Pular para o conteúdo
TakoAPI
Voltar às skills

agentaudit-skill

por starbuck100

Browser & Automation
—Ainda sem avaliações
Sua avaliação

Resumo

Automatic security gate that checks packages against a vulnerability database before installation.

Descrição

O que esta skill faz

A security gate that intercepts package manager commands and checks packages against the AgentAudit vulnerability registry before allowing installation. It assigns a trust score (0-100) and returns PASS, WARN, BLOCK, or UNKNOWN based on known findings. Supports npm, pip, yarn, pnpm, and clawhub.

It automates pre-install vulnerability checks across multiple package managers using a shared community registry with multi-agent consensus scores, replacing manual package research.

Quando usá-la

  • Checking an npm package for known vulnerabilities before adding it to a project
  • Blocking a low-trust pip library from being auto-installed by an agent
  • Auditing an unknown MCP server skill before first use
  • Catching typosquatted packages that mimic popular libraries
  • Reviewing telemetry or broad file-access findings before proceeding with a WARN-level package

Exemplo de fluxo de trabalho

ENTRADA

User asks: Install the chromadb package for my vector database project

AGENTE
  1. 1

    Run gate script: node $AGENTAUDIT_HOME/scripts/gate.mjs pip chromadb

  2. 2

    Gate queries agentaudit.dev/api for existing findings on chromadb

  3. 3

    Registry returns trust score of 52 with medium-severity findings

  4. 4

    Agent presents findings summary: telemetry collection enabled by default, broad file system access

  5. 5

    User decides whether to proceed or choose a different package

SAÍDA

WARN result (score 52/100) with findings summary displayed; user makes the final install decision

Requisitos

  • AGENTAUDIT_API_KEY environment variable (generated by running the registration script, which creates an account on agentaudit.dev)

Instalação

clawhub install starbuck100/agentaudit-skill

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links