool anywhere in this comparison, and its numbers are real -- just measured a different way: an estimate table instead of a task benchmark. Same underlying idea Atelier bets on (verbose command output wastes tokens), checked against an actual leaderboard instead of a stated project profile.