Skip to main content
Connect your OpenStack proxy to PostHog to automatically capture rich LLM analytics — no PostHog SDK or client-side changes required.

​
How it works

  • OpenStack emits a $ai_generation server-side event after every routed request.
  • The distinct ID is resolved from the provided user id, so make sure you pass a stable user identifier such as the X-Openstack-User header.
  • Events include latency, token usage, cost breakdowns, request / response metadata, and error flags drawn from the proxy state.
Events are recorded even if the upstream model call fails, enabling funnel and reliability analysis.

​
Setup

  1. In the OpenStack dashboard, open the paywall you want to analyse.
  2. Navigate to Configuration → Analytics and enable the PostHog toggle.
  3. Paste your PostHog Project API Key, and add a Project ID to annotate events.
  4. Save configuration. No code changes are required in your app.
This is especially helpful for low-code and no-code stacks (Zapier, n8n, Bubble, Make) where you cannot easily install PostHog libraries.

​
Event payload

Each $ai_generation event contains the consolidated proxy context. Key properties include:
PropertyDescription
$ai_modelModel name recorded by the upstream provider.
$ai_providerDetected provider (openai, anthropic, gemini, openrouter, etc.).
$ai_inputSanitized request messages captured by OpenStack (tools removed when privacy requires).
$ai_output_choicesSanitized assistant outputs or tool deltas when available.
$ai_input_tokens, $ai_output_tokensToken counts (prompt / completion) from provider usage metadata.
$ai_latencyTotal provider latency in seconds.
$ai_http_statusUpstream HTTP status or proxy error code.
$ai_is_error, $ai_errorFlags and messages for failed generations.
$ai_total_cost_usdRequest cost in USD.

​
Building dashboards

  • Use the $ai_is_error flag to build error funnels and monitor vendor reliability.
  • Segment by $ai_model, or $ai_provider to understand usage patterns.
  • Combine with PostHog notebooks or dashboards to compare LLM costs versus revenue, and trigger alerts when latency spikes.
Make sure every request to OpenStack includes a stable user value. Without it, events are dropped because PostHog requires a distinct ID for analytics.