User repository on GitHub
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Agents & MCP
Jew Guard
An opencode security plugin powered by Jev that risk-scores every tool call before it runs and scans every result for prompt injection afterwards. Deny blocks, ask prompts and allow runs, with configurable thresholds and an optional fail-closed mode.
- Stars
- 1
- Language
- JavaScript
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- surfaceflinger
- Last verified
- Oct 2026
- Languages
- JavaScript
- Capabilities
- Open sourceRequires API key
Highlights
- Jev risk-scores every tool call 0-3 plus three yes/no questions before it runs.
- Deny blocks, ask prompts, allow runs; explicit user requests override ask but never deny.
- Results are scanned for injections and canaries; flags are prepended and remembered for later calls.
- Guards agent instruction files like AGENTS.md and SKILL.md, caching results by content hash.
- From-scratch opencode-only port of jev-guard with one dependency; MIT; tunable JEV_GUARD_ threshold variables.
Quickstart
json
{ "plugin": ["jew-guard"] }Watch out
README and GitHub metadata reviewed 2026-10-17; the digest truncates the after-call and instruction-scanning sections. The guard is documented as a judge, not a sandbox, and it needs an OpenCode login plus a Console workspace to pay for the Jev calls.
Questions about this entry
Short answers drawn from the linked project or page.
A security guard plugin for opencode, powered by Jev as TypeSafe's System One model. Every tool call is risk-scored before it runs, and every result is scanned for prompt injections and canaries after it returns.
Jev returns a risk score from 0 to 3 plus three yes/no answers: should a human approve this, did the user actually ask for it, and does it follow planted instructions? Deny blocks, ask prompts and allow runs.
They compose: with a tool set to ask, a Jev allow auto-approves and a Jev ask opens a real prompt; under allow, an ask verdict only shows a warning. The README suggests making everything ask for the most predictable security.
Add { "plugin": ["jew-guard"] } to your opencode configuration and restart. The first run needs /connect for your normal OpenCode login, then /jev-workspace to pick the Console workspace that pays for the Jev calls.
By default the guard allows the call with a warning; setting JEV_GUARD_FAIL_CLOSED makes it block instead. An OPENROUTER_API_KEY can act as a fallback judge, with the README's note that the free tier trains on prompts.
No. The commands /jev-workspace, /jev-model, /yolo and /soft are human-only by construction; yolo and soft deliberately relax blocking, and soft turns deny verdicts into prompts.
It is a from-scratch port of leepokai/jev-guard, opencode-only, with one dependency, the opencode plugin API. Thresholds are tunable through JEV_GUARD_ variables with documented defaults such as deny at 2.5 and ask at 1.5.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
YouTube: Kubernetes admission control with a decision model, fail-closed.
Hacker News: SafeDep Gryph plus Jev checks for compromised coding-agent behavior.
2 pointsDetecting Compromised AI Coding Agents with Jev and GryphRead the thread on Hacker NewsX: Jev and Laya on security queries
X: Testing Jev for prompt injection
More Jev resources like this
An MCP proxy and hooks adapter that screens an agent's tool calls before they run and the tool results before the agent reads them, with thresholds in a policy file.
Jev Agents & MCP#community#mcp#security
A memory plugin for the DeepSeek Harness that binds knowledge and handoff documents to projects in one local Markdown directory. Its settings include a decision-model page for JEV-style models over the /v1/systemone protocol, used for active injection and end-of-turn scoring.
Jev Agents & MCP#community#github#javascript
Community
jojoman2024/dsh-memjJavaScript
jojoman2024/dsh-memjJavaScriptA Claude Code configuration CLI shipping Jev modules for model routing, skill suggestion, tool guardrails, auto-mode gates, and sandbox hooks.
Jev Agents & MCP#community#typescript#coding-agent
Community
davila7/claude-code-templatesPython
davila7/claude-code-templatesPythonA guard for coding agents that puts Jev in front of every step: it judges tool calls for intent and risk, scans tool outputs for instructions aimed at AI agents, and checks replies, mapping probabilities to allow, ask or warn.
Jev Agents & MCP#community#catalog-import#logicrw
Community
harshwasan/pi-jev-sentinelTypeScript
harshwasan/pi-jev-sentinelTypeScriptA local stdio MCP server that adds a jev_decide tool to OpenCode agents, supporting noul, choice and score questions so an agent can route or gate work on structured answers and probabilities instead of prose.
Jev Agents & MCP#community#catalog-import#logicrw
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.
Jev Agents & MCP#community#catalog-import#logicrw
Community
connectedgraph/claude-jev-wardenHTML
connectedgraph/claude-jev-wardenHTMLFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.