Installation and setup

The Langfuse Python SDK installs with pip. Without keys it switches itself off and your code still runs; with keys it sends traces to Langfuse Cloud.

pip install langfuse 4.15.4

Three packages: the SDK, the OpenAI client that model calls go through, and pytest.

pip install "langfuse==4.15.4" openai pytest

from importlib.metadata import version

print("langfuse", version("langfuse"))
print("opentelemetry-sdk", version("opentelemetry-sdk"))

python versions.py

The SDK brings the OpenTelemetry packages with it. Langfuse builds its tracing on them, which is why a Langfuse trace can also hold steps recorded by other OpenTelemetry tools.

What the SDK does without keys

from langfuse import observe


@observe()
def answer(ticket):
    return "Order A17 shipped on 3 March."


print(answer("Where is my order A17?"))

python no_keys.py

@observe asks the SDK to record the function; lesson 3 explains it. With no LANGFUSE_PUBLIC_KEY set, the client logs an authentication warning and disables itself, and the function still returns its answer. The SDK is designed so that a tracing problem never breaks the application it traces.

With keys set, the default destination is https://cloud.langfuse.com, Langfuse Cloud's EU region. Lesson 2 shows that happening, and points the SDK at a server on your own machine instead.

Langfuse Cloud keys and a free modelOptional

No lesson needs these. To see traces in Langfuse's own interface, create a free Langfuse Cloud project, copy its public and secret keys from the project settings, and set three variables. For a hosted model instead of the local one, any provider with an OpenAI-compatible endpoint works with lesson 7's code; these have a free tier.

export LANGFUSE_PUBLIC_KEY="pk-lf-..."
export LANGFUSE_SECRET_KEY="sk-lf-..."
export LANGFUSE_BASE_URL="https://cloud.langfuse.com"
export GROQ_API_KEY="gsk_..."

Other Langfuse Cloud regions have their own address, such as https://us.cloud.langfuse.com; a self-hosted server has whatever address you gave it.