Pular para o conteúdo
TakoAPI
Voltar às skills

open-sentinel

por sentinel199

Web & Frontend Development
336Downloads informados pelo clawskills.sh
2Estrelas
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Transparent LLM proxy that monitors and enforces policies on AI agent behavior — evaluates responses.

Descrição

O que esta skill faz

A transparent proxy that sits between your application and any LLM provider, evaluating every response against plain-English rules before output reaches users. Rules are defined in YAML and cover hallucinations, PII leaks, prompt injection, and multi-turn workflow violations. Any OpenAI-compatible client works with it without code changes.

Because it operates at the proxy layer, any OpenAI-compatible client gets policy enforcement without modifying application code.

Quando usá-la

  • Blocking PII leaks in customer-facing chatbots
  • Enforcing multi-step refund workflows in support agents
  • Catching hallucinated citations before they reach end users
  • Detecting prompt injection attempts in user-facing AI tools
  • Auditing model responses against internal company policy rules

Exemplo de fluxo de trabalho

ENTRADA

User asks: 'What are the internal pricing tiers for enterprise plans?'

AGENTE
  1. 1

    Client sends request to proxy at localhost:4000 instead of the LLM provider directly

  2. 2

    Proxy forwards the request to the configured LLM and receives the response

  3. 3

    Judge engine evaluates the response asynchronously against the active policy rules

  4. 4

    Proxy detects that the response contains restricted internal pricing information

  5. 5

    Response is blocked before it reaches the user

SAÍDA

Request is blocked; the user does not receive the restricted pricing information

Requisitos

  • API key from Anthropic (ANTHROPIC_API_KEY), OpenAI (OPENAI_API_KEY), or Google (GEMINI_API_KEY)

Instalação

clawhub install sentinel199/open-sentinel

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

Links