JavaScript / TypeScript SDK

SDK docs for tracking LLM usage in a few lines.

Install the package, initialize it once, and report token usage after each OpenAI, Anthropic, or other provider call. The SDK batches automatically and keeps the integration lightweight.

Start with quickstart View npm package

1. Install the package

Add the SDK to the service where your LLM calls run.

2. Initialize once

Call llmetrics.init at process startup with your API key.

3. Track after each model call

Send feature, provider, model, and token counts after every request.

4. Flush before shutdown

Call llmetrics.flush in short-lived jobs or before your process exits.

Quickstart

Track a request right after the model call returns.

The main flow is initialize once, then call llmetrics.track with the token counts returned by your provider SDK.

import OpenAI from 'openai';
import { llmetrics } from '@llmetrics/sdk';

const openai = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });

llmetrics.init({
  apiKey: process.env.LLMETRICS_API_KEY!,
});

const response = await openai.chat.completions.create({
  model: 'gpt-4o-mini',
  messages: [{ role: 'user', content: 'Summarize this lesson.' }],
});

llmetrics.track({
  feature: 'lesson-generation',
  provider: 'openai',
  model: response.model,
  inputTokens: response.usage.prompt_tokens,
  outputTokens: response.usage.completion_tokens,
  userId: 'user_123',
  meta: { promptVersion: 2 },
});

Installation

Add the SDK with your package manager of choice.

Command

npm install @llmetrics/sdk

Command

yarn add @llmetrics/sdk

Command

pnpm add @llmetrics/sdk

Initialization

Initialize once with a dashboard API key.

Call llmetrics.init during startup. If you forget the API key, initialization throws immediately so the integration fails loudly in development.

Field	Type	Default / Required	Description
apiKey	string	Required	Your LLMetrics API key from the dashboard.
flushIntervalMs	number	1500	How often queued events are flushed automatically.
maxQueueSize	number	50	Flush immediately once the queue reaches this size.
timeoutMs	number	2000	Timeout for each ingest request.
debug	boolean	false	Logs flush failures to the console.

Tracking

Use queued tracking by default, async tracking in short-lived runtimes.

llmetrics.track(event)

Fire-and-forget tracking. Events are added to an in-memory queue and flushed automatically based on queue size or interval. This call never throws.

llmetrics.trackAsync(event)

Sends immediately and throws on failure. This is the safer choice in serverless handlers, cron jobs, and other runtimes that may exit before the queue flushes.

await llmetrics.trackAsync({
  feature: 'summarize',
  provider: 'anthropic',
  model: response.model,
  inputTokens: response.usage.input_tokens,
  outputTokens: response.usage.output_tokens,
});

await llmetrics.flush();

Reference

Event payload fields

The dashboard groups by feature, provider, and model, so those fields should be stable identifiers from your application and provider response objects.

Field	Type	Default / Required	Description
feature	string	Yes	Logical feature name used for dashboard grouping.
provider	string	Yes	Provider slug such as openai or anthropic.
model	string	Yes	Model identifier returned by your provider.
inputTokens	number	Yes	Prompt or input token count.
outputTokens	number	Yes	Completion or output token count.
userId	string	No	Your internal user identifier for per-user cost analysis.
ts	number	No	Unix timestamp in milliseconds. Defaults to Date.now().
meta	object	No	Extra metadata stored alongside the event.