@governmyai/sdk-anthropic

GovernMy.ai adapter for Anthropic's SDK tool-use loop. Wraps anthropic.messages.create() so that gated tool calls are checked against the regulatory rules engine before being surfaced to your code.

Companion package to @governmyai/sdk.

Install

npm install @anthropic-ai/sdk @governmyai/sdk @governmyai/sdk-anthropic

Usage

const { Anthropic } = require('@anthropic-ai/sdk');
const { GovernMy } = require('@governmyai/sdk');
const { wrapAnthropic } = require('@governmyai/sdk-anthropic');

const anthropic = wrapAnthropic(
  new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY }),
  {
    governmy: new GovernMy({ apiKey: process.env.GOVERNMY_API_KEY }),
    context: {
      riskTier: 'high',
      role: 'provider',
      annexIiiCategory: ['employment'],
      lifecyclePhase: 'deployment',
      systemId: 'hiring-ai-v2',
    },
    mode: 'warn', // 'warn' (default), 'block', 'log-only'
    gatedToolNames: ['submit_decision', 'send_offer', 'reject_candidate'],
    onObligations: async ({ toolName, obligations, blocked }) => {
      console.log(`[${toolName}] ${obligations.length} obligations${blocked ? ' (blocked)' : ''}`);
    },
  }
);

// Use the Anthropic SDK normally. When the model returns a tool_use block
// matching gatedToolNames, the adapter runs the rules engine check first.
const response = await anthropic.messages.create({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  tools: [
    { name: 'submit_decision', description: 'Submit hiring decision', input_schema: {...} },
  ],
  messages: [{ role: 'user', content: '...' }],
});

// Inspect regulatory results attached to the response
if (response._governmy) {
  for (const [toolId, info] of Object.entries(response._governmy)) {
    if (info.humanReviewRequired) {
      // Route to human review queue before executing the tool call
    }
  }
}

Mode behavior

| Mode | cannotBeAutoSatisfied match | Any humanReview.required match | Other matches | |---|---|---|---| | warn (default) | throws HumanReviewRequired | returns with obligations attached | returns | | block | throws HumanReviewRequired | throws HumanReviewRequired | returns | | log-only | throws HumanReviewRequired | returns with obligations attached | returns |

cannotBeAutoSatisfied obligations (EU AI Act Art. 14 et al.) always throw — this is structural and cannot be configured off.

When to use which mode

• warn: your code has a review queue or escalation path for flagged tool calls. Default.
• block: strict enforcement — any regulated tool call without prior review aborts.
• log-only: you want the event trail written (for audit) without any control-flow impact.

Catching HumanReviewRequired

const { HumanReviewRequired } = require('@governmyai/sdk');

try {
  const response = await anthropic.messages.create({ ... });
  // use response
} catch (err) {
  if (err instanceof HumanReviewRequired) {
    // err.obligations, err.unoverridable, err.context
    enqueueForReview(err);
    return;
  }
  throw err;
}

Notes

• Pass-through for non-tool-use flows. Messages that don't return a tool_use block run unchanged.
• The adapter is a thin Proxy — it preserves the rest of the Anthropic SDK surface (e.g. client.models.list()).
• Per-tool context layering: the adapter appends toolName to context.useCases for each gated call, so rules engine matching can be tool-specific.

License

MIT