Traccia Documentation

Observe, evaluate, govern, and enforce policy on your AI agents in production.

Traccia is an AI agent control plane for observability, evaluation, governance, and runtime policy enforcement. Instrument once with the open-source OpenTelemetry SDK, then send traces to your own stack or to the hosted Traccia platform.

Enforce, not just observe. The same spans that show you what your agents did also power cost attribution, evaluations, policy checks, and audit evidence.

The Platform

Traccia is organized around four pillars that build on one another.

Pillar 1

Visibility

Full traces of every agent run, with sessions, tool calls, prompts, and errors captured from Python and TypeScript.

Traces Dashboard
Pillar 2

Intelligence

Accurate cost attribution under sampling, plus prompts, datasets, scorers, and experiments to prove quality before you promote.

Costs Experiments
Pillar 3

Control

Spend limits, retry caps, and custom rules enforced at the agent boundary so a policy decision blocks the run instead of logging it after the fact.

Policies Guardrails
Pillar 4

Certification

An AI system registry, reviews, incidents, and audit-ready evidence exports generated from the traces you already collect.

Governance Hub EU AI Act

Choose Your Path

The SDK is open source and works standalone. The platform adds the dashboard, evaluations, enforcement, and governance surfaces.

Use Traccia Platform

Hosted

The complete control plane: agent-first dashboards, cost attribution, evaluations, policy enforcement, governance evidence, and team collaboration.

  • Policy enforcement
  • Cost attribution
  • Governance and evidence

Use SDK Locally

Open Source

Export traces to your own observability stack. Works with Jaeger, Grafana Tempo, Zipkin, and any OpenTelemetry-compatible backend.

  • No account required
  • OTLP standard export
  • Full SDK control

How It Works

Traccia sits between your application and your backend, enriching traces with AI-specific metadata and routing them wherever you choose.

Your Agent CodePython or TypeScript
Traccia SDKobserve() and govern()
OTLP Over HTTPOpenTelemetry

Your Own Backend

Jaeger, Grafana Tempo, Zipkin, or any OTLP-compatible collector. No account required.

Local Setup Guide

Traccia Platform

Quick Example

Install the SDK, then add tracing to any function with the @observe() decorator. Platform users can add @govern() to enforce policies before a run starts.

bash
pip install traccia
agent.py
python
from traccia import observe, init


# Initialize Traccia
init()


@observe()
def my_agent(query: str) -> str:
    """Your agent logic here."""
    response = call_llm(query)
    return response


# All LLM calls are automatically traced
result = my_agent("What is the weather today?")

What You Get

Full Trace Visibility

See every step of agent execution, from prompts to tool calls.

Cost Attribution

Track token usage and costs per agent, model, and tool.

Policy Enforcement

Spend limits, retry caps, and custom rules applied at the agent boundary.

Governance Hub

AI system registry, reviews, incidents, and evidence exports.

Team Collaboration

Org-level visibility with role-based access and workspaces.

Auto-Instrumentation

Automatic tracing for OpenAI, Anthropic, and FastAPI.

OpenTelemetry Native

Export to any OTLP-compatible backend, with no framework lock-in.

PII Redaction

Optional pattern masking before spans leave your process.

© 2026 Traccia.