Send CrewAI Traces to KloudMate

CrewAI doesn’t export OpenTelemetry traces to your own backend. OpenLLMetry, from Traceloop, adds them: one Traceloop.init() call instruments CrewAI and the model SDKs it uses. Send those spans to KloudMate to see each crew run in AI Observability.

Prerequisites

A KloudMate API key, from Settings → Ingest Keys.
Python 3.11 or later.

Install the packages

pip install crewai traceloop-sdk

Initialize OpenLLMetry

Call Traceloop.init() at startup, before you import CrewAI:

import os

from traceloop.sdk import Traceloop

Traceloop.init(
    app_name="research-crew",
    api_endpoint="https://otel.kloudmate.com:4318",
    headers={"Authorization": os.environ["KLOUDMATE_API_KEY"]},
)

from crewai import Agent, Crew, Task  # noqa: E402

Use the base URL without /v1/traces, because the SDK appends it. KloudMate expects the raw API key in the Authorization header, with no Bearer prefix. If the KloudMate agent runs on the same machine, set api_endpoint to http://localhost:4318 and leave out headers.

You can set the endpoint and header with environment variables instead. These take priority over the values passed to init():

export TRACELOOP_BASE_URL="https://otel.kloudmate.com:4318"
export TRACELOOP_HEADERS="Authorization=YOUR_KLOUDMATE_API_KEY"

Turn off CrewAI’s own tracing

CrewAI has built-in telemetry and tracing that send data to CrewAI, and on the first run it can stop and ask whether to share the trace. Neither reaches KloudMate. Turn them off so only OpenLLMetry spans are sent and the run doesn’t wait for an answer:

export CREWAI_DISABLE_TELEMETRY="true"
export CREWAI_TRACING_ENABLED="false"

Group a conversation

Set a session_id association property before each run. KloudMate uses it as the conversation ID:

Traceloop.set_association_properties({"session_id": "chat-42", "user_id": "user-17"})
result = crew.kickoff(inputs={"topic": topic})

Control what’s captured

OpenLLMetry records prompts and responses by default. To leave them out, set TRACELOOP_TRACE_CONTENT=false. Token counts, models, and latency are still recorded.

Known limitations

CrewAI doesn’t create spans for its tools, including tools you define with @tool, so those calls appear only inside the model’s messages, not on the Tools tab. Tool calls through an MCP server are traced.
Calls that CrewAI sends to OpenRouter carry a bare model name without the provider, so they can’t be priced and count as calls not priced.

Check the data

Run the crew once, then open AI Observability → Traces. Each crew run appears with its model calls, tokens, and estimated cost, and each CrewAI agent appears in the Agents table on the Overview tab.