vigil
от robinoppenstam
Git & GitHubКратко
AI agent safety guardrails for tool calls.
Описание
Что делает этот навык
Vigil validates AI agent tool calls before they execute, blocking operations like destructive shell commands, SSRF, SQL injection, path traversal, and credential leaks. It runs as a drop-in npm package with 22 rules, zero runtime dependencies, and under 2ms latency per check.
Zero runtime dependencies and sub-2ms latency mean it can run inline on every tool call without adding measurable overhead to the agent pipeline.
Когда его использовать
- Blocking rm -rf commands issued by autonomous agents
- Preventing SSRF in agent-driven API and HTTP calls
- Catching SQL injection before a database tool executes
- Auditing all tool calls made by a shell-executing agent
- Adding a safety layer to an existing MCP server
Пример сценария
User asks: validate a shell command before my agent runs it
- 1
Import checkAction from vigil-agent-safety
- 2
Pass the tool name and params object to checkAction
- 3
Receive a decision (ALLOW, BLOCK, or ESCALATE) with a reason string
- 4
Block execution or proceed based on the returned decision
BLOCK: Destructive command pattern — the rm -rf / call is rejected before execution
Требования
- vigil-agent-safety npm package
Установка
clawhub install robinoppenstam/vigilВыполните команду CLI или вставьте промпт в Claude Code, Cursor, Codex — любой кодинг-агент, которым пользуетесь.