Jevgrep (Code Search)

They are the author's own single-run measurement, and the README is careful about scope: ten tuned Python SWE-bench tasks, one frozen installed package, the exact public skill in the repo, both sides solving 8 of 10. A later rerun measured 25.8% lower total cost including Jev, and a further release cut Jev's own cost by about 59% while combined Sol-plus-Jev ran 2 to 3% higher. No speed claim.

No. The CLI alone does not teach the agent anything; `jg skill` installs the agent skill separately and needs npm or npx plus network access. Provider selection requires 0.3.0 or newer. Updating the npm package does not overwrite skill files already in your projects, so rerun `jg skill` after upgrades.

Searches send eligible source content to Jev through the provider chosen at `jg auth`, and the README says the ignore-file, hidden, dependency, binary and credential filters are not a guarantee that all sensitive information has been removed. `jg files [root]` counts what a search may read with no key and no network call, and --exclude skips paths for one search.

When you already know the symbol or path, where a direct read or rg is all you need. Jevgrep is aimed at questions that span unfamiliar files. What comes back is evidence for the agent, not a generated answer, and it does not guarantee every relevant file was found. Environment-based credentials are ignored; saved auth is what counts.

Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen, or a custom TypeSafe-compatible endpoint; provider selection requires 0.3.0 or newer, and jg auth saves the chosen provider's key in an owner-only config file.

Run jg skill, which delegates to the skills CLI and needs npm or npx plus network access, or run npx skills add dzhng/jevgrep --skill jevgrep directly; updating the npm package does not overwrite skill files already in your projects.

A summary and compact file list first, then selected source with line references, then detailed declaration and call locations; declaration parsing covers Python, TypeScript/JavaScript, Go, and Rust, with a fallback for other text.