Configuration

latenzy is configured with a YAML file (-c/--config). It has three sections: providers, probe, and exporter. API keys are read from environment variables only and are rejected if placed in the config.

providers (required, ≥1)

Each entry describes one provider and the models to probe on it.

Field	Default	Meaning
provider	—	anthropic, openai, or gemini
models	—	list of model IDs (≥1); charset-validated
endpoint	direct	free label distinguishing the path (direct / bedrock-us-east-1 / vertex / a gateway) — separates the same model over different routes
base_url	provider default	override the API base; must be an http(s) URL with a host, no query/fragment
api_key_env	per provider	env var holding the key (ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY by default)

probe

Field	Default	Bounds	Meaning
interval_seconds	300	≥10	delay between probe cycles
timeout_seconds	60	1–600	per-probe overall deadline
max_output_tokens	16	1–1024	cap on generated tokens per probe (keeps cost negligible)
prompt_classes	[small]	—	any of small, medium, large; latency scales with input size
concurrency	4	1–32	max simultaneous in-flight probes

exporter

Field	Default	Meaning
host	127.0.0.1	bind address; hostname/IP literal
port	9877	bind port
auth_token_env	none	env var holding a bearer token

Fail-closed rule: binding a non-loopback host without auth_token_env refuses to start. Loopback is zero-config; any routable bind requires a token, which Prometheus then sends as Authorization: Bearer <token>. See Security.

otel (optional)

Mirror measurements to OpenTelemetry in addition to Prometheus. Requires the otel extra: pip install 'latenzy[otel]'.

Field	Default	Meaning
enabled	false	turn the OpenTelemetry bridge on
endpoint	none	OTLP/HTTP collector URL (http(s)); omit to export to the console

otel:
  enabled: true
  endpoint: https://collector.internal:4318/v1/metrics

When enabled, latenzy run fans every measurement out to both Prometheus and OpenTelemetry. Instrument names follow the OTel GenAI conventions (gen_ai.client.operation.duration, gen_ai.client.token.usage) plus latenzy-namespaced TTFT/throughput. Apps that already run OpenTelemetry can skip this and use latenzy.otel.OTelBridge with their own meter (see the API reference).

Endpoints and model snapshots

endpoint is a first-class label so you can compare the same model over different paths (direct API vs Bedrock vs Vertex) — a real routing/procurement decision. Track dated snapshot IDs (e.g. claude-haiku-4-5-20251001) alongside aliases to catch a silent alias re-point, a common cause of a mystery latency shift.