Skip to content
plugmcp
DirectorySkillsCategoriesBlog
EN/中文
Sign inSign up
Home/Servers/Jev

Jev

I
by io.github.Aitejiu

Jev judgments for agents: injection scanning, shell risk gating, ranking

Power your agent

Add this server

The publisher's own command. Not a PlugMCP endpoint.

Local command
uvx jev-mcp==0.1.2
Credentials from the publisher
  • TYPESAFE_API_KEY — TypeSafe API key used for Jev judgments
Provider
Claude Code
TYPESAFE_API_KEY=<secret> claude mcp add io.github.aitejiu-jev -- uvx jev-mcp==0.1.2

Run in a terminal.

Claude Desktop
{
  "mcpServers": {
    "io.github.aitejiu-jev": {
      "command": "uvx",
      "args": [
        "jev-mcp==0.1.2"
      ],
      "env": {
        "TYPESAFE_API_KEY": "<secret>"
      }
    }
  }
}

claude_desktop_config.json

Step-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Related MCP

View more
  • Firecrawl🔥 Official Firecrawl MCP Server - Adds powerful web scraping and search to Cursor, Claude and any other LLM clients.
  • N
    NotionToken-auth Notion MCP: pages, databases, blocks, comments, files. Runs headless in CI and agents.
  • Context7Context7 Platform -- Up-to-date code documentation for LLMs and AI code editors

Related skills

View all
  • reticleAI agents can generate code, but still struggle to understand what they build. Reticle brings Jev-style machine-native runtime perception to web & desktop applications.
  • Intent-RouterIntent compiler for AI agents — converges vague requests into typed IntentSpec contracts (probe, ask, or halt before routing), the input layer for routers and typed-decision models like Jev & Laya
  • awesome-typesafe-jevAwesome Jev: a source-backed field guide to TypeSafe's System One model, with SDKs, live demos, agent tools, and independent evaluations.

Jev judgments for agents: injection scanning, shell risk gating, ranking

jev-harness-lab

用 TypeSafe Jev(System One 决策模型)做 agent harness 工程实验的技术仓库:评测框架、基准报告、可运行的集成工具(MCP server、skill router)。

skills.sh

安装

① MCP server(给 agent 加三个 Jev 工具)

用 uv 一条命令,无需克隆:

# 从 PyPI
uvx jev-mcp

# 直接从 GitHub 源码运行
uvx --from git+https://github.com/Aitejiu/jev-harness-lab jev-mcp

配置到任意 MCP 客户端(opencode / Claude Desktop 等):

{
  "mcp": {
    "jev": {
      "type": "local",
      "command": ["uvx", "--from", "git+https://github.com/Aitejiu/jev-harness-lab", "jev-mcp"],
      "environment": { "TYPESAFE_API_KEY": "your-key" },
      "enabled": true
    }
  }
}

提供的工具:

工具作用
scan_injection扫描工具/网页/邮件输出里的注入指令,返回 block / review / pass
bash_riskshell 命令四维风险打分(破坏性 / 触密 / 外发 / 不可逆),返回 deny / review / allow
rank_candidates候选片段按相关性打分排序(RAG 精排)

MCP Registry 归属标记(勿改格式):

mcp-name: io.github.Aitejiu/jev

② Agent skill(让 Jev 帮主模型选 skill)

npx skills add Aitejiu/jev-harness-lab --skill jev-skill-router

用 Jev 把任务路由到已安装的 skill,并且只加载选中那一个的完整指令,避免把整个 skill 目录塞进主模型上下文。技能页:https://www.skills.sh/aitejiu/jev-harness-lab/jev-skill-router

环境变量

两个集成都需要:

export TYPESAFE_API_KEY=<your-key>

目录

.
├── examples/            # 最小可运行示例(三原语、state、路由、护栏、抽取)
├── eval/                # 评测框架:11 个基准脚本 + 24 份报告 + 原始结果
│   └── datasets/        # 手工构造的数据集(如 130 条 shell 命令风险集)
├── integrations/
│   └── opencode/        # opencode 插件:jev_route_skill(skill 门控)
├── skills/
│   └── jev-skill-router/  # 可安装的 agent skill(SKILL.md + 独立脚本,skills.sh)
├── docs/
│   └── REPORT.md        # 技术评估报告(数据与结论)
├── src/jev_mcp/         # MCP server 包(PyPI: jev-mcp)
├── pyproject.toml       # Python 打包配置(console script: jev-mcp)
├── server.json          # MCP Registry 元数据(mcp-name: io.github.Aitejiu/jev)
├── mcp_server.py        # 兼容 shim:不安装也可 python mcp_server.py 运行
├── skill_router.py      # 本地 skill 目录路由(Jev 选择并加载 SKILL.md)
├── common.py            # .env 加载 + 共享 client
└── requirements.txt

快速开始

uv venv --python 3.12 .venv
uv pip install -r requirements.txt
echo "TYPESAFE_API_KEY=<your-key>" > .env
# 最小示例
.venv/bin/python examples/quickstart.py
.venv/bin/python examples/noul.py          # criteria 对判断的影响
.venv/bin/python examples/routing.py       # 阈值路由

# 跑评测(示例:skill router,501 个真实 skills)
.venv/bin/python eval/run_skillretbench.py --variant hybrid --per-setting 100

评测结果缓存在 eval/results/*.jsonl(已随仓库提交,可 --report-only 直接出报告);原始数据集在 eval/data/,需按 eval/README.md 的说明下载(已 gitignore)。

集成工具

MCP server

.venv/bin/python mcp_server.py   # stdio
工具作用输入 → 输出
scan_injection扫描工具输出中的注入指令tool_output → action(block/review/pass) + 概率
bash_riskshell 命令四维风险打分command → action(deny/review/allow) + 破坏性/触密/外发/不可逆分数
rank_candidates候选片段相关性重排query + candidates(≤10) → 排序后的 index/score

接入 opencode 的配置示例(项目 .opencode/opencode.json):

{
  "mcp": {
    "jev": {
      "type": "local",
      "command": ["/absolute/path/to/jev-harness-lab/.venv/bin/python", "/absolute/path/to/jev-harness-lab/mcp_server.py"],
      "enabled": true
    }
  }
}

Skill router

.venv/bin/python skill_router.py --query "帮我查飞书文档" --load

opencode 插件在 integrations/opencode/jev-skill-router.ts:注册 jev_route_skill(task) 工具,请求进来时用 Jev 从本地 skill 目录选出最合适的一个并返回其完整指令。参考做法是配合 agent.build.tools.skill = false 关闭内置 skill 工具,使主模型上下文不再携带整个 skill 目录。

主要基准结果(详见 docs/REPORT.md)

任务数据结果
间接注入检测InjecAgent 1,105 条阈值 0.10:P/R 100%,良性误报 0%
检索重排BEIR SciFact 900 对BM25 → Jev:MRR 0.622 → 0.843,Hit@1 50% → 78.3%
意图分类SNIPS / Banking777 类 97.9% / 77 类 80.3%
工具目录路由MetaTool 199 工具相似干扰 k=5 96.5%
Skill routerSkillRetBench 501 库hybrid 架构 R@1 75.8%(最强基线 38.0%)
命令风险门控自建 130 条危险拦截 100%、正常放行 98.2%
模型难度路由RouterBench51%(无信号,负结果)
轨迹失败归因Who&When 1,403 步AUROC 0.56(负结果)

总计约 22,500 次 API 调用、52.2M input tokens、$2.19。

说明

  • 所有 benchmark 使用公开数据集;原始结果 JSONL 已提交,报告可复现。
  • 部分官方基线为模拟实现(在报告中已标注)。
  • Jev 已知限制:choice 最多 255 个选项、仅文本输入、state+questions 共享约 32k tokens、英语为主。
  • 版本:jev-1.13.0,模型升级后建议重新评测。
How do I install Jev?
Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.
Does Jev need an API key?
The publisher lists these environment variables: TYPESAFE_API_KEY.
What transport does Jev use?
Jev runs locally over stdio.
When was Jev last checked?
Last checked 2026-10-16. Jev judgments for agents: injection scanning, shell risk gating, ranking

No tool list yet. Remote servers get one after the first health check.

plugmcp

Plug any tool into any agent. MCP servers for people who do not write servers. Gateway for volume, n8n for apps.

  • github.com/TODO
  • x.com/TODO
  • linkedin.com/company/TODO

PlugMCP, 2026

Product
  • Browse
  • Install
  • Host your own
  • Voice mode
  • Browser
Directory
  • All MCP servers
  • Agent skills
  • Categories
  • New
  • Voice
  • Submit a server
Developers
  • Docs
  • Blog
  • What is MCP
  • What is WebMCP
  • Status
Company
  • About
  • Contact
  • Terms
  • Privacy
  • DPA