Jevgrep lets a coding agent find files in a codebase by asking a plain question, and cut Sol's cost by 40 percent in a ten-task test

Coding agents often spend the first minutes of every task hunting through folders they have never opened before. David Zhang, an engineer on GitHub since 2012, published a CLI called jevgrep on 2026-09-26 that gives the agent a shortcut: ask a plain question of the repository, and the tool returns relevant files, reading leads and verbatim source excerpts in one response. The repository has gained 686 stars since it went public, and the package @dzhng/jevgrep is on npm.

What jg does

The command is jg, and it takes a question in plain English and a folder to search. The examples in the project's README ask things like "Where is authentication checked before a request reaches a handler?" and "How are database connections created, pooled, and closed?". Jevgrep walks the repository, uses content previews to pick candidate files, then identifies useful source units and surrounding context. The output is a summary, then file locations, reading leads, and selected source with line references. Python and TypeScript or JavaScript files get declaration-level parsing; other text falls back to a generic reader.

The relevance judgement runs through Jev, TypeSafe's decision model, called through the Vercel AI Gateway, TypeSafe directly, OpenRouter or OpenCode Zen. The developer supplies one of those provider keys through jg auth, which stores it in an owner-only config file. Filters skip ignored files, hidden folders, dependency and build directories, binary files and obvious credential files, though the project notes those filters are not a guarantee that no sensitive content leaves the machine.

The number, and the tradeoff

Zhang reports one ten-task SWE-bench repeat where jevgrep dropped the full cost of the Sol coding agent from $7.62 to $4.52, about 40 percent lower, with Jev charges reported separately. The solve rate on that same run was 7 out of 10 with jg, against 8 out of 10 for the saved baseline. The README calls this a cost reduction with a quality tradeoff, and warns readers to treat it that way; it is one project's own measurement, on one run. An earlier run of the same corrected runtime solved 6 out of 10 at $5.54, and both runs failed the original quality gate. The sample is a tuned Python subset, so its numbers do not carry across to other coding agents or general savings.

The agent needs a matching skill

The package ships with an agent skill file that has to be installed separately with jg skill. Zhang writes that installing the CLI alone does not teach an agent to use it. The installer detects Claude Code, Codex and OpenCode and asks where to install; the skill tells the agent when to call jg and how to use its output. The current skill version also installs the CLI if it is missing. The project needs Node.js 22 or newer and runs on macOS or Linux.

For a team using Claude Code, Codex or another agent through the day, most of the tokens and most of the failed edits sit in the first pass, when the agent is still guessing where the change belongs. A cheap, focused search that returns real file paths and source excerpts moves that guessing out of the model. The measured saving above is one project's number on ten Python tasks. The wider pattern is worth trying anyway: put a small retrieval step in front of the big edit step, so the edit step starts with a smaller and better-chosen amount of text to reason over.

Source

Primary source: dzhng/jevgrep on GitHub.