Langfuse Tracing for OpenAI Agents

Drop in Langfuse's OpenAI wrapper to capture prompts, completions, latency, and token usage for every agent LLM call — then nest tool steps under one trace.

langfuseobservabilitytracingopenaiopentelemetry

Setup time15-45 minutes (estimate)

Estimated costLangfuse Cloud Hobby is free (50k units/month); Core starts at $29/month. Self-host is free of Cloud fees. You still pay your OpenAI (or other) token prices — see langfuse.com/pricing and your provider's pricing page.

ToolsLangfuse, OpenAI, Python

VerificationLast verified 2026-10-16

Use Langfuse as the tracing backend for agents that already call the OpenAI Python SDK. The drop-in wrapper records each chat completion; wrap tool steps with start_as_current_observation so one user request becomes one nested trace. For dollar budgets and hard spend caps, see Agent Cost Monitoring with LiteLLM. For a tool-agnostic span checklist, see Agent Tracing and Evaluation Setup.

Requirements

A Langfuse Cloud project (Hobby plan is free) or a self-hosted Langfuse instance
Public and secret API keys from the Langfuse project settings
Python with the OpenAI SDK (openai) already in use

Install and credentials

pip install langfuse openai
export LANGFUSE_PUBLIC_KEY=pk-lf-...
export LANGFUSE_SECRET_KEY=sk-lf-...
export LANGFUSE_BASE_URL=https://cloud.langfuse.com   # EU; use us/jp/hipaa URLs if your project is there
export OPENAI_API_KEY=sk-...

Drop-in OpenAI tracing

from langfuse.openai import openai
from langfuse import get_client

completion = openai.chat.completions.create(
    name="agent-chat",
    model="gpt-6-luna",  # re-check OpenAI's current model list before shipping
    messages=[
        {"role": "system", "content": "You are a support agent. Be concise."},
        {"role": "user", "content": "Summarize ticket 4821 in one sentence."},
    ],
    metadata={"agent": "support-bot", "ticket": "4821"},
)
print(completion.choices[0].message.content)

# Short-lived scripts must flush or traces can be lost
get_client().flush()

Open the Langfuse UI → Traces after the run. You should see the generation with input, output, model, latency, and usage.

Nest tool calls under one agent trace

from langfuse import get_client, propagate_attributes
from langfuse.openai import openai

langfuse = get_client()

with langfuse.start_as_current_observation(
    as_type="span",
    name="support-ticket-agent",
    input={"ticket_id": "4821"},
) as root:
    with propagate_attributes(
        user_id="end-user-99",
        session_id="ticket-4821",
        tags=["agent:support"],
        metadata={"env": "staging"},
    ):
        with langfuse.start_as_current_observation(
            as_type="span", name="fetch-ticket"
        ) as tool_span:
            ticket = {"id": "4821", "body": "..."}  # your CRM/API call
            tool_span.update(output={"chars": len(ticket["body"])})

        with langfuse.start_as_current_observation(
            as_type="generation",
            name="draft-reply",
            model="gpt-6-luna",
        ) as gen:
            completion = openai.chat.completions.create(
                model="gpt-6-luna",
                messages=[
                    {"role": "system", "content": "Draft a short reply."},
                    {"role": "user", "content": ticket["body"]},
                ],
            )
            reply = completion.choices[0].message.content
            gen.update(output=reply)

    root.update(output={"reply_preview": (reply or "")[:120]})

langfuse.flush()

Cost note

Langfuse Cloud Hobby includes 50k billable units/month (traces, observations, scores) with 30-day data access and 2 users. Paid plans start at Core $29/month. Self-hosting the open-source stack is free of Langfuse Cloud fees; you still pay your LLM provider. See Langfuse pricing. Do not treat Hobby limits as guaranteed forever — re-check the pricing page.

Sources (official docs, checked 2026-10-16)

Failure modes

Short-lived scripts that exit without get_client().flush() (or shutdown) can drop buffered traces.
Logging full prompts stores PII and secrets in Langfuse; redact or mask before export, and restrict who can view the project.
Wrong LANGFUSE_BASE_URL (EU vs US vs JP vs HIPAA) sends credentials to the wrong region and auth fails or data lands in the wrong project.
This is tracing and usage visibility, not a hard spend cap; pair with LiteLLM budgets (Agent Cost Monitoring) when you need enforcement.

Markdown version: /blueprints/langfuse-openai-tracing.md

Want a deeper build of this?

Vote for an extended, fully tested version (more tools, a full workflow file). Votes decide what we build next.

Related blueprints

AI Agent WorkflowsNot yet verified

AI Lead Follow-Up and Appointment Booking Agent (GoHighLevel)

Text every new lead within minutes, let a Conversation AI bot qualify them and book a call on your calendar, then hand off to a human, all inside GoHighLevel.

Moderate · 1-2 hoursgohighlevelcrmagency

AI Agent WorkflowsNot yet verified