Jev in production › Agent guardrails and approvals

jev-enforce

Claude Code plugin whose Stop and PostToolUse hooks ask Jev one yes/no question per CLAUDE.md rule about each reply and file edit, then block with the broken rules quoted. 93.3% precision and recall at the default threshold on 59 labelled examples (author).

Open on GitHub ↗

93.3%measured, as published by the source
Use
Agent guardrails and approvals
Industry
Developer tools
Form
Open-source tool
Stage
In production
Plugs into
Claude Code
Listed
2026-09-24
Found via
github
Repository
erkamyaman/jev-enforce
Stars
10
Forks
1
Last push
2026-09-27
Language
TypeScript
License
MIT
jev-enforce screenshot
demo/demo.gif in the erkamyaman/jev-enforce README, MIT; shown from GitHub.

The README opens with

Claude Code plugin that checks every reply and every file edit against your CLAUDE.md (or AGENTS.md), using TypeSafe's Jev. Break a rule and Claude gets it quoted back and fixes it, in the same turn.

CLAUDE.md is context, not a constraint. It holds for a few turns, then a handler shows up with raw SQL in it and a test quietly becomes it.skip. Nothing checks the output against the rules, so this does.

Jev makes it cheap enough to run every turn: it answers typed yes/no questions instead of writing text, all of them in one request. One question per rule, about 350ms, about 3 cents per 1,000 checks.

Badge

For the project's own README, linking back here:

Listed in Jev in production

Also used for agent guardrails and approvals