Integrations
Works with LangChain, CrewAI, OpenAI Agents SDK, AutoGen and LlamaIndex. OpenAI and Anthropic get automatic cost tracking. Exports via OTLP to Jaeger, Grafana Tempo, Zipkin, SigNoz and other OpenTelemetry backends.
Traccia 2.0
Observe, evaluate and hard-block AI agents across any framework
Advanced API
What is Traccia 2.0?
Traccia 2.0 is an AI agent control plane that unifies OpenTelemetry tracing, cost attribution, prompt evaluation, runtime policy enforcement and compliance evidence across frameworks such as LangChain, CrewAI and the OpenAI Agents SDK.
Traccia 2.0 is a control plane for AI agents running in production. It combines tracing, cost tracking, prompt and evaluation tooling, policy enforcement and compliance reporting in one place, so teams that operate many agents can see what each one is doing and constrain what it is allowed to do. The core is an OpenTelemetry-native Python SDK, installed with pip and started with a single init call. It traces LLM calls, tool use and agent decisions across LangChain, CrewAI, the OpenAI Agents SDK, AutoGen and LlamaIndex. Cost and token totals are computed locally when each span ends, so they stay accurate even when traces are sampled at 10%. OpenAI and Anthropic calls get automatic cost tracking, and other HTTP-based providers are traced through OpenTelemetry. The SDK can be used on its own, exporting to backends such as Jaeger, Grafana Tempo, Zipkin or SigNoz without an account, or paired with the hosted platform dashboard. The platform adds a registry of agents with health, environment and ownership, plus alerts for policy violations. Its guardrail engine detects problems through explicit annotations, provider-native signals and heuristics, and can hard-block an agent mid-execution when a spend cap, model restriction or tool call limit is breached. A prompt registry keeps immutable versions with a protected production label, while datasets, scorers (built-in, LLM-as-judge or custom) and saved experiments let teams compare a candidate prompt against a baseline before promoting it. Compliance exports map evidence to EU AI Act articles and optional HIPAA controls. It suits platform and ML engineering teams, and security or compliance owners, who have outgrown framework-specific tooling. Compared with LangSmith, which fits best alongside LangChain, Traccia positions itself as framework-neutral and focused on enforcement rather than alerting after the fact. Some enterprise items, such as SOC 2 Type II and SSO, are still in progress or on the roadmap.
How do you use Traccia 2.0?
1Install the SDK
Run pip install traccia in the Python environment where your agents run. No account is needed to use the open-source SDK locally.2Initialize tracing
Call traccia.init() once at startup. Traces from supported frameworks such as LangChain, CrewAI or the OpenAI Agents SDK begin exporting without further code changes.3Choose a destination
Send OTLP traces to your own backend such as Jaeger or Grafana Tempo, or connect to the Traccia platform for the dashboard and agent registry.4Set policies and redaction
Define rules such as a cost guard, restricted models or tool call limits, and enable init(redact_pii=True) to mask sensitive patterns before export.5Evaluate and promote prompts
Version prompts in the registry, run experiments against a dataset with scorers, and promote the winning version to production with evidence attached.
Pros and cons
Pros
- Policy engine can hard-block agents mid-execution, going beyond after-the-fact alertingAI
- OpenTelemetry-native SDK covers LangChain, CrewAI, OpenAI Agents SDK, AutoGen and LlamaIndexAI
- Cost and token totals stay accurate under trace sampling because they are computed locally at span endAI
- Open-source SDK can export to Jaeger, Grafana Tempo, Zipkin or SigNoz with no accountAI
- Prompt registry, datasets and experiments sit alongside tracing, with EU AI Act and HIPAA evidence exportsAI
Cons
- SOC 2 Type II certification is still in progress, which may slow enterprise procurementAI
- OIDC/SAML SSO is only on the roadmap, so single sign-on is not available yetAI
- PII redaction in the SDK is pattern-based; ML-based PII detection is only plannedAI
- Automatic cost tracking is stated only for OpenAI and Anthropic; other providers get basic tracingAI
- Pricing for the hosted platform is not published on the homepageAI
How much does Traccia 2.0 cost?
Features
Unified agent registry, OpenTelemetry tracing, cost attribution across 2,500+ models, 3-tier guardrail detection with hard-block enforcement, PII detection and redaction, prompt registry and playground, datasets, scorers and experiments, and EU AI Act and HIPAA evidence exports.
Frequently asked questions about Traccia 2.0
Does Traccia 2.0 have an API?
Yes, Traccia 2.0 offers an API.How do you use Traccia 2.0?
The walkthrough on this page covers 5 steps: 1. Install the SDK 2. Initialize tracing 3. Choose a destination 4. Set policies and redaction 5. Evaluate and promote prompts.What platforms does Traccia 2.0 support?
Traccia 2.0 is available on Web App.What does Traccia 2.0 integrate with?
Works with LangChain, CrewAI, OpenAI Agents SDK, AutoGen and LlamaIndex. OpenAI and Anthropic get automatic cost tracking. Exports via OTLP to Jaeger, Grafana Tempo, Zipkin, SigNoz and other OpenTelemetry backends.What are the limitations of Traccia 2.0?
SOC 2 Type II certification is still in progress, which may slow enterprise procurement. OIDC/SAML SSO is only on the roadmap, so single sign-on is not available yet. PII redaction in the SDK is pattern-based; ML-based PII detection is only planned.
