Metrics

Every tool call is counted and timed and exposed as OpenTelemetry metrics:

Instrument	Type	Description
mcp.tool.calls	counter	Number of tool calls.
mcp.tool.call.duration	histogram (seconds)	Duration of the proxied upstream request.

Both carry the attributes tool (the tool/operation name) and outcome — and nothing else, so metric cardinality stays bounded. outcome is one of:

Value	Meaning
success	The upstream answered with a non-error status.
error	The upstream answered with a 4xx/5xx, or the request could not be built or sent.
auth_error	The upstream OAuth token could not be obtained, so no request was made. Points at the provider or the credential, not at the API.
cancelled	The client cancelled the call or disconnected before it completed; the upstream request in flight was aborted and no result was sent.

To break activity down by caller, log the relevant JWT claims with --trace-claim (see Tracing the caller’s JWT claims) and aggregate them in your logging backend, rather than turning a per-user identifier into a metric label.

Enable either exporter, both, or neither — they are independent:

# OTLP push to a collector + a Prometheus scrape endpoint, at once.
oas2mcp http \
  --openapi-file ./examples/petstore.yaml \
  --bind-addr 0.0.0.0:8000 \
  --otlp-endpoint http://otel-collector:4318 \
  --metrics-addr 0.0.0.0:9090
# Push: POST http://otel-collector:4318/v1/metrics  (HTTP/protobuf, every 30s)
# Pull: GET  http://0.0.0.0:9090/metrics            (Prometheus text format)

The Prometheus endpoint runs on its own HTTP server (the --metrics-addr address), separate from the MCP transport, so it works under stdio too. OTLP honours the standard OTEL_EXPORTER_OTLP_* environment variables.