@cernova/sdk

The detection layer for LLM pipelines. Wraps Anthropic's messages.create/messages.stream and OpenAI's chat.completions.create to automatically capture tokens, latency, and cost — Cernova runs per-step anomaly detection on every call and alerts you when something silently regresses.

Install

npm install @cernova/sdk @anthropic-ai/sdk

Quick start

import Anthropic from '@anthropic-ai/sdk';
import { Tracer } from '@cernova/sdk';

const tracer = new Tracer({ apiKey: process.env.CERNOVA_API_KEY! });
const anthropic = tracer.wrapAnthropic(new Anthropic());

const response = await anthropic.messages.create({
  model: 'claude-haiku-4-5-20251001',
  max_tokens: 256,
  messages: [{ role: 'user', content: 'Hello!' }],
  _trace: { stepName: 'my-step' },
});
// response is the normal Anthropic Message — your code is unchanged

Multi-step runs

Group multiple LLM calls into a single traced run so you can see the full pipeline in the dashboard:

const run = anthropic.run();

const step1 = await run.messages.create({
  model: 'claude-haiku-4-5-20251001',
  max_tokens: 16,
  messages: [{ role: 'user', content: 'Classify this: "refund request"' }],
  _trace: { stepName: 'classify' },
});

const step2 = await run.messages.create({
  model: 'claude-haiku-4-5-20251001',
  max_tokens: 256,
  messages: [{ role: 'user', content: `Reply to a ${text(step1)} inquiry.` }],
  _trace: { stepName: 'generate-reply' },
});

console.log(run.runId); // same run_id groups both steps in the dashboard

Streaming

messages.stream is fully supported — tokens and latency are captured after the stream ends with zero impact on streaming latency:

const stream = run.messages.stream({
  model: 'claude-haiku-4-5-20251001',
  max_tokens: 512,
  messages: [{ role: 'user', content: 'Tell me a story.' }],
  _trace: { stepName: 'story' },
});

for await (const event of stream) {
  if (event.type === 'content_block_delta' && event.delta.type === 'text_delta') {
    process.stdout.write(event.delta.text);
  }
}
// trace is ingested automatically once stream completes

API

new Tracer(config)

| Field | Type | Required | Description | |---|---|---|---| | apiKey | string | yes | Your Cernova project API key | | apiUrl | string | | Override ingest URL — defaults to Cernova's servers | | runId | string | | Custom run ID — auto-generated UUID if omitted |

tracer.wrapAnthropic(client)

Returns a wrapped client with .messages.create(), .messages.stream(), and .run().

anthropic.run()

Creates a TracedRun — a fresh run_id that groups all steps called on it. Each call to run() resets the step index to 0.

_trace option

_trace: {
  stepName?: string;  // label shown in the dashboard (default: step_1, step_2, …)
}

Stripped before forwarding to Anthropic — the provider never sees it.

What gets captured

| Field | Source | |---|---| | run_id | run.runId or tracer.runId | | step_name | _trace.stepName | | model | response.model | | input_tokens / output_tokens | response.usage | | latency_ms | wall-clock ms | | cost | computed from built-in pricing table | | status_success | true on success, false on thrown error | | output_code | full text content from response | | error | error message if the call threw |

Ingest is fire-and-forget

The POST to /ingest never blocks your app. Network failures are logged to console.warn and silently dropped — your LLM calls always complete normally.

Dashboard

View traces, anomaly scores, AI-powered run analysis, and cost breakdowns at cernova.dev.