How does the JEV safety gateway fit between nginx and an LLM backend?

It is a pre-filtering reverse proxy: nginx forwards to the gateway, the gateway extracts the user input from each request and asks the JEV model to judge it, then either returns a 403 or transparently forwards the request to an OpenAI-compatible upstream.

Which API protocols and paths can it parse?

OpenAI, Anthropic and Gemini formats: /v1/chat/completions, /v1/completions, /v1/embeddings, /v1/moderations, /v1/messages, /v1/rerank, /v1/images/* and Gemini :generateContent, with a generic text-collection fallback for unknown paths.

How does it handle encoded or injected bypass attempts?

Before judging, it strips client-injected context such as blocks and decodes base64 segments in place; base64 decoding is on by default and can be turned off in the console.

What happens when JEV is down or a key is rejected?

Multiple TypeSafe API keys can be configured and rotated automatically on 401, 429 and 529 responses. On JEV failure the gateway can be set to fail open (allow) or fail closed (block), and abusive IPs get a temporary ban that spends no tokens.

Does the gateway store the text it sends for judging?

record_snippet defaults to off, so audit records keep the verdict, score and reason but not the submitted text. Enable it in the console when debugging; all records live in SQLite alongside the admin password login.

Should I expose the gateway ports directly to the internet?

No. The README binds both ports to 127.0.0.1 intentionally: the proxy port is meant to sit behind a local nginx and the console behind an SSH tunnel, because publishing 8080 would let scanners bypass nginx and clients forge source IPs.

What is the license and can I try it first?

It is AGPL-3.0-or-later, written in Go with a Vue 3 console. A public demo runs at jev-gateway.clim.asia with the password 123456, but the README warns it is preview-only: never put real JEV keys or real traffic through it.

Jev directory

CommunityJev Integrations & Adapters

JEV Safety Gateway

A pre-filtering reverse proxy that sits between nginx and an OpenAI-compatible LLM backend: it extracts user input from each request, asks TypeSafe's JEV whether it is harmful, blocks risky ones with a 403 and passes normal ones through unchanged, streaming included.

See highlights Open github.com

Stars: 19
Language: Go
License: AGPL-3.0
Reading time: 3 min
Published by: Community
Author: dark-hxx
Last verified: Oct 2026

dark-hxx/jev-safety-gateway

User repository on GitHub

View on GitHub

位于 nginx 与大模型后端之间的前置过滤反向代理：逐请求提取用户输入交给 JEV 判定，有害拦截、正常透明放行

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity catalog-import logicrw gateway moderation nginx

Works forJev for Moderation Jev for Decision gating

Used inJev for Security & Privacy

Languages: Go
Platforms: API SQLite
Capabilities: Open sourceSelf-hostedRequires API key

Highlights

Covers OpenAI, Anthropic and Gemini request shapes: chat completions, messages, rerank and generateContent, plus a fallback for other paths.
Strips client-injected context like <system-reminder> blocks and decodes base64 segments before judging, so encoded payloads cannot dilute the safety signal.
Rotates multiple TypeSafe API keys on 401/429/529 and can fail open or fail closed when JEV itself is unavailable.
record_snippet defaults to off: audit records keep verdicts, scores and reasons but not submitted text, all in SQLite behind an admin password login.
Deploys with one docker run; both ports publish to 127.0.0.1 only so a local nginx fronts the proxy and source-IP spoofing stays blocked.

Quickstart

bash

mkdir -p ./data ./geoip && sudo chown -R 10001:10001 ./data
docker run -d --name jev-safety-gateway --restart unless-stopped \
  -p 127.0.0.1:8080:8080 -p 127.0.0.1:8081:8081 \
  -v "$PWD/data:/data" -v "$PWD/geoip:/geoip:ro" \
  ghcr.io/dark-hxx/jev-safety-gateway:latest

Watch out

Grounded in the repository README (primarily Chinese) and GitHub metadata reviewed on 2026-10-15; the README publishes no latency or accuracy benchmarks, the public demo uses the shared password 123456 and must not carry real JEV keys or traffic, and the project is AGPL-3.0-or-later.

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.

X: Jev 1.14 is now on OpenRouter

YouTube: Building a contextual brand-safety scorer on Jev from scratch.

X: A website slop detector that checks 35 tells in 243 ms

X: Breaking, golden nugget, or slop

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

Decision gates in agent frameworks: what each integration ships

Oct 16, 2026

How Pydantic AI, Mastra, TanStack AI, DSPy, Spring AI, and LiteLLM expose Jev: where the typed decision sits in the loop, what each integration returns, and which big frameworks are only adjacent.

Jev Integrations & Adapters#agents#frameworks#integrations