🛡️ @aegis-kernel/anthropic

Anthropic Claude Tool-Use Deterministic Safety Clearance Guard
Pre-Execution Clearance • Claude tool_use / tool_result Interception • Structured Self-Healing Feedback

🚀 Overview

@aegis-kernel/anthropic intercepts Anthropic Claude tool_use content blocks and evaluates deterministic security invariants before tool execution occurs. If a tool call violates a policy invariant, it constructs a standardized tool_result error block containing actionable self-healing guidance for Claude to re-ask safely.

📦 Installation

npm install @aegis-kernel/anthropic @aegis-kernel/core

⚡ Quickstart

import Anthropic from '@anthropic-ai/sdk';
import { AegisAnthropicGuard } from '@aegis-kernel/anthropic';

const anthropic = new Anthropic();
const guard = new AegisAnthropicGuard({
  mode: 'enforce',
  packs: ['@aegis/sql-guard', '@aegis/data-guard'],
});

const response = await anthropic.messages.create({
  model: 'claude-3-5-sonnet-20241022',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Delete records older than 30 days' }],
  tools: myClaudeTools,
});

for (const content of response.content) {
  if (content.type === 'tool_use') {
    // Intercept, evaluate invariants, and execute
    const resultBlock = await guard.handleToolUse(content, async (input) => {
      return await executeMyTool(content.name, input);
    });

    // Feed resultBlock back to Claude in next turn
    nextTurnMessages.push({ role: 'user', content: [resultBlock] });
  }
}

📄 License

Distributed under the MIT License. Copyright (c) 2026 Sneh Gabani.