dzhng/jevgrep

jevgrep – AI‑powered code‑search for coding agents

What it does

• jevgrep is a command‑line tool that lets a coding‑assistant (Claude Code, Codex, OpenCode, etc.) ask natural‑language questions about a code base and get back a concise list of the most relevant files, plus the exact source snippets and declaration locations that answer the query.
• It works by sending the repository’s filtered source (respecting .gitignore, hidden files, binaries, etc.) to a large‑language‑model endpoint – Vercel’s Jev model or any compatible TypeSafe‑style provider – which scores relevance across folders, files and code declarations.
• The CLI returns a structured stdout output: a short summary, a compact file list, then the matching code excerpts with line numbers and call‑site information. The output is meant for the agent to consume, not to be shown directly to a human.

Why it matters

• In many AI‑driven development workflows the agent first needs to locate the right piece of code before it can generate a change. Traditional tools (rg, git grep) require exact symbols or paths, which defeats the purpose when the agent only knows the behaviour it is looking for.
• The README reports a benchmark (SWE‑bench) where using jevgrep kept the same success rate (8/10 tasks) while cutting the total LLM‑inference cost by roughly 30 % compared with a baseline that performed no AI‑guided search.

How to get started

1. Install globally with npm (requires Node 22+):

   npm install -g @dzhng/jevgrep
   

2. Authenticate to a supported provider (Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen, or a custom compatible endpoint):

   jg auth
   

3. Install the skill that teaches your coding agent to invoke the CLI:

   jg skill   # or: npx skills add dzhng/jevgrep --skill jevgrep
   

4. Run a natural‑language query against a repository:

   jg "How are telemetry events recorded and sent?" ./my-project
   

   The tool will output a list of candidate files and the matching code excerpts.

Supported languages

• Full declaration parsing for Python, TypeScript/JavaScript, Go, and Rust.
• Other text files fall back to a simple content‑preview matcher.

Key commands

• jg auth – store provider key.
• jg skill – install the agent‑side skill.
• jg <question> <path> – perform a relevance search.
• jg files <root> – list how many files would be scanned (no network call).
• jg doctor – verify authentication.
• jg --help – see cache and search‑override options.

Installation notes

• No external ripgrep or Python runtime is needed; the tool is pure Node.
• Works on macOS, Linux, and Windows.
• Updates are done via npm; the skill must be re‑installed separately.

Development

• The repo uses TypeScript, Bun workspaces, and Turborepo.
• Tests run in Docker to ensure the published npm package works without extra dependencies.

License – MIT.

---

jevgrep is a genuine AI‑enhanced code‑search utility aimed at reducing the inference cost of coding agents while keeping their problem‑solving ability intact.