RTK cut git status output 91.88% on our fixture, and ls -la only 68.62%

RTK is a single Rust binary that sits in front of your agent's shell commands and returns a shorter version of their output. Its README says it "cuts up to 90% of the bash output your agent reads," and in the same breath warns that this is different from cutting your bill by 90%. We installed rtk v0.52.0 with no keys and measured two commands on a 30-file git fixture. git status dropped 91.88%, ls -la dropped 68.62%, and every token figure in this post is an estimate of bytes divided by four.

The catch

Keep two things apart. The 90% figure is RTK's claim about bash output. The table below is our measurement of two commands on one fixture. Neither is a billing number. rtk counts tokens as bytes/4, and so does our report, which labels every token value "estimate (bytes/4), measured on fixture". We ran no tokenizer.

The savings depend on the command. A short git status compresses well. ls -la kept almost a third of its bytes. On Linux the ls -la byte counts shift with dates and file owners, so the tests assert exact rtk git status and rtk ls lines and only check that rtk's git status output is smaller than the raw one.

Agent setup has a trap too. At 0.52.0, rtk init --agent cursor prints that Cursor hooks are global-only and writes nothing. The global variant writes under your home directory, so we left it out of tests and CI. The project-scope rtk init for Claude Code writes CLAUDE.md and .rtk/filters.toml in the project, and our tests confirm HOME stays empty.

What we ran (keyless)

export RTK_TELEMETRY_DISABLED=1 DO_NOT_TRACK=1 DISABLE_TELEMETRY=1
bash scripts/install-rtk.sh
cargo test --all
bash scripts/report.sh        # writes reports/token-diet.json

The install script downloads the release tarball and checksums.txt, verifies the SHA256 against the file and a hard-coded value, and only then installs bin/rtk. The report builds a fresh fixture (30 files, one dirty file1.txt, branch master) and runs each command raw and through rtk.

From reports/token-diet.json, run on October 14, 2026:

To measure your own tree instead of the fixture:

cargo run -- --rtk bin/rtk --dir /path/to/repo --json --report reports/token-diet.json -- git status

The token-diet binary only runs allowlisted ls and git commands and exits with code 2 on anything else.

How it is wired

rtk sits between the agent and the shell. The agent calls rtk git status, rtk runs git status, and the model reads the compact result. The project-scope rtk init writes the instructions Claude Code reads, and the lab skill read-less-before-deciding works in both Claude Code and Cursor from .claude/skills/. Telemetry stays off with RTK_TELEMETRY_DISABLED=1, which rtk telemetry status reports as blocked.

Minilab (open source)

Upstream is rtk-ai/rtk: 83,376 stars, Apache-2.0, Rust, as of October 14, 2026. Clone the lab, run bash scripts/report.sh, and compare your reports/token-diet.json against the table above before you quote any savings.

bash scripts/install-rtk.sh && cargo test --all && bash scripts/report.sh