Kit de Segurança para Agentes IA Locais
a42189399-ux/local-ai-agent-safety-kit
Checklists de segurança defensivas para agentes IA em desenvolvimento local
Ver no GitHubUm conjunto de checklists defensivas e diretrizes de isolamento para desenvolvedores que usam agentes de IA com acesso ao terminal em seus ambientes locais. Fornece procedimentos práticos para proteger credenciais, implementar sandboxes, auditar mudanças e validar conexões MCP, reduzindo riscos de exposição de segredos e execução não autorizada.
Instalação
npx convez-ai a42189399-ux/local-ai-agent-safety-kit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Proteger credenciais e API tokens antes de autorizar agentes IA no repositório
- Implementar isolamento de sandbox usando Docker/VM e permissões granulares
- Auditar logs de sessão de agentes e criar workflows de rollback via git
- Validar segurança de servidores MCP conectados localmente
Veredito

Recurso educacional prático e bem estruturado para mitigar riscos reais de agentes IA em ambientes de desenvolvimento. Especialmente valioso para equipes adotando agents como Cursor e Cline, mas não substitui auditorias formais de segurança nem oferece garantias de proteção absoluta.
README
A tradução para o português deste README ainda está na fila.
Local AI Agent Safety Kit
[!NOTE] Non-Official Educational Checklist This repository is a community-driven, defensive educational resource. It is not affiliated with, authorized, or endorsed by OpenAI, Anthropic, Google, GitHub, Cursor, Docker, or any other vendor. This kit does not provide security guarantees.
🎯 Target Audience & Scope
This kit provides practical security checklists, isolation guidelines, and credential-guarding workflows for:
- Individual developers integrating terminal-based AI agents into their local workflows.
- Small engineering teams auditing security implications of agent adoption in their workspaces.
🛠️ Supported Contexts & Tools
While not claiming complete coverage, the concepts in this kit apply broadly to local AI coding agents and execution environments, including:
- CLI Agents: Codex, Claude Code, Gemini CLI, Antigravity.
- Editor Extensions: Cursor, Cline, Roo Code.
- Extensible protocols: MCP (Model Context Protocol) servers and custom tool harnesses.
🚀 Quick-Start Checklist (Top 10 Actions)
Before launching or authorizing an AI agent inside a local repository, run through these ten rapid checks:
- Isolation: Run the agent in an isolated directory structure (or Docker/VM sandbox) rather than your root user directory.
- Branch Protection: Initialize a clean git state and checkout a separate scratch/sandbox branch before launching the agent.
- Secret Hiding: Add all
.envfiles,.git/config, SSH keys, and cloud credentials to the project.gitignore. - Prompt Approvals: Turn on write-approval prompts in the agent configuration (do not run with blanket bypass permissions by default).
- Keychains: Keep active API tokens out of directories the agent can read.
- History Cleanup: Check that shell history configuration does not write keys in plain-text to files the agent can inspect.
- Log Verification: Truncate or monitor agent session logs to verify no credentials are being output in plain text.
- MCP Allowlist: Only connect to trusted MCP servers running on your local machine; verify port boundaries.
- Pre-commit Hooks: Run pre-commit secret scanners to detect keys before pushing any code written by agents.
- Commit Small: Commit files in small, logical chunks so you can easily review diffs and roll back accidental filesystem writes.
📂 Kit Documents
- Agent Permission Checklist: Questions to answer before running an agent in a project.
- Secrets & Env Guardrails: Defensive rules to isolate credentials, keychains, and tokens.
- Sandbox Modes Comparison: Conceptual overview of filesystem containment modes.
- Audit & Rollback Workflow: Git-based workflows to trace changes and undo mistakes.
- Workspace Isolation Template: Safe directory structures and docker configs.
- MCP Server Risk Checklist: How to verify local tool integration safety.
- Disclaimer: Responsibility statements and liability disclaimers.