@sentrial/mastra

Automatic instrumentation for Mastra AI agents. One function call tracks sessions, LLM steps, tool calls, tokens, cost, and latency.

Install

npm install @sentrial/mastra @sentrial/sdk
# or
pnpm add @sentrial/mastra @sentrial/sdk

Quick Start

import { instrumentAgent } from '@sentrial/mastra';
import { myAgent } from './mastra/agents/my-agent';

const agent = instrumentAgent(myAgent, {
  apiKey: process.env.SENTRIAL_API_KEY,
});

// Every generate() and stream() call is now tracked
const result = await agent.generate('What is the weather in NYC?');

Multi-Turn Conversations

Mastra's memory options are automatically mapped:

• memory.resource → Sentrial userId
• memory.thread → Sentrial convoId

const result = await agent.generate('Remember my name is Alice', {
  memory: { resource: 'user-123', thread: 'convo-456' },
});

What Gets Tracked

| Mastra Event | Sentrial Tracking | |---|---| | Each generate()/stream() call | Session with duration, cost, tokens | | Each LLM step (onStepFinish) | Tool call event: llm:{provider}:{model} | | Each tool call within a step | Tool call event with input/output | | Errors | Error event + failed session | | Final output text | Session assistantOutput |

Per-Call Overrides

const result = await agent.generate('Hello', {
  sentrial: {
    userId: 'custom-user',
    convoId: 'custom-convo',
    metadata: { experiment: 'v2' },
  },
});

Options

instrumentAgent(agent, {
  apiKey: 'sentrial_live_...',   // or SENTRIAL_API_KEY env var
  apiUrl: 'https://...',         // custom API URL
  agentName: 'my-agent',        // defaults to agent.name or agent.id
  failSilently: true,           // default: true
  pii: true,                    // auto-fetch PII redaction config
});