dzhng/jevgrep

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

What it solves

Coding agents often struggle and waste tokens (and money) searching for the correct files in an unfamiliar codebase. Jevgrep provides a way for agents to find relevant files, reading leads, and source excerpts by asking natural language questions about what the code does, rather than searching for specific keywords.

How it works

Jevgrep uses the Jev model to judge the relevance of folders, files, and declarations across a repository hierarchy. It explores the repository, selects files based on content previews, and identifies useful source units and their surrounding context. It supports declaration parsing for Python and TypeScript/JavaScript, with a fallback for other text files. The tool outputs evidence (file locations and source excerpts) to stdout for a coding agent to use as a starting point for implementation.

Who it’s for

It is designed for developers using coding agents (such as Claude Code, Codex, or OpenCode) who want to reduce the cost of agent-driven development by streamlining the initial research phase of a task.

Highlights

  • Cost Reduction: Demonstrated a ~30% reduction in coding-agent costs on specific SWE-bench tasks while maintaining the same success rate.
  • Agent Integration: Includes a "skill" installer (jg skill) that teaches coding agents how to invoke the tool.
  • Language Support: Specialized declaration parsing for TypeScript, JavaScript, and Python.
  • Provider Flexibility: Supports multiple providers including Vercel AI Gateway, TypeSafe, OpenRouter, and OpenCode Zen.

Related

  • Project
    NVIDIA/cccl

    A unified collection of core CUDA C++ libraries (Thrust, CUB, and libcudacxx) providing high-performance parallel algorithms and standard library abstractions for GPU development.

  • Project
    NVIDIA/CUDALibrarySamples

    A collection of GPU-accelerated examples demonstrating the use of NVIDIA CUDA libraries for high-performance math, image processing, and linear algebra.

  • Project
    boxlite-ai/boxlite

    A compute substrate for AI agents that provides hardware-isolated micro-VMs to securely run OCI images and execute code.

  • Project
    data-infra/cube-studio

    A cloud-native, one-stop MLOps platform that manages the entire AI lifecycle from data labeling and distributed training to model deployment, with extensive support for heterogeneous domestic AI hardware.