@authensor/connector-anthropic

Safety connector for the Anthropic API. Wraps @anthropic-ai/sdk so every messages.create() call goes through Authensor policy evaluation and Aegis content scanning.

Quickstart

npm install @authensor/connector-anthropic @anthropic-ai/sdk

import { createSafeAnthropicClient } from '@authensor/connector-anthropic';

const client = await createSafeAnthropicClient({
  apiKey: process.env.ANTHROPIC_API_KEY,
  policy: {
    id: 'my-policy',
    name: 'My Policy',
    version: '1.0.0',
    rules: [
      { id: 'allow-all', effect: 'allow' },
    ],
  },
});

const response = await client.messages.create({
  model: 'claude-sonnet-4-5-20250514',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello' }],
});

What it does

1. Before each messages.create() call, the request content is evaluated against your policy using the Authensor engine.
2. If the policy says deny, an AuthensorDeniedError is thrown.
3. If the policy says escalate (require_approval), an AuthensorEscalationError is thrown.
4. If the policy says allow, the request is forwarded to the Anthropic API.
5. After receiving a response, Aegis scans the output for threats (PII, injection, credentials, exfiltration).
6. If Aegis flags the response, an AuthensorContentThreatError is thrown.

Configuration

| Option | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | env var | Anthropic API key | | policy | Policy \| string | required | Authensor policy object or YAML string | | principalId | string | 'anthropic-connector' | ID for audit trail | | aegisEnabled | boolean | true | Enable Aegis response scanning | | onEvaluation | function | - | Callback after policy evaluation | | onAegisThreat | function | - | Callback when Aegis flags content |

License

MIT