Skip to main content
Langfuse provides observability for LLM applications — traces, token usage, latency, and cost analytics. Octo integrates Langfuse as an optional callback on every LLM call.

​
Setup

1

Create a Langfuse account

Sign up at cloud.langfuse.com or self-host.
2

Create a project and get API keys

In the Langfuse dashboard, create a project and copy the Public Key and Secret Key.
3

Add to .env

LANGFUSE_ENABLED=true
LANGFUSE_PUBLIC_KEY=pk-lf-...
LANGFUSE_SECRET_KEY=sk-lf-...
LANGFUSE_HOST=https://cloud.langfuse.com
4

Restart Octo

On startup you’ll see “Langfuse tracing enabled” if configured correctly.

​
Environment Variables

VariableRequiredDefaultDescription
LANGFUSE_ENABLEDYesfalseToggle tracing on/off (true, 1, or yes)
LANGFUSE_PUBLIC_KEYWhen enabled—Langfuse project public key
LANGFUSE_SECRET_KEYWhen enabled—Langfuse project secret key
LANGFUSE_HOSTNohttps://cloud.langfuse.comLangfuse server URL (for self-hosted)

​
What Gets Traced

Every LLM call through the LangGraph supervisor and worker agents is traced, including:
  • Input/output messages — full conversation context sent to the model
  • Tool calls — which tools were invoked, with arguments and results
  • Token usage — input/output/total tokens per call
  • Latency — time per LLM call and per tool execution
  • Session grouping — all calls within a conversation thread are grouped by session ID

​
Session Tracking

Traces are automatically grouped by Octo’s conversation thread ID. Each session in Langfuse corresponds to one Octo chat session, making it easy to analyze full conversation flows.

​
Analyzing Traces

Common things to look for in the Langfuse dashboard:
Check the input tokens per call. If a simple question triggers 200K+ input tokens, context is bloated. Look at the input messages to identify large tool results or redundant memory.
Count the number of LLM calls per user message. More than 5-6 calls for a simple question suggests the agent is over-investigating. The supervisor’s efficiency instructions should prevent this.
Use Langfuse’s cost tracking to identify expensive patterns. Agent handoffs, deep research, and large context windows are the main cost drivers.
Identify slow tool calls or LLM responses. Bedrock cross-region inference can add latency — check if calls are timing out.

​
Disabling

Set LANGFUSE_ENABLED=false in .env (or remove it entirely). No code changes needed — Octo skips the Langfuse callback when disabled.
Langfuse is a core dependency (langfuse>=2.0 in pyproject.toml), so no extra install step is required.

​
Self-Hosting

For teams that need data privacy, Langfuse supports self-hosting. Point LANGFUSE_HOST to your instance: 
LANGFUSE_HOST=https://langfuse.internal.yourcompany.com