Overview & result
JevGrep is an MCP server and CLI that lets a coding agent ask a natural-language question about a codebase, ranks relevant code fragments, and returns the original source with file paths and line numbers. The author reports lower context usage and search time in repository exploration tests, without publishing benchmark values.
From the original post
Claude Code can spend a surprising amount of context on repeated text searches and full-file reads.
So i thought a good use of Jev could be a "JevGrep", an MCP tool that lets Claude ask a natural-language question about a codebase, ranks code fragments by relevance, and returns the original source with file paths and line numbers.
Instead of doing multiple grep / rg searches and opening large files, Claude can directly ask things like:
"Where is symlink access controlled, and what prevents path escapes?"
and get back only the most relevant code.
In my tests, this noticeably reduced both context usage and search time on repository exploration tasks.
It works as a CLI, or as an MCP server connected directly to Claude Code.
GitHub: https://github.com/nassim-arifette/jevgrep
How Jev fits in the loop
- Accept a natural-language codebase question.
- Use Jev to rank candidate code fragments by relevance.
- Return original source with file paths and line numbers through the CLI or MCP server.
How to reproduce
- Install the CLI or MCP server from https://github.com/nassim-arifette/jevgrep.
- Ask a codebase question such as where symlink access is controlled and what prevents path escapes.
- Compare returned context and search time with repeated text searches, while treating the author’s qualitative result as unbenchmarked.
Why this build matters
Shows Jev used as a bounded relevance layer for coding-agent retrieval, with a concrete MCP surface and an explicit benchmark gap.
Limitations
- The source post publishes no retrieval precision, recall or latency numbers.
- A Reddit comment points to Graphify as an alternative code-indexing approach.