Jevgrep

Jevgrep is what a growing set of community CLIs call themselves: semantic code search that finds code by what it does, powered by Jev. You ask in plain English — or pipe in a stack trace — and Jev scores each line or chunk, returning ranked, grep-style file:line hits with calibrated confidence. These are community projects, not TypeSafe and not us.

Regular grep matches characters; jevgrep matches meaning. Instead of an embedding index you have to build and keep fresh, the tools send code chunks and your plain-English question to Jev — TypeSafe AI's System One model — and get back a typed judgment per chunk in a single non-autoregressive pass. That's what makes "find where we record telemetry" resolve to real file:line hits in one command, ranked by calibrated confidence rather than by keyword overlap.

jevgrep turns your question into per-chunk Jev judgments, then prints grep-style file:line hits ranked by confidence.

Install one — mind the name

"jevgrep" is a shared name: several community authors ship a tool by that name, and the bare unscoped package is not the same as any of them. Install by the exact scoped package. Two that are actively documented:

# dzhng's build (MIT) — installs an agent skill for your coding agent:
npm install -g @dzhng/jevgrep
jg auth        # store your Jev / OpenRouter key
jg skill       # drop the skill into your coding agent
jg "how are telemetry events recorded and sent?" ./my-project

# nassim-arifette's build — choose your provider explicitly:
npm install -g @nassim-arifette/jevgrep
jevgrep init --global --provider typesafe   # or: openrouter | vercel

Both need Node 22+ and a Jev key. They reach Jev through TypeSafe directly, or through OpenRouter's Decisions endpoint / the Vercel AI Gateway — whichever provider you configure.

How it uses Jev

One typed judgment per line or per code chunk — a noul (does this satisfy the question?) or a score — so ranking is calibrated, not keyword-based
No embedding index to build or invalidate; the model reads the actual chunks each run
grep-style file:line output and exit codes, so it drops into scripts and coding agents
Runs in ~70–500ms per Jev pass, cheap enough to point at a whole repo

Community projects (not ours)

These are independent, community-built tools that use Jev — none are made by TypeSafe or by us. The clearest map of the space is the community awesome-jev list, which catalogs several grep-style variants (a per-line noul tool, a chunked semantic grep, a Rust scorer with no index):

awesome-jev — community-curated index of projects built on Jev, including several jevgrep-style tools
dzhng/jevgrep (@dzhng/jevgrep) — MIT CLI: jg auth / jg skill, natural-language code search for coding agents
nassim-arifette/jevgrep — scoped CLI with a provider switch (TypeSafe / OpenRouter / Vercel)

Build your own semantic grep

The core is a small loop: split a file into chunks, ask Jev one typed question per chunk, keep the high-confidence hits. It's the same POST https://jevtypesafeai.com/api/v1/decide call (official upstream POST https://api.typesafe.ai/v1/systemone) with a jv_live_ key — a noul per chunk, sorted by confidence. See the MCP server and the AI-agent guide for the request shape.

FAQ

Is jevgrep made by TypeSafe or by you?

No. jevgrep is a name several community developers use for their own Jev-powered code-search CLIs. We don't publish jevgrep and neither does TypeSafe — this page just explains what it is and links the real authors.

Which package should I install?

Install by the exact scoped name of the build you want — for example @dzhng/jevgrep or @nassim-arifette/jevgrep — not the bare unscoped "jevgrep", which is a different project. All need Node 22+ and a Jev key.

How is it different from grep or an embedding search?

grep matches text; embedding search matches vector similarity from an index you maintain. jevgrep asks Jev a typed question about each chunk at query time, so it ranks by calibrated confidence in what the code does, with no index to build or refresh.

Do I need a TypeSafe key, or can I use OpenRouter?

Either. The tools let you point at TypeSafe directly, at OpenRouter's Decisions endpoint, or at the Vercel AI Gateway — you set the provider during init.

Roll your own with the decision API

Semantic grep is just one typed question per chunk. Run one free in the browser — no signup — then grab a jv_live_ key.

▶ Try Jev free Get an API key →