79k stars, one big claim

RTK filters terminal output before an AI agent reads it. One X post claiming it cuts Claude Code tokens by 60% hit 313,000 views.

The README says otherwise

RTK's own docs draw the line the social posts erase. JetBrains ran SkillsBench and found no savings at all.

“RTK cuts up to 90% of the bash output your agent reads. […] it is not the same as cutting your bill by 90%.”

What RTK actually does

It rewrites Git, test, package and file commands into terser output. An ls -la loses the owner and date, keeping names, sizes and permission bits.

The test setup

Claude Code with Fable 5.0 and OpenCode with DeepSeek V4 Pro, five runs each way per task on Terminal-Bench 2.1. Final count: 1,740 attempts, over $1,500 in tokens.

A 5% swing each way

Total spend fell 5% for Fable and rose 5% for DeepSeek. Pass rates dipped slightly with RTK, by 1% and 2%.

Weight every task equally

One expensive task can drown out dozens of cheap ones. Averaging per-task changes, Fable was 1% more expensive and DeepSeek rose 17%.

One task carried it all

Almost all of Fable's savings came from winning-avg-corewars, where RTK finished in half the turns. Across every other task, savings were under 1%.

The savings counter lies

RTK's gain metric is raw bytes minus filtered bytes, divided by four. It reported 349 million tokens saved across DeepSeek runs, an 89% cut, with no cheaper tasks to show for it.

Credit for output never asked for

In train-fasttext the agent ran head -1 twice. RTK compared those two one-line reads against the whole file and claimed 120.5 million tokens saved each time.

“Those two calls accounted for 69% of the comparison's savings counter”

A rewrite loop that cost 9x

A find flag RTK 0.45.0 did not support got rewritten into a command that refused to run. The agent hit 339 consecutive errors in 12 minutes. Fixed in 0.46.0, after these runs.

Terminal output is a sliver

Without RTK, terminal output was 7% of Fable's input tokens and 26% of DeepSeek's. Read, Grep and Glob are separate tools that bypass RTK entirely.

Caching already did the work

After each turn the context is cached, so later reads cost a tenth of normal input for Fable and a thirtieth for DeepSeek. Cache reads were 94% of input but 30% of the bill.

More turns eat the savings

DeepSeek turns held 7% less input each, but there were 18% more of them. One extra turn can cost more than all the compression saved.