Native OTel (CrewAI Enterprise)

CrewAI Enterprise ships an OpenTelemetry exporter built into the platform. Point it at our endpoint, set two env vars, and every crew run lands in ANTS Platform with zero code changes.

Prerequisites

• A CrewAI Enterprise account with permission to edit Settings → Organization and Environment Variables.
• ANTS Platform API keys. Generate them under Project Settings → API Keys in the ANTS Platform dashboard (opens in a new tab). You need both the public_key (pk-ap-…) and secret_key (sk-ap-…).

Setup

Generate the Authorization header

CrewAI's OTel collector authenticates with a single header. The value is HTTP Basic auth: base64 of <public_key>:<secret_key>.

macOS / LinuxPythonNode

PK=pk-ap-xxxxxxxxxxxx
SK=sk-ap-yyyyyyyyyyyy
echo "Basic $(printf '%s:%s' "$PK" "$SK" | base64)"

Copy the full output (including the Basic prefix). Treat it as a secret.

Add the OTel collector in CrewAI

In CrewAI Enterprise, go to Settings → Organization → OpenTelemetry Collectors → Add Collector. Fill in:

Field	Value
Type	Traces
Endpoint	https://api.agenticants.ai/api/public/otel/v1/traces
Service Name	crewai (or any identifier; we use it for diagnostics only)
Custom Headers → Key	Authorization
Custom Headers → Value	The Basic … value you generated above
Certificate	Leave blank

Hit Test Connection to confirm the endpoint accepts the credentials, then Save.

Set the GenAI semconv env vars

By default, CrewAI's exporter emits span structure but no message content or canonical token attributes. Two env vars unlock both. Go to Environment Variables in CrewAI Enterprise and add:

Key	Value
OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT	span_and_event
OTEL_SEMCONV_STABILITY_OPT_IN	gen_ai_latest_experimental

⚠️

Without these two env vars you'll see spans in ANTS Platform with empty input/output panels and missing per-call token usage. They're cheap to set and required for full fidelity.

Run a crew

Trigger any crew from CrewAI Studio or via your usual deployment. Within a few seconds, the trace appears under Tracing → Traces in ANTS Platform named after crewai.crew.name (e.g. MarketIntelligenceContentPipelineCrew).

What's automatic

You don't have to configure any of these — they happen on every trace:

• Trace name = the crew name from crewai.crew.name.
• Tags = framework:crewai plus crew:<crew name>.
• Session ID = crewai.crew.id (groups runs of the same crew). Override with gen_ai.conversation.id if you have a user-defined session.
• Per-agent rollup = every LLM and tool span is tagged with a stable, deterministic agent ID derived from the agent's role. The AI Command Center groups cost / latency / requests by this ID automatically.
• Cost = computed on our side. We map gen_ai.request.model against our pricing table and multiply by gen_ai.usage.*. If a model is missing, file a ticket.
• Tool labels = humanised from gen_ai.tool.name (e.g. search_the_internet_with_serper → Search the internet with serper).

End-user identity (optional but recommended)

CrewAI's OTel exporter doesn't emit user identity. To attribute traces to the human who invoked the crew, attach user.id from the app that calls crew.kickoff(). Three patterns, pick whichever fits your stack — none need a new pip install since CrewAI already ships opentelemetry-api.

Option 1: W3C baggage (recommended)

Baggage rides on every child span automatically. One call attaches identity to the crew, every task, every agent, every LLM call.

from opentelemetry import baggage, context
 
ctx = baggage.set_baggage("user.id", current_user.id)
with context.attach(ctx):
    result = crew.kickoff(inputs={"topic": topic})

We read baggage.user.id (and baggage.langfuse.user.id) into the trace's userId field.

Option 2: Wrap kickoff in a parent OTel span

If you already wrap each automation run in a span for your own APM:

from opentelemetry import trace
 
tracer = trace.get_tracer(__name__)
with tracer.start_as_current_span("run_market_research") as span:
    span.set_attribute("user.id", current_user.id)
    result = crew.kickoff(inputs={"topic": topic})

The parent span becomes the trace root and the crew is nested under it. Slightly noisier hierarchy in the UI — pick Option 1 if you don't already have your own span.

Option 3: Kickoff inputs (zero OTel)

result = crew.kickoff(inputs={"topic": topic, "user_id": current_user.id})

The identifier shows up in the trace's Input panel under crewai.crew.inputs but does not populate the per-user metrics — use only as a stopgap.

Recognised attribute keys

We read user identity from any of these on any span in the trace (we walk up to the root):

Next

• Stuck on missing input / cost / user attribution? See Troubleshooting.
• Self-hosted CrewAI or want full Python-SDK control? Use the SDK approach instead.