Langfuse¶

What it is¶

An open-source LLM engineering platform for tracing, observability, metrics, and evaluation.

What problem it solves¶

It allows developers to debug complex LLM interactions, track costs, monitor latency, and run evaluations (scores) on agent outputs.

Where it fits in the stack¶

Category: Process & Understanding / Observability

Key Features¶

Tracing: Visualizes the nested calls in an agentic loop, including tool calls, LLM latency, and token usage.
Evaluation: Automates model-based grading of responses using LLM-as-a-judge or manual scoring.
Prompt Management: Version-controlled prompt management that allows decoupling prompts from application code.
OpenRouter Integration: Native support for receiving logs from OpenRouter without additional instrumentation.

Getting started¶

Installation¶

pip install langfuse

Basic Integration (OpenAI)¶

Langfuse provides a wrapper for the OpenAI SDK that automatically captures traces.

from langfuse.openai import openai

# Standard OpenAI call, now automatically traced
response = openai.chat.completions.create(
  model="gpt-4o",
  messages=[{"role": "user", "content": "How does Langfuse help with AI observability?"}],
  name="obs-test-run" # Optional: name the trace
)

print(response.choices[0].message.content)

CLI examples¶

langfuse api traces list¶

Lists recent traces from your project:

langfuse api traces list --limit 10

langfuse api prompts get¶

Retrieves a specific prompt version from the Langfuse prompt management system:

langfuse api prompts get --name "my-prompt-name"

langfuse integration claudecode enable¶

Enables tracing for Claude Code sessions:

langfuse integration claudecode enable

API examples¶

Python (LangChain Integration)¶

from langfuse.callback import CallbackHandler

langfuse_handler = CallbackHandler()

# Pass the handler to your LangChain run
chain.invoke({"input": "Hello!"}, config={"callbacks": [langfuse_handler]})

Python (Scoring a trace)¶

from langfuse import Langfuse

langfuse = Langfuse()

# Send a score for an existing trace
langfuse.score(
    trace_id="existing-trace-id",
    name="user-feedback",
    value=1,
    comment="Very helpful response"
)

Sources / references¶

Contribution Metadata¶

Last reviewed: 2026-04-26
Confidence: high