Connect AI assets

Connect every AI asset you run — agents, MCP servers and skills — so Boidra Platform can show what each one did, what it cost and how it performed. Boidra Platform ingests standard OpenTelemetry spans that follow the GenAI semantic conventions. If your asset already emits those, you only need to point its exporter at Boidra Platform. This page covers the data model and the one piece of config every path shares.

What you can connect

Asset	How it shows up	What to do
Agent — a custom Joule agent, an A2A agent on BTP, a domain agent	Its turns as invoke_agent spans and its LLM calls as chat spans, under its agent name	Point the agent's OpenTelemetry exporter at Boidra Platform (below)
MCP server — custom, or exposed through the MCP Gateway	Every call as an execute_tool span with the tool's name, and tool type mcp	Emit tool spans from the agent or client that calls it, or from the server itself
Skill — a reusable capability run by an agent or client	Through the agent or client that runs it: as the agent turn or the tool call it performs	Instrument the agent or client that runs the skill

Once connected, the same asset can be registered and published in the AI Asset Dashboard, so people can find it and its usage and cost are tracked from the first call.

The flow

Your AI asset (agent · MCP server · skill runner)
   │  (OpenTelemetry SDK)
   ▼
OTLP exporter  ──OTLP/HTTP──▶  POST /otel/v1/traces  ──▶  Boidra Platform backend  ──▶  Dashboard

The three span types

Boidra Platform stores every span as one of three types, decided from the GenAI gen_ai.operation.name attribute (with span-name fallbacks):

gen_ai.operation.name	Span type	Meaning
invoke_agent	invoke_agent	A top-level agent turn
chat / text_completion	chat	An LLM call
execute_tool	execute_tool	A tool or MCP call

Spans that aren't GenAI-shaped (HTTP client spans, framework "workflow" wrappers, zero-duration empty agent wrappers) are skipped — so you can export everything and Boidra Platform keeps only the meaningful agent/LLM/tool spans.

The attributes Boidra Platform reads

These are the OpenTelemetry GenAI attributes that populate dashboard fields. Fallback keys are accepted where shown.

Dashboard field	Attribute (with fallbacks)
Span type	gen_ai.operation.name
Agent name	gen_ai.agent.name → agent_name → (service.name for agents)
Model	gen_ai.request.model → gen_ai.response.model → model
Input tokens	gen_ai.usage.input_tokens → gen_ai.usage.prompt_tokens
Output tokens	gen_ai.usage.output_tokens → gen_ai.usage.completion_tokens
Tool name	gen_ai.tool.name → tool_name
Tool type	gen_ai.tool.type → tool_type
Conversation id	gen_ai.conversation.id → a2a.context_id → conversation_id
Tenant id	sap.tenancy.tenant_id → tenant_id
User id	user.id → user_id
Error	OTel span status ERROR, or error = true / outcome = "error"

The full authoritative table is on the OTLP attributes reference.

The shared config: point the exporter at Boidra Platform

Every path sets the OTLP endpoint to your Boidra Platform host, with the /otel prefix:

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

The OpenTelemetry SDK appends /v1/traces, so spans land at /otel/v1/traces.

Use OTLP/HTTP, not gRPC

Boidra Platform's ingest is OTLP/HTTP. Some SDKs (including the SAP Cloud SDK for Python) default to gRPC. When that's the case, also set:

export OTEL_EXPORTER_OTLP_PROTOCOL="http/protobuf"

Choose your path

• SAP Cloud SDK for Python — recommended; auto-instruments LLM calls and provides agent/tool span helpers.
• SAP Cloud SDK for AI (JS) — the JavaScript/TypeScript path.
• OpenTelemetry SDK — vendor-neutral, any language.
• Raw OTLP / curl — the lowest level, great for testing.