@sapper-ai/openai

OpenAI Agents SDK integration for SapperAI - drop-in tool input/output guardrails.

Installation

pnpm add @sapper-ai/openai

Quick Start

import { createToolInputGuardrail, createToolOutputGuardrail } from '@sapper-ai/openai'
import { RulesDetector, DecisionEngine } from '@sapper-ai/core'
import { Agent } from '@openai/agents'

// 1. Set up detection pipeline
const detector = new RulesDetector()
const engine = new DecisionEngine([detector])

// 2. Create guardrails
const inputGuardrail = createToolInputGuardrail(engine, {
  mode: 'enforce',
  defaultAction: 'allow',
  failOpen: true,
})

const outputGuardrail = createToolOutputGuardrail(engine, {
  mode: 'monitor',
  defaultAction: 'allow',
  failOpen: true,
})

// 3. Attach to OpenAI Agent
const agent = new Agent({
  model: 'gpt-4',
  tools: [myTool],
  inputGuardrail,
  outputGuardrail,
})

// Tool calls are now automatically scanned
const response = await agent.run('Execute command: rm -rf /')

API Summary

• createToolInputGuardrail(engine, policy) - Creates input guardrail for OpenAI Agents
• createToolOutputGuardrail(engine, policy) - Creates output guardrail for OpenAI Agents

Both functions return guardrail functions compatible with @openai/agents SDK.

How It Works

1. Input Guardrail: Scans tool arguments before execution
2. Output Guardrail: Scans tool results before returning to LLM
3. Action: Blocks malicious calls in enforce mode, logs in monitor mode
4. Fail-Open: On detector errors, allows execution to maintain availability

License

MIT