jevgrep's Own Numbers Are the Honest Kind: 8 of 10 Tasks, and a Smaller Saving Than the Headline

jevgrep is an open-source, MIT-licensed command line tool that helps coding agents find relevant code by asking what it does. A model judges relevance across a repository, and the calling agent then does the implementation. Its README claims about 30 percent lower cost.

Our take: read the small print, because the author does. In a ten-task SWE-bench comparison, jevgrep and the baseline each solved 8 of 10. The 30 percent figure covers one model's cost only, falling from $7.62 to $5.44. A later rerun that includes the cost of the search model measured 25.8 percent lower, with the same 8 out of 10 solved.

The author also says these are single-run observations, not statistical proof, and that results may not hold for every repository. That candour is rarer than it should be in tooling claims.

One practical warning: eligible source content is sent to the model provider. Default filters skip hidden, dependency, build, binary and obvious credential files, but the README says that is not a guarantee against leaking sensitive data. Run its files command first to see what a search would read.

Before adopting any agent tool on client code, measure it on your own repositories and check where the code goes.