open-guardrail

Open-source guardrail engine for LLM applications.

Provider-agnostic text input/output middleware. Works in Node.js, browsers, and edge runtimes.

Install

npm install open-guardrail

Quick Start

import { pipe, promptInjection, pii, keyword } from 'open-guardrail';

const result = await pipe(
  promptInjection({ action: 'block' }),
  pii({ entities: ['email', 'phone'], action: 'mask' }),
  keyword({ denied: ['hack', 'exploit'], action: 'block' }),
).run('user input text here');

if (!result.passed) {
  console.log('Blocked:', result.action);
}

YAML Config

version: "1"
pipelines:
  input:
    mode: fail-fast
    guards:
      - type: prompt-injection
        action: block
      - type: pii
        action: mask
        config:
          entities: [email, phone]

import { OpenGuardrail } from 'open-guardrail';
const engine = await OpenGuardrail.fromConfig('./guardrail.yaml');
const result = await engine.run(text);

30 Built-in Guards

| Category | Guards | |----------|--------| | Security | promptInjection, regex, keyword | | Privacy | pii | | Content | toxicity, topicDeny, topicAllow, bias, language | | Format | wordCount, schemaGuard | | AI Delegation | llmJudge, hallucination, relevance, groundedness | | Operational | costGuard, rateLimit, dataLeakage, sentiment | | Agent Safety | toolCallValidator, codeSafety | | Advanced | copyright, watermarkDetect, multiTurnContext | | Korea/ISMS | piiKr, profanityKr, residentId, creditInfo, ismsP, pipa |

Key Features

• StreamingPipeline — chunk-level real-time guard validation
• GuardRouter — risk-based pipeline routing (low/medium/high)
• AuditLogger — EU AI Act / Korean AI Basic Act compliance logging
• Vercel AI SDK — middleware adapter (open-guardrail-vercel-ai)
• 8 Presets — default, strict, korean, security, content, ai-basic-act-kr, eu-ai-act
• Event hooks — guard:before, guard:after, guard:blocked, guard:error
• Dry run mode — test guards without blocking
• Provider agnostic — works with any LLM, any framework

Packages

| Package | Description | |---------|-------------| | open-guardrail | All-in-one (core + 30 guards) | | open-guardrail-core | Core engine only | | open-guardrail-guards | Guards only | | open-guardrail-cli | CLI tools | | open-guardrail-vercel-ai | Vercel AI SDK adapter |

License

MIT