Standard export path
Point an OTLP/HTTP trace exporter at the project endpoint. Use protobuf or JSON and the batching behavior already supported by your SDK or collector.
OTLP/HTTP ingestion
Turn GenAI spans into usable cost analytics.
Keep the provider calls in your backend. Export completed LLM spans through your existing OpenTelemetry pipeline and analyze requests, tokens, latency, errors and estimated cost in HyperProxy.
HTTP protobuf or JSONGzip supportedIdempotent span ingestion
Use your telemetry stack
Add LLM analytics without changing request routing.
The OTLP adapter is for teams that already instrument backend AI calls and want a focused operational view of model usage.
GenAI metrics only
HyperProxy reads supported provider, model, token, status and timing attributes from completed client spans rather than storing arbitrary application traces.
Operational cost view
Join span usage with the model price catalog, keep incomplete cost visible and compare models or sessions in the same dashboard as gateway traffic.
Exporter configuration
One trace endpoint per project.
Create a server API key with ingest permission, then configure the OTLP traces endpoint and authorization header. Keep this token on a trusted backend or collector.
- Up to 512 spans and 2 MiB per batch.
- 600 submissions per project per minute.
- Trace IDs can group calls into sessions.
otel.envOTLP/HTTP
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=\
https://app.hyperproxyai.com/api/v1/admin/projects/PROJECT_ID/v1/traces
OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf
OTEL_EXPORTER_OTLP_TRACES_HEADERS=\
"Authorization=Bearer hp_obs_YOUR_SERVER_TOKEN"Reliable ingestion
Count each AI call once.
Finish the span
Export a completed GenAI client span with provider, model, duration, status and the token totals that are available.
Reuse span identity
Retries with the same trace and span IDs are idempotent, so an exporter retry does not duplicate usage or consume another request.
Choose one path
Do not export a gateway call again as external telemetry. Send each AI call through gateway metering or OTLP ingestion once.
Purpose-built scope
LLM metrics, not a second general APM.
| Input | Outcome |
|---|---|
| Completed GenAI client span | Request, provider, model, tokens, latency, status and estimated cost |
| Trace and span IDs | Session grouping and idempotent retry handling |
| Unsupported or unrelated spans | Reported through OTLP partial-success counts rather than stored as general traces |
| Prompt and completion bodies | Not part of the metrics record |
FAQ
OpenTelemetry and LLM questions.
Does HyperProxy store all span attributes?
No. The adapter extracts supported completed LLM metrics and does not serve as a general trace store.
Which OTLP formats are supported?
OTLP over HTTP with protobuf or JSON, including gzip-compressed requests.
Can I put the ingest token in a browser or mobile app?
No. It is a project-level server token and must remain on a trusted backend or collector.
Related