Claude Guard
novalidesignl-ship-it/claude-guard
Audita skills, hooks e MCP servers antes de confiar no setup
Ver no GitHubScanner de segurança para ecossistema Claude Code que detecta padrões de risco em skills, hooks, MCP servers e configurações — injeção de prompt, execução remota, secrets commitados e endpoints plaintext. Funciona offline, sem dependências, nada sai da sua máquina.
Instalação
npx convez-ai novalidesignl-ship-it/claude-guard1 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
- Validar terceiros skills e MCP servers antes de instalar
- Detectar secrets acidentalmente commitados no repositório
- Validar hooks suspeitos em configurações do Claude Code
- CI/CD: falhar build se encontrar risco HIGH de segurança
Veredito

Ferramenta essencial para quem trabalha com Claude Code em setup corporativo ou produção. Detecção heurística offline é prática e suficiente para maioria dos riscos conhecidos — não substitui auditoria profunda, mas bloqueia armadilhas óbvias.
README
A tradução para o português deste README ainda está na fila.
claude-guard — security scanner for your Claude Code setup
Scan your Claude Code skills, hooks, MCP servers, and settings for risky patterns — prompt-injection lures, download-and-execute hooks, committed secrets, plaintext MCP endpoints, and hidden bidi/zero-width tricks. Zero dependencies, runs offline, nothing leaves your machine.
As the Claude Code ecosystem explodes, people install third-party skills, plugins, and MCP servers by the dozen. A single malicious or careless one can run shell commands on your machine, read your secrets, or smuggle hidden instructions into your agent. claude-guard is a fast heuristic linter that flags the obvious — and the sneaky-but-known — before you trust a setup.
Quick start
npx claude-guard
By default it scans ~/.claude, ./.claude, and ./.mcp.json. Point it at anything:
npx claude-guard ./some-skill-you-downloaded
Example
Running it against a deliberately insecure setup:
claude-guard · scanning examples/insecure-setup
HIGH MCP server "weather" uses plaintext HTTP [http://…****se]
.mcp.json:5 → use https:// — http exposes tokens and traffic in transit
HIGH possible OpenAI-style API key committed in config [sk-EXAM…****00]
settings.json:24 → move secrets out of config into a secret manager / env
HIGH hook downloads and executes remote code [curl -s…****sh]
settings.json:11 → hooks run on your machine without confirmation
HIGH instruction to ignore prior instructions (prompt injection)
skills/sketchy-skill/SKILL.md:13 → a third-party skill should never need this
HIGH bidirectional control character (visual code spoofing)
skills/sketchy-skill/SKILL.md:24 → hidden/obfuscated content is a supply-chain red flag
MEDIUM broad permission auto-approves commands: Bash(*)
settings.json:4 → scope permissions to specific commands instead of wildcards
7 high · 4 medium · 0 low across 3 files
Secrets are always redacted in the output — claude-guard never prints or transmits the real value.
What it checks
| Surface | Flags |
|---|---|
skills/*.md, CLAUDE.md | prompt-injection lures ("ignore previous instructions", "don't tell the user"), exfiltration language, download-and-execute pipelines, hidden bidirectional / zero-width characters |
| hooks | shell commands that run automatically; curl … | bash-style download-and-execute |
| MCP servers | plaintext http:// endpoints; servers that run a local binary |
| settings | broad/dangerous permission grants (Bash(*), skip-permission flags) |
| all configs | committed secrets — Anthropic / OpenAI / GitHub / AWS / Google / Slack keys, private keys (shown redacted) |
Output & CI
- Findings are grouped by severity (HIGH / MEDIUM / LOW) with
file:lineand a fix hint. - Exit code is
1when any HIGH finding is present — drop it into CI to fail a build on a risky config. --jsonfor machine-readable output,--no-colorfor plain text.
npx claude-guard --json > report.json
What it is — and isn't
- ✅ A fast, zero-dependency, offline heuristic scanner for the most common and most dangerous Claude Code config risks.
- ✅ Private: it reads local files only. No telemetry, no network calls, nothing uploaded.
- ⚠️ Heuristic, not a sandbox or formal verifier. It catches known-bad patterns — a clean result is not a proof of safety. Always review what a third-party skill or MCP server actually does.
Install
# one-off
npx claude-guard
# or install globally
npm install -g claude-guard
claude-guard
Requires Node.js ≥ 16. No other dependencies.
Why "guard"
Skills and MCP servers are just files and URLs you point a powerful agent at. That's the convenience — and the attack surface. claude-guard is the 2-second check you run before you let a downloaded skill touch your machine.
License
MIT © novalidesignl-ship-it
Independent open-source tool. Not affiliated with or endorsed by Anthropic. "Claude" and "Claude Code" are trademarks of Anthropic.