`clownware/bouncer` 提供目前最硬的第三方独立标定数据
Claude Code PreToolUse layer where Jev answers YAML-defined questions about each tool call, with 19 hard rules and calibration runs.
Overview
Bouncer is a judgment layer for agents, starting with Claude Code tool calls. A PreToolUse hook asks Jev a handful of plain-English questions about each proposed call, cheap enough to run on every call (about four cents a day per the README). Questions and thresholds live in a YAML file you own.
It is designed for people running with permissions skipped: seatbelt mode never prompts, staying silent unless one of nineteen deterministic rules matches something never acceptable, such as printing a private key, a live credential on the command line or git stash clear. Judgments still run and are logged but only interrupt if you enable a deny threshold. The gate and a batch judge have shipped. Accuracy numbers come from live calibration runs against Jev, which the listing describes as some of the strongest third-party calibration data (Brier 0.036 to 0.149 across six questions); the README cautions that clearing the bar is not the same as behaving well.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Plain-English policy questions about each tool call (defined in YAML) | yes · no |
How it works
- Hook receives a proposed tool call
- Deterministic rules check for never-acceptable actions
- Jev answers the YAML questions
- Log judgments; deny only if a threshold is enabled
Strengths
- Zero interruptions by default with a hard floor
- Policy and thresholds in your YAML
- Published calibration runs
Caveats
- Calibration on the repo's own policy
- Judgments only block if you enable thresholds
Getting started
Install the Claude Code plugin and edit the YAML policy; see docs for modes.At a glance
- Type
- Project
- Status
- beta
- Language
- TypeScript
- Runtime
- Claude Code hook
- Has code
- yes
Use cases and integrations
How curated lists describe it
Accuracy/Brier/reliability by bin for 6 questions: Brier **0.036–0.149**;
wh000wh000/awesome-jev-live
Frequently asked questions
Will it interrupt me?
Not in seatbelt mode unless a hard rule matches or you enable a deny threshold.
Where are the questions defined?
In a YAML file you own.
How well calibrated is it?
Its published runs report Brier scores of 0.036-0.149 across six questions.
Related projects
Safety gate plugin for Claude Code, Codex CLI and Antigravity: local sensitive-file checks, then Jev judges blast radius.
PluginbetaTypeScriptClaude Code mod that answers auto mode's tool-call permission question with eight Jev yes/no checks behind a shell-aware blocklist.
ProjectbetaTypeScriptClaude Code hook that blocks known key formats locally and sends unknown high-entropy strings to Jev only in masked form.
HookbetaPythonUnix-pipeable CLI for Jev and other System One models: ask questions, gate shell commands via exit codes, calibrate thresholds.
CLIbetaGoClaude Code hook where Jev reviews each tool call against risk, authorization and scope questions and a local policy returns allow or deny.
ProjectbetaGoJev Sentinel: guard for Pi, Claude Code and Codex that checks tool calls, tool outputs and replies, failing closed.
PluginbetaTypeScript