Automatic Tracing for LLM Applications

Add observability to your AI applications with one line of code. Tracium automatically traces OpenAI, Anthropic, Google AI, LangChain, and more.

Get Started View on GitHub

Quick Install

pip install tracium

One-Line Setup

Add tracing to your entire application with a single function call:

app.pypython

1
2
3
4
5
6
7
8
9
10
11
12
13
14
import tracium
 
# Initialize with auto-instrumentation
tracium.trace(api_key="sk_live_...")
 
# That's it! All LLM calls are now traced
from openai import OpenAI
client = OpenAI()
 
response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "Hello!"}]
)
# ^ This call is automatically traced

Why Tracium?

Zero Configuration

Auto-instrumentation for OpenAI, Anthropic, Google AI, LangChain, and LangGraph out of the box.

Web Framework Support

Built-in support for FastAPI, Flask, Django, and Celery with automatic request tracing.

Fail-Safe Design

SDK errors never break your application. All operations are non-blocking and fail gracefully.

Token Tracking

Automatic token usage tracking for all supported LLM providers. Monitor costs in real-time.

Context Propagation

Automatic context propagation across threads and async boundaries. No manual work required.

Queue Management

Background queue with configurable settings to prevent event loss under high load.

Supported Integrations

Tracium automatically detects and instruments these libraries:

OpenAI

Anthropic

Google AI

LangChain

LangGraph

Manual Tracing

For more control, use explicit traces and spans:

manual_tracing.pypython

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
import tracium
 
client = tracium.init(api_key="sk_live_...")
 
with client.agent_trace(agent_name="support-bot") as trace:
    # Record a planning span
    with trace.span(span_type="plan", name="analyze_request") as span:
        span.record_input({"query": "How do I reset my password?"})
        plan = analyze_user_request(query)
        span.record_output({"plan": plan})
    
    # Record an LLM call span
    with trace.span(span_type="llm", name="generate_response") as span:
        span.record_input({"prompt": plan})
        response = generate_response(plan)
        span.record_output({"response": response})
        span.set_token_usage(input_tokens=150, output_tokens=200)

Ready to get started?

Set up Tracium in your project in under 5 minutes.

View Installation Guide