Back
developer-toolsagentsOpen source

jev-bouncer

Add an independent Jev risk check before Claude Code tool calls.

What Jev decides

Action class and risks such as irreversible effects, external writes, or secret exposure.

A Claude Code hook asks Jev about shell commands, edits, and MCP calls, then applies local policy and keeps a replayable decision log.

CHOICENOUL
Jev role
Provide an independent risk judgment on proposed coding-agent tool calls.
Input
The proposed shell command, edit, or MCP call with its execution context.
Output
Typed class probabilities and independent risk probabilities.
Decision frequency
Before eligible tool calls; deterministic allowlist checks skip the model.
Fallback
Local allowlists and tripwires handle known shapes; configured policy applies the model result.
Author reported · Source reviewed

How it works

Input
The proposed shell command, edit, or MCP call with its execution context.
Jev
CHOICE + NOUL
Output
Typed class probabilities and independent risk probabilities.

Overview

jev-bouncer evaluates eligible Claude Code tool calls before execution. It asks Jev for a typed action class and independent risk probabilities for shell commands, file edits, and MCP calls. A local policy turns those answers into allow, warning, or deny outcomes and logs each verdict. Read-only command shapes and suspicious URLs have deterministic local checks.

This is an additional probabilistic gate, not a security boundary. The repository explicitly keeps sandboxes and deny rules as the boundary and notes that hostile content can fool a model judgment.

Language
Python
License
MIT
Author
alsoleg89
Sources checked Sep 23, 2026

Sources

Related projects