Observability
KnoxCall is OTel-native for AI observability: every gateway call emits an OpenTelemetry span using the GenAI semantic conventions, plus KnoxCall-specific attributes for the guardrail outcomes. Point it at your own collector and use whatever tracing/eval backend you already run — Langfuse, Braintrust, Datadog, Honeycomb, and so on. There is no proprietary eval store to lock you in.Enabling export
Configure the OpenTelemetry OTLP endpoint on the operator observability Integration (KnoxCall’s config lives in the admin UI, not environment files). Once an endpoint is set, spans export automatically; until then, tracing is off.Span attributes
Security and cost events you can alert on
Spans are for tracing a call. The events below are written to your audit log, which is where you set alert rules —POST /admin/alerts with
alert_type: "ai_gateway_event" and the action in condition_json.audit_action.
See alert rules for the rule shape.
Configuration changes are audited too, with the same
ai_gateway. prefix:
gateway_create / _update / _archive, agent_create / _update /
_archive, token_mint, token_revoke, firewall_policy_create / _update /
_delete, mcp_server_create / _update / _archive, and mcp_tool_upsert /
_update / _delete.
ai_gateway_event is not yet in the admin UI’s alert-type picker, so these rules
have to be created through the API today.