ENPT

jevgrep: CLI uses Jev to cut coding agent costs by 40%

Last updated:

Source · Model card: Jev

On 26 September 2026, David Zhang (founder of Duet, ex-NASA JPL) released jevgrep — a CLI that uses Jev to help coding agents find relevant source files before implementing changes.

The tool answers repository questions like “Where is authentication checked?” by exploring the codebase hierarchy, using Jev to judge file relevance, and returning file locations with source excerpts. It exposes a skill system that teaches coding agents (Claude Code, Codex, OpenCode) when to call jg for context collection.

The claim

Zhang reports ~40% lower coding agent cost on a ten-task SWE-bench repeat: $4.52 with jevgrep vs $7.62 baseline (using Sol agent).

The catch: solve rate dropped from 8/10 to 7/10. Zhang frames this explicitly as “a cost reduction with a quality tradeoff, not evidence of equal or better solve quality.” The sample is a tuned Python subset, not the full SWE-bench.

Metric With jevgrep Baseline
Cost $4.52 $7.62
Solve rate 7/10 8/10

All numbers are self-reported and task-specific. Treat as UnverifiedClaim.

How it works

npm install -g @dzhng/jevgrep
jg auth          # configure provider (Vercel AI Gateway, TypeSafe, OpenRouter, OpenCode Zen)
jg skill         # install the agent skill
jg "How are telemetry events recorded?" ./my-project

The CLI sends eligible source content to Jev through the selected provider. It respects .gitignore, excludes dependency/build folders, and skips obvious credential files — but these filters are not a security guarantee.

Why it matters

jevgrep is a production harness validating the “Jev as relevance judge” pattern for agentic workflows. The explicit quality vs cost tradeoff disclosure is unusually honest for a launch post. The skill system approach — teaching agents when to use specialized tools — may be more interesting than the benchmark numbers.

High engagement on launch: 244k views, 3.5k likes, ~5k bookmarks in under 24 hours.