Jevrail hooks into terminal coding agents like Claude Code and Codex. Before a shell command actually runs, it asks a fast Jev decision model to score the command on things like destroying uncommitted work, causing irreversible data loss, touching production, writing outside the project, or exfiltrating data.
Each score comes back as a probability, not a guess, and a verdict of allow, confirm, or deny follows from the highest-risk score and its category, with a plain-English reason attached.
It ships as a CLI with hook, explain, install/uninstall, doctor, log, eval and exec commands, and can run in a deterministic, no-model mode when there is no key configured.
What you get
- Full TypeScript source code under the MIT license
- The jevrail CLI, its policy thresholds, and the evaluation benchmark docs
Jev
TypeScript
Node.js
Claude Code
Codex
