COMMUNITY OBSERVEDMetric claims: AUTHOR REPORTEDSOURCE AUDITED🤖 Jev Agent

Jev Semantic Code Reading for Claude Code

A Claude Code plugin that uses Jev to locate relevant code and narrow large reads before the agent consumes context.

Overview & result

A Claude Code plugin backed by Jev answers bounded codebase questions and can narrow large file reads before Claude Code receives them.

From the original post

i can't believe this plugin is open source

one developer checked 9 Claude Code sessions and logged 422,164,852 cached input tokens

earlier file reads kept coming back with the conversation. each later request carried that context again

so he built a Jev plugin that selects the useful code before Claude reads it

the author's paired tests measured:

> finding code → 30% fewer billed input tokens

> checking files → 38% fewer

> narrowing large reads → 41% fewer

the search can work even when your words don't match the code

“where is the register flow” → 127 files checked → 4 matches

register appeared nowhere. the code called it signup, and the interface was in French

what you get:

→ jev find: locate code by what it does

→ jev ask: check a question across multiple files

→ read hook: select the relevant part of a large file

→ jev gain: track tokens examined, returned and spent

in 3 paired file-reading tests:

whole files → 12,190 tokens added to context

selected sections → 2,722 tokens

78% less code entering the conversation. both setups answered all 3 questions correctly

the measurements come from small author-run tests. your savings depend on the files and tasks

exact name → normal search

unknown name → Jev finds the relevant lines

Claude → reads them and does the work

the quoted article walks you through getting access to Jev and creating an API key

the repo has the full plugin installation and connection steps for Claude Code ↓

https://t.co/X7gZQ64dh4

— Mika

How Jev fits in the loop

  1. Jev scores code relevance or a bounded codebase question.
  2. The plugin returns a typed answer or a smaller relevant window.
  3. Claude Code reads the selected evidence and performs the coding work.

How to reproduce

  1. Read the exact X post and the BorisLeMeec/jev README.
  2. Review the find, ask and Read hook commands before using them on a repository.
  3. Treat the reported token results as author-reported small-sample measurements.

Why this build matters

It targets repeated context transfer, a practical bottleneck in long Claude Code sessions.

Limitations

  • The source post is a third-party summary of the repository rather than the repository author’s own X post.
  • The reported tests are small and depend on repository shape and query type.

Patterns