jevgrep: ask where a behavior lives

jevgrep lets your coding agent ask 'where does this behavior live?' and get back relevant source, with a claim of about 30% lower cost.

Coding agents spend part of every unfamiliar task just finding the right files. jevgrep (MIT, about 2.8k stars, created last month) is a CLI that takes a question about a repository and returns relevant files, reading leads and verbatim excerpts in one response. Under the hood it uses Jev, a model that judges relevance across folders, files and declarations, and your agent does the implementing.

The README's measurement: on a ten-task SWE-bench comparison, jevgrep and the baseline both solved 8 of 10 tasks. Full cost fell from $7.62 to $5.44, a 28.6% reduction that the README rounds to about 30%, counting failed attempts and not counting the cost of Jev itself.

Read that carefully. Ten tasks is a small sample, and Jev has its own cost that isn't in the headline. The README also says outright that when you know the exact symbol or path, a direct read or ripgrep is all you need. It's for questions that span unfamiliar files.

It needs Node 22+ and a key for a provider such as Vercel AI Gateway or OpenRouter. Install with npm install -g @dzhng/jevgrep, run jg auth to add your key, then jg skill to teach your agent to use it.

What does your agent do first on a repo it's never seen: search, or ask?