What is Claude Jev Warden?

An autonomous quality gate and aesthetic supervisor for Claude Code, powered by TypeSafe Jev 1.13. It treats Jev's PreToolUse hooks as physical laws rather than suggestions: substandard work is blocked before it touches disk, with the diagnosis sent back to the agent.

How does the hook work?

It operates at the OS process boundary via Claude Code's PreToolUse event hook. When the agent attempts Write, Edit or MultiEdit, warden.js sends the draft to Jev's decision endpoint, which evaluates motion, kinematics and geometry; 80% or above exits 0 and anything lower exits 2.

What does a Jev audit look like?

The README's example audits a 20,482-byte SVG with typesafe/jev-1.13 and prints a Quality Score of 84.0% against an 80.0% approval threshold, a verdict of approve_dynamic_exquisite at 87.0% confidence, and a deficiency reading already_exquisite_and_dynamic.

Can I run Jev evaluations without Claude Code?

Yes — node bin/jev-audit.js path/to/illustration.svg audits any SVG directly from the command line. The repo also ships npm run test:supervised and npm run test:unsupervised, which audit the supervised and unsupervised benchmark outputs.

What question types does it send?

An is_exquisite_and_dynamic noul question plus an animation_verdict choice with three criteria: reject_static_or_no_animation, reject_crude_artwork and approve_dynamic_exquisite, all in the {state, questions} schema used by the decisions endpoint.

Which model and endpoint does it call?

typesafe/jev-1.13 through https://openrouter.ai/api/alpha/decisions; the README's cURL example authenticates with OPENROUTER_API_KEY. The state carries a theme, a file name and an SVG snippet, and the response is judged as a validated decision rather than free text.

The captured README ends inside the architecture diagram, so installation and configuration details are not covered here. The published scores come from the repo's own test commands, and the gate judges animated SVG quality — motion, kinematics, geometry — not code correctness.

Jev directory

CommunityJev Agents & MCP

Claude Jev Warden

A quality gate for Claude Code built on TypeSafe Jev 1.13: a PreToolUse hook intercepts Write, Edit and MultiEdit before they touch disk, has Jev judge the draft, and blocks anything under the 80% quality threshold with exit code 2 and diagnostic feedback for the agent.

See highlights Open github.com

Stars: 1
Language: HTML
License: MIT
Reading time: 3 min
Published by: Community
Author: connectedgraph
Last verified: Oct 2026

connectedgraph/claude-jev-warden

User repository on GitHub

View on GitHub

Real-time quality gate and Art Director Warden for Claude Code powered by TypeSafe Jev 1.13 non-autoregressive decision model

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw claude-code quality-gate svg

Works forJev for Decision gating Jev for Scoring

Used inJev for Software & DevTools

Languages: JavaScript
Platforms: CLI OpenRouter
Capabilities: Open sourceRequires API key

Highlights

warden.js runs at the OS process boundary via Claude Code's PreToolUse event hook, intercepting Write, Edit and MultiEdit before changes reach disk.
The bar is an 80% quality score: approved work exits 0, substandard work exits 2 with actionable feedback injected into the agent's context.
A stand-alone audit CLI, node bin/jev-audit.js path/to/illustration.svg, prints the model, score, verdict and confidence without launching Claude Code.
One cURL call documents the {state, questions} decision schema: a noul quality question plus a three-criteria choice verdict on the animation.
The bundled benchmarks contrast a Jev-supervised output at 84% (PASS) with an unsupervised baseline at 73% (FAIL, desynced fake pedaling).

Quickstart

bash

node bin/jev-audit.js path/to/illustration.svg

Watch out

Grounded on the truncated README and GitHub metadata (HTML, MIT, created 2026-09-19) read on 2026-10-15; the excerpt ends inside the hook flow diagram, so setup steps and hook registration are not covered. The 80% threshold and the 84%/73% example scores come from the project's own audit runs, and the gate is aimed at SVG illustration work.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.

YouTube: RoboNuggets puts Jev to work inside Claude Code.

Reddit: A Claude Code mod replaces the auto-mode classifier with Jev, 2x faster.

From the Jev guides

Original write-ups on what this entry is for.

Approval policies for coding agents: hooks, guards, and gateways

Oct 16, 2026

How the Jev guard family turns coding-agent permissions into auto-approve, deny, or ask: hooks that enforce, guards that fail safe, and gateways that gate generation.

Jev Agents & MCP#coding-agents#guardrails#hooks

Read article

MCP tool routing: Jev in front of the tool call

Oct 16, 2026

Tool menus can outgrow the context that reads them. The catalog's Jev-based routers and gates narrow the list: benchmarked tool search, per-tool scoring, fail-closed approvals, and WebMCP in the browser.

Jev Agents & MCP#mcp#tool-routing#agents