Telemetry¶

Status: Current | Last reviewed: 2026-09-10

Route your tenant's AI telemetry — LLM calls, speech services, and agent session traces — to your own observability backend (Arize Phoenix, LangFuse, or any OpenTelemetry/OTLP collector). This runs alongside any platform-level telemetry the operator has configured: enabling it never disables platform monitoring, and disabling it never affects other tenants.

You need the tenant Admin role — Settings → Tenant Settings is visible only to tenant admins and platform admins.

Configure the destination¶

Everything is configured on one screen — there's no separate integration to set up first.

Open Settings → Tenant Settings → Telemetry.
Turn on Enable tenant telemetry.
Fill in the destination:
- OTLP Endpoint URL — the base URL of your backend's OTLP ingest.
- Protocol — http/protobuf for most hosted backends; grpc for Phoenix's :4317 ingest.
- Project name — optional label your backend uses to group traces (e.g. a Phoenix project).
- Headers — add one row per header sent with every export. Tick Secret on any header that carries a credential (like an auth token); its value is stored encrypted and is never shown back to you — the field displays a masked placeholder, and leaving it masked keeps the stored value unchanged.
Click Save settings.

Backend examples¶

Backend	Endpoint	Protocol	Headers
Arize Phoenix	`http://your-phoenix-host:4317`	`grpc`	usually none (self-hosted)
LangFuse	`https://cloud.langfuse.com/api/public/otel`	`http/protobuf`	`Authorization` = `Basic <base64 of publicKey:secretKey>` (mark Secret)
OTel collector	`https://collector.example.com:4318`	`http/protobuf`	per your collector's auth (mark tokens Secret)

Changes take effect within about a minute — no restart or redeploy. New agent sessions and workflow runs start exporting immediately; a session already in progress when you change the setting may have its trace split across the old and new destinations.

What gets exported¶

Every AI-related trace for your tenant:

Agent sessions — a session trace per conversation with per-turn spans, tool calls, and per-service usage (tokens, characters, audio seconds) for every AI service in the pipeline (OpenAI, Deepgram, ElevenLabs, Google GECX, AWS, and the rest).
Workflow runs — a run trace with per-node spans, LLM calls, document OCR, content redaction, and transcription provider calls.

Infrastructure telemetry (API errors, HTTP requests) is platform-operated and is not sent to tenant destinations.

Turning it off¶

Toggle Enable tenant telemetry off and save. Export to your destination stops within about a minute. Deactivating or deleting the integration has the same effect.