Getting started with Authensor

Authensor is an open-source safety stack for AI agents. It provides policy evaluation, content scanning, approval workflows, behavioral monitoring, and cryptographic audit trails. This guide gets you from zero to enforcing your first policy.

Prerequisites

• Node.js 18+ or Python 3.10+
• pnpm (for TypeScript) or pip (for Python)
• An AI agent that makes tool calls

Install the SDK

For TypeScript:

pnpm add @authensor/sdk

For Python:

pip install authensor

Write your first policy

Create a file called policy.yaml in your project root:

version: "1"
rules:
  - tool: "*"
    action: allow
  - tool: "shell.execute"
    action: block
    when:
      args.command:
        matches: "rm -rf|shutdown|reboot"
    reason: "Destructive shell commands are blocked"
  - tool: "*.delete"
    action: escalate
    reason: "Delete operations require human approval"

This policy allows all tool calls by default, blocks destructive shell commands outright, and escalates any delete operation to a human reviewer.

Initialize and guard

import { createGuard } from '@authensor/sdk';
import policy from './policy.yaml';

const guard = createGuard({ policy });

// In your agent's tool execution loop:
const decision = guard('shell.execute', { command: 'ls -la' });
// decision.action === 'allow'

const blocked = guard('shell.execute', { command: 'rm -rf /' });
// blocked.action === 'block'
// blocked.reason === 'Destructive shell commands are blocked'

Enable content scanning

Add Aegis to scan tool arguments for prompt injection, credential exposure, and PII:

import { createGuard } from '@authensor/sdk';

const guard = createGuard({
  policy,
  aegis: { enabled: true }
});

Enable monitoring

Add Sentinel to track behavioral patterns and detect anomalies:

const guard = createGuard({
  policy,
  aegis: { enabled: true },
  sentinel: { enabled: true }
});

Next steps

• Read writing YAML policies for AI agents for advanced policy patterns
• Set up the control plane for centralized policy management
• Connect approval workflows for escalation handling