@triage-sec/sdk

Security-focused observability for AI agents. Captures telemetry from LLM calls automatically using OpenLLMetry and sends traces to the Triage backend via standard OTLP/HTTP.

Install

npm install @triage-sec/sdk
# or
pnpm add @triage-sec/sdk

Requires Node.js >= 18.

Quick start

import * as triage from "@triage-sec/sdk";
import OpenAI from "openai";

// Initialize once at app startup
triage.init({
  apiKey: "tsk_...",
  appName: "my-chatbot",
  environment: "production",
});

// Annotate application context before LLM calls
triage.setUser("user_123", "admin");
triage.setTenant("org_456");
triage.setSession("session_789", 1);

// Use any LLM provider normally — telemetry is automatic
const client = new OpenAI();
const response = await client.chat.completions.create({
  model: "gpt-4o",
  messages: [{ role: "user", content: "Hello" }],
});

OpenLLMetry auto-instruments OpenAI, Anthropic, Cohere, Mistral, Groq, Bedrock, Vertex AI, LangChain, LlamaIndex, CrewAI, Pinecone, ChromaDB, and 20+ other providers — zero provider-specific code needed.

Configuration

triage.init({
  apiKey: "tsk_...",        // or TRIAGE_API_KEY env var
  endpoint: "https://...",  // or TRIAGE_ENDPOINT (default: https://api.triageai.dev)
  appName: "my-app",        // or TRIAGE_APP_NAME
  environment: "production",// or TRIAGE_ENVIRONMENT (default: "development")
  enabled: true,            // or TRIAGE_ENABLED (default: true)
  traceContent: true,       // or TRIAGE_TRACE_CONTENT — capture prompts/completions (default: true)
});

All options can be set via environment variables. Explicit arguments take precedence over env vars.

Context helpers

Six helpers annotate traces with application-layer context that OpenLLMetry can't see:

// User identity (from your auth layer)
triage.setUser(userId: string, role?: string)

// Multi-tenancy
triage.setTenant(tenantId: string, name?: string)

// Conversation tracking
triage.setSession(sessionId: string, turnNumber?: number, historyHash?: string)

// Input before/after sanitization
triage.setInput(raw: string, sanitized?: string)

// Prompt template tracking
triage.setTemplate(templateId: string, version?: string)

// RAG chunk access control
triage.setChunkAcls(acls: Record<string, unknown>[])

Scoped context

Use withContext to scope annotations to a callback:

triage.withContext({ userId: "user_123", tenantId: "org_456" }, async () => {
  // All LLM calls here get these annotations
  await client.chat.completions.create({ ... });
});
// Context is restored after the callback

Shutdown

// Flush pending traces before process exit
await triage.shutdown();

License

See LICENSE.