jevgrep.

Semantic code search · powered by Jev

Let your coding agents find code by what it does.

Describe a behavior. Get the exact files, line ranges and source that justify the answer — with the coverage to know what was actually searched.

Get startedRead the CLI guide

~/acme-api

A terminal session. Running jg --all "where are expired CLI tokens rejected?" apps/webjudges 120 snippets and returns one match: apps/web/app/api/[...path]/route.ts lines 11 to 15, the function authenticate, which loads the token row filtered on revoked_at and expires_at. Runningjg exact -n --glob '*.ts' 'expires_at' apps/web lists every literal occurrence through ripgrep.

Replayed from this repository. 120 snippets judged, one match returned — and jg exact hands off to ripgrep unchanged.

How all-mode judges a match

With --all, every snippet gets two independent judgments — whether it acts on the entity you named, and whether it performs the operation you described — multiplied into one score. Related code that shares the vocabulary scores close, and still falls short.

Real scores from this repository, stable across three runs. The score ranks results; it is not a calibrated probability, and the threshold is tuned on a small labelled corpus rather than a representative one.

How the searches compare

Speed, model cost, and first-result accuracy on 40 development lookups.

Function lookup speed, cost, and first-result accuracy: Jev-only development prototype and Opus
Search methodSpeedCostAccuracy
JevgrepJev-only · development benchmark0.66s~29× faster in development~$0.000427~251× lower model cost92.5%First result · 37 of 40
Claude Opus 5Agent with grep + read19.0sRecorded baseline$0.107213Recorded model cost92.5%Selected answer · 37 of 40

Prototype timing excludes production overhead; speed and cost ratios use historical Opus measurements. Additional sample: Jevgrep 73.7%, Opus 92.1% first-result accuracy.

September 20, 2026

Behind the benchmarks

What we measured

40 lookups from scrapy, a repository we did not write. Each query is a function’s own docstring; the docstrings are stripped from the copy being searched, so nothing can match a query against itself. Cases a term-ranked grep already answers were removed. Jevgrep results are from one uncached development run on September 20, 2026, using Jev for both scope selection and comparison of complete function source. This prototype uses no generative query expansion. The backend now calls TypeSafe directly with no Gateway or model fallback; the updated CLI release is pending npm publication. The figures here remain the prototype’s development measurements.

The 0.66s figure is the mean local diagnostic time per lookup, with source inventory prepared before timing. It includes local ranking and real AI Gateway calls, but excludes source discovery and parsing, hosted application overhead, and database work. Opus’s 19.0s and $0.107213 are historical measurements; its accuracy comes from a later rerun. These are not simultaneous, isolated, end-to-end production measurements. Timings have not been remeasured for the direct TypeSafe connection.

Accuracy and its limits

Both systems selected an accepted implementation in 37 of 40 original-sample queries (92.5%). This measures the first result only. Both use the same frozen, reviewed labels, which accept equivalent implementations instead of requiring only the original function.

On the additional sample, Jevgrep selected an accepted implementation in 28 of 38 queries (73.7%), versus Opus’s 35 of 38 (92.1%). Two underspecified queries were excluded for both systems. Both samples were inspected during development, and label review was not blinded. Jevgrep had no failed searches in these two runs. The prototype has not reached 95% first-result accuracy or demonstrated accuracy parity across samples or repositories.

How cost is calculated

Jevgrep’s mean model cost is $0.000427263 per lookup in the same 40-query run, using AI Gateway list pricing of $0.042 per million Jev input tokens. It includes both Jev stages and four failed request attempts, each assigned the full input cost of its successful retry because failed-attempt usage was unavailable. All 40 lookups ultimately completed. Dividing the recorded Opus cost of $0.107213 by this mean gives approximately 251× lower model cost. Hosting and database costs are excluded. Account credits, promotional pricing, and cache discounts are not used in this comparison.

Your working tree, not an index

Reads the files on disk right now, including uncommitted edits, and respects your ignore rules. Nothing to build, sync or invalidate.

Judged one behavior at a time

Default search compares complete functions from likely code scopes. All-mode can evaluate individual blocks for queries about a guard or branch inside a larger function.

Answers an agent can act on

Up to three selected choices with a bounded output budget. --json and the MCP server share one versioned schema, and coverage says what was judged and what was not.

Start searching

npm install -g @remotehost/jg@preview
jg login
jg "reject expired sessions" .

Register it with a coding agent over MCP with jg mcp ., or keep it in the terminal. Exact searches still belong to ripgrep — jg exact passes your arguments straight through.

Create an accountSee every flag

Preview release. Search quality is evaluated on a small labelled corpus; agent speedups are not yet measured.