Developer David has open-sourced Jevgrep, a research tool built on Jev to help coding agents locate code inside a repository before passing that context to agents such as Claude Code and Codex for editing and testing.
According to the project description, a user can ask a question like where login validation is implemented, and Jevgrep will use Jev, described as a TypeSafe decision model, to search the repository layer by layer. It then returns relevant files and source snippets for the downstream agent to work with.
Built to cut token use in the code-finding step
The tool is aimed at the part of a coding agent workflow that spends heavily on tokens: finding the right code.
Jevgrep does not start by stuffing an entire repository into a model, and it is not limited to a single semantic search. It first decides which directories are worth exploring, then checks related files and code declarations, and finally returns source snippets, line numbers, and follow-up reading clues. The repository also ships with a Skill so an agent knows when to call jg to collect context.
SWE-bench test solved the same number of tasks, with lower Sol cost
The latest SWE-bench experiment used 10 Python tasks. Runs with Jevgrep and runs without it both completed 8 tasks.
Cost was different. Total GPT-5.6 Sol spending fell from $7.62 to $5.44 when Jevgrep was used, a reduction of 28.63%. David had initially written 40% on X, but the repository later updated the experiment result, and the current wording is about 30%.
Jev cost was not included in that headline figure
The reported reduction only counts Sol fees and does not include Jev usage. Based on the experiment logs, Jev call costs that can be confirmed add up to at least $1.57. Some calls do not have complete billing records, so the actual total remains unknown.
The test sample was also small: 10 tasks, with each task run only once.