Send Google ADK Traces to KloudMate

The Agent Development Kit (ADK) has OpenTelemetry tracing built in. It creates a span for each agent run, model call, and tool call, following the GenAI semantic conventions. Set the standard exporter variables and ADK sends them to KloudMate, where every run appears in AI Observability.

Prerequisites

A KloudMate API key, from Settings → Ingest Keys.
ADK for Python 2.13 or later.

Install the packages

Install ADK with the otel extra, which adds the OTLP exporter:

pip install "google-adk[otel]"

Without the extra, ADK logs a warning and doesn’t export anything.

Point ADK at KloudMate

Set these variables in the environment that runs your agent:

export OTEL_SERVICE_NAME="support-agent"
export OTEL_EXPORTER_OTLP_PROTOCOL="http/protobuf"
export OTEL_EXPORTER_OTLP_ENDPOINT="https://otel.kloudmate.com:4318"
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=YOUR_KLOUDMATE_API_KEY"

KloudMate expects the raw API key in the Authorization header, with no Bearer prefix. If the KloudMate agent runs on the same machine, set the endpoint to http://localhost:4318 and leave out the header. With OTEL_EXPORTER_OTLP_ENDPOINT set, ADK also exports its metrics and logs to the same place.

Start exporting

For adk web and adk api_server, set the variables and start the server as usual. In your own code, call maybe_set_otel_providers() once, before you create a runner:

from google.adk.telemetry.setup import maybe_set_otel_providers

maybe_set_otel_providers()

Capture prompts and responses

By default, ADK doesn’t put prompts and responses in the GenAI attributes that AI Observability reads. To add them, set:

export OTEL_SEMCONV_STABILITY_OPT_IN="gen_ai_latest_experimental"
export OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT="SPAN_ONLY"

Use SPAN_ONLY, not true. In ADK, true writes content to log events only, which AI Observability doesn’t read.

Group a conversation

ADK uses the session ID as the conversation ID, so every turn in one ADK session is grouped into the same conversation. You don’t need to set anything.

Known limitations

When you use a non-Gemini model through LiteLLM, the provider comes from the model name prefix, such as openrouter/ or anthropic/. Some LiteLLM providers report reasoning tokens twice, which inflates output tokens and cost.

Check the data

Run the agent once, then open AI Observability → Traces. The run appears as a turn named after the root agent, with its model, steps, tokens, and estimated cost.