SAP Cloud SDK for Python

The SAP Cloud SDK for Python has a telemetry module that auto-instruments LLM calls and gives you context managers for agent, chat, and tool spans — all emitting the OpenTelemetry GenAI attributes Boidra Platform reads. This is the recommended path for Python agents.

Source of truth: the SDK's telemetry user guide. APIs can change upstream — check there if a signature differs.

1. Point the exporter at Boidra Platform

Boidra Platform ingests OTLP/HTTP, but this SDK defaults to gRPC. Set both the endpoint and the protocol:

export OTEL_EXPORTER_OTLP_ENDPOINT="https://<your-dashboard-host>/otel"
export OTEL_EXPORTER_OTLP_PROTOCOL="http/protobuf"

Debug locally first

Set OTEL_TRACES_EXPORTER=console to print spans to your terminal before sending them to a dashboard — handy for confirming the right attributes are set.

2. Enable telemetry

Call auto_instrument() before importing your AI libraries so their calls are traced:

from sap_cloud_sdk.core.telemetry import auto_instrument

auto_instrument()          # LLM calls are now automatically traced

from litellm import completion   # import AI libs AFTER auto_instrument()

auto_instrument(middlewares=None, disable_batch=False) — pass disable_batch=True to export spans immediately instead of batching (useful in short-lived scripts).

3. Wrap an agent turn

Use invoke_agent_span(...) around a top-level agent turn. This becomes an invoke_agent span in Boidra Platform, and the auto-instrumented LLM/tool calls nest underneath it:

from sap_cloud_sdk.core.telemetry import (
    auto_instrument, invoke_agent_span, set_tenant_id,
)

auto_instrument()
from litellm import completion

def handle_request(query: str, user_id: str):
    set_tenant_id("your-tenant-id")   # set at the request boundary

    with invoke_agent_span(
        provider="openai",
        agent_name="SupportBot",
        attributes={"user.id": user_id},
    ):
        return completion(
            model="gpt-4",
            messages=[{"role": "user", "content": query}],
        )

Span helpers → Boidra Platform

SDK call	Boidra Platform span	Populates
invoke_agent_span(provider, agent_name, conversation_id?, attributes?, propagate?)	invoke_agent	agent name, conversation id
chat_span(model, provider, conversation_id?, attributes?)	chat	model (tokens come from auto-instrumentation)
execute_tool_span(tool_name, tool_type?, tool_description?, attributes?)	execute_tool	tool name, tool type
set_tenant_id(tenant_id)	—	tenant id on the spans in this request
context_overlay(operation, attributes?)	per operation	generic operations without a dedicated helper

• conversation_id on invoke_agent_span maps to Boidra Platform's context id, which groups a multi-turn conversation across traces.
• Tokens are filled in automatically for auto-instrumented LLM calls (via gen_ai.usage.*); if you create a chat_span manually around a non-instrumented call, set those attributes yourself (see the OTel SDK guide).

Manual chat and tool spans

When autoinstrumentation doesn't cover a call, wrap it explicitly:

from sap_cloud_sdk.core.telemetry import chat_span, execute_tool_span

with chat_span(model="gpt-4o", provider="openai"):
    ...   # your LLM call

with execute_tool_span(tool_name="get_incident", tool_type="mcp"):
    ...   # your tool / MCP call

Verify

Run your agent, then open the dashboard's Traces page (or the Live tab). You should see an invoke_agent trace for your agent with nested chat/tool spans, the model, and token counts. If nothing appears, temporarily set OTEL_TRACES_EXPORTER=console to confirm spans are being produced, and double-check OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf.