@agenttrust/sdk

@agenttrust/sdk

Official JavaScript/TypeScript SDK for AgentTrust — prompt injection detection, agent identity verification, and audit trails for autonomous AI agents.

Installation

npm install @agenttrust/sdk
# or
yarn add @agenttrust/sdk
# or
pnpm add @agenttrust/sdk

Quick Start

import { AgentTrust } from '@agenttrust/sdk'

// From environment variable (recommended)
const at = AgentTrust.fromEnv() // reads AGENTTRUST_API_KEY

// Or explicitly
const at = new AgentTrust({ apiKey: 'atk_YOUR_API_KEY' })

Get your API key at agenttrust.ai. Keys are prefixed with atk_.

InjectionGuard

Scan any text for prompt injection, command execution attempts, and social engineering before your agent acts on it.

const result = await at.injectionGuard(text, {
  capabilities: ['send_messages', 'open_links'],
  channel: 'email',
  use_llm_review: true,
})

if (result.blocked) {
  // Hard block — do not proceed
  console.log('Triggers:', result.triggers)
  console.log('Mitigations:', result.mitigations)
} else if (result.requiresHuman) {
  // Queue for human review
} else {
  // Safe to execute
}

Options

Option	Type	Required	Description
capabilities	string[]	✅	What your agent can do (e.g. send_messages, open_links, data_access)
channel	'email' | 'chat' | 'api'	—	Source channel for better contextual analysis
use_llm_review	boolean	—	Enable additional LLM-based risk assessment
exclusions	string[]	—	Terms that suppress trigger detection
custom_keywords	CustomKeyword[]	—	Your own keyword triggers with severity

// Custom keywords example
custom_keywords: [
  { keyword: 'powershell', severity: 'high' },
  { keyword: 'wget', severity: 'medium' },
]

Dashboard-configured keywords and exclusions are automatically merged with request values.

Response

Field	Type	Description
risk_level	'low' | 'medium' | 'high'	Overall risk assessment
suggested_mode	'allow_execute' | 'draft_only' | 'require_human' | 'block'	Recommended handling
triggers	string[]	Rules that matched
trigger_excerpts	Record<string, string[]>	Supporting excerpts per trigger
mitigations	string[]	Suggested mitigation actions
ruleset_version	string	Detection ruleset version
llm_review	LLMReview | undefined	AI review block (when use_llm_review: true)
blocked	boolean	Convenience: true when suggested_mode === 'block'
requiresHuman	boolean	Convenience: true when suggested_mode === 'require_human'

TrustCode — Identity Verification

Issue a TrustCode

Prove your agent's identity when reaching out to users or other systems.

const trustCode = await at.issue({
  preset: 'draft_only',
  ttl_seconds: 3600, // 1 hour (default: 86400)
  payload: {
    intent: 'schedule_meeting',
    message: 'Scheduling a meeting on behalf of Acme Corp.',
  },
})

console.log(trustCode.code)        // e.g. "gFs2-jbQE-GddW"
console.log(trustCode.verify_url)  // shareable verification link
console.log(trustCode.badge_html)  // embeddable HTML trust badge

Verify a TrustCode

No API key required — anyone can verify.

const identity = await at.verify('gFs2-jbQE-GddW')

if (identity.verified) {
  console.log('Agent:', identity.issuer_agent_id)
  console.log('Org:', identity.issuer_org_id)
  console.log('Intent:', identity.payload.intent)
}

Framework Examples

Express middleware

import express from 'express'
import { AgentTrust } from '@agenttrust/sdk'

const at = AgentTrust.fromEnv()
const app = express()
app.use(express.json())

app.post('/agent/run', async (req, res) => {
  const check = await at.injectionGuard(req.body.message, {
    capabilities: ['data_access'],
    channel: 'api',
  })

  if (check.blocked) {
    return res.status(400).json({
      error: 'Prompt injection detected',
      triggers: check.triggers,
    })
  }

  if (check.requiresHuman) {
    return res.status(202).json({ status: 'queued_for_review' })
  }

  // Continue with agent logic...
})

LangChain tool wrapper

import { AgentTrust } from '@agenttrust/sdk'

const at = AgentTrust.fromEnv()

async function secureToolCall(input: string, capabilities: string[]) {
  const check = await at.injectionGuard(input, {
    capabilities,
    use_llm_review: true,
  })

  if (check.blocked) {
    throw new Error(`Blocked by AgentTrust: ${check.triggers.join(', ')}`)
  }

  // Call your tool...
}

Configuration

const at = new AgentTrust({
  apiKey: 'atk_YOUR_API_KEY', // required — get from dashboard
  baseUrl: 'https://agenttrust.ai/api', // optional, override for testing
})

Environment variable

Set AGENTTRUST_API_KEY and use AgentTrust.fromEnv() for cleaner setup.

Requirements

• Node.js 18+ (uses native fetch)
• Works with Bun, Deno, and any modern JS/TS runtime

Links

• Dashboard
• Documentation
• npm