posthog-replicate

PostHog LLM observability for the Replicate SDK. Drop-in replacement that automatically tracks all your AI model calls.

Install

npm install posthog-replicate posthog-node

Usage

import { Replicate } from 'posthog-replicate';
import { PostHog } from 'posthog-node';

const posthog = new PostHog('<your-api-key>');
const replicate = new Replicate({ posthog });

const output = await replicate.run('stability-ai/sdxl', {
  input: { prompt: 'A sunset over mountains' },
  posthogDistinctId: 'user_123'
});

await posthog.shutdown();

Every call sends a $ai_generation event to PostHog with model, latency, and input/output.

What's Tracked

• run() - full execution with output
• stream() - streaming responses
• predictions.create() - async prediction creation
• predictions.get() - prediction status polling (captures output when complete)

Async Predictions

Tracking options are automatically linked between create() and get() calls:

// Pass tracking options once at creation
const prediction = await replicate.predictions.create({
  model: 'stability-ai/sdxl',
  input: { prompt: 'A sunset' },
  posthogDistinctId: 'user_123'
});

// Options are automatically inherited - no need to pass them again
const result = await replicate.predictions.get(prediction.id);

What's Not Tracked

• predictions.list(), predictions.cancel()
• models.*, deployments.*, hardware.* and other non-generation methods

Caveats

• Call posthog.shutdown() before your app exits to flush pending events
• Input/output is tracked by default - use posthogPrivacyMode: true to disable
• Requires Node.js 20+