@anticlaude/engine

Core library for AntiClaude: agent red-team evals, report comparison, local guard, and runtime policy.

npm install @anticlaude/engine

Features

• Attack runner — bundled payloads with adaptive variant strategies
• Eval suites — deterministic selection by seed, category, severity, tags, or ids
• Detection — rule-based + global patterns; optional LLM judge
• Report compare — baseline/current diff with score, breach, error, and category gates
• Guard SDK — local policy evaluator for prompt, tool-call, and output risk
• Runtime policy — per-agent tools, risk levels, allow/block/review
• Audit trace — redacted JSONL writer/parser and incident index
• MCP / skill audit — configuration and definition scanners
• Target adapters — generic JSON, OpenAI chat, Anthropic messages, custom templates
• Fixtures — local mock agents for eval labs

Usage

import {
  DEFAULT_GUARD_POLICY,
  DEFAULT_RUNTIME_POLICY_PROFILE,
  evaluateGuardPolicy,
  evaluateRuntimeToolRequest,
  runScan,
} from '@anticlaude/engine'

const report = await runScan({
  endpoint: 'https://your-agent.example/api/chat',
  target: {
    adapter: 'generic-json',
    bodyField: 'message',
  },
  payloadCount: 10,
})

console.log(`Report contract: v${report.reportVersion}`)
console.log(`Score: ${report.score}/100`)

const decision = evaluateGuardPolicy(DEFAULT_GUARD_POLICY, {
  surface: 'tool-call',
  toolCall: { name: 'refund_user', arguments: { amount: 9999 } },
})
console.log(decision.action)

const runtimeDecision = evaluateRuntimeToolRequest(DEFAULT_RUNTIME_POLICY_PROFILE, {
  agentId: 'support-agent',
  toolCall: {
    name: 'export_customer_data',
    arguments: { destination: '[email protected]' },
  },
})
console.log(runtimeDecision.action)

CLI wrapper: anticlaude
Docs: github.com/TacticSpaceTech/AntiClaude

License

AGPL-3.0-only