Traccia

Traccia is an AI agent control plane built by Algen AI Private Limited, designed to give engineering and compliance teams full visibility, cost attribution, policy enforcement, and audit-ready governance over autonomous AI agents in production. The open-source Python SDK is licensed under Apache 2.0 and is listed as an external tracing processor in the OpenAI Agents SDK documentation. The hosted Traccia platform adds dashboards, evaluations, guardrail enforcement, and a Governance Hub on top of the same SDK instrumentation.

What It Is

Traccia positions itself as the "trust and governance layer" for autonomous AI systems. It sits between application code and any OTLP-compatible observability backend, enriching OpenTelemetry spans with AI-specific metadata: token counts, LLM costs, prompt versions, guardrail findings, and EU AI Act / HIPAA compliance evidence. The SDK instruments any Python function with a single @observe() decorator and auto-patches OpenAI, Anthropic, Gemini, LangChain, CrewAI, and the OpenAI Agents SDK with zero code changes.

Four-Pillar Architecture

The platform is organized around four layers that build on one another:

• Visibility — Full OpenTelemetry-native traces of every LLM call, tool use, and agent decision, with session grouping, error capture, and a unified agent registry across frameworks.
• Intelligence — Accurate cost attribution computed locally at span-end, independent of sampling rate; a prompt registry with immutable versioning; and an offline evaluation engine (evaluate()) that runs datasets, scorers, and experiment comparisons.
• Control — A three-tier guardrail detection engine (explicit annotation, provider-native signals, heuristic) that runs as a passive span processor. The @govern() decorator adds runtime policy enforcement that can hard-block agents mid-execution before unsafe actions complete.
• Certification — Article-mapped EU AI Act evidence (Art. 12, Art. 14, Art. 50), an opt-in HIPAA Controls module with PHI inventory and labeled exports, and one-click evidence pack exports from the Governance Hub.

Framework and Backend Integrations

The SDK auto-instruments the following frameworks when installed:

• LangChain — via a callback handler (pip install traccia[langchain])
• CrewAI — automatic crew, task, agent, and LLM span creation
• OpenAI Agents SDK — automatic detection and instrumentation
• AutoGen and LlamaIndex — traced via OpenTelemetry

For LLM providers, OpenAI and Anthropic include automatic token and cost tracking; Gemini (google-genai) is supported as of v0.1.28; any HTTP-based provider is traced via OpenTelemetry. The SDK exports to Grafana Tempo, Jaeger, Zipkin, SigNoz, or any OTLP-compatible backend with no account required.

Update: v0.1.28 — Gemini SDK Integration

The latest release (v0.1.28, published August 2026) adds Gemini (google-genai) auto-instrumentation, bringing Google's model family into the same automatic token and cost tracking pipeline as OpenAI and Anthropic. The repository was created in January 2026 and has seen continuous releases, with the GitHub project accumulating 102 stars and 20 forks as of the last update. The SDK is available on PyPI as traccia.

Deployment Model

Traccia offers two paths:

• SDK locally / self-hosted — Install pip install traccia, point the OTLP endpoint at any compatible backend (Jaeger, Grafana Tempo, etc.), and get full tracing with no account required. The SDK is open source under Apache 2.0.
• Traccia Platform (hosted) — Send traces to api.traccia.ai with an API key to unlock the agent-first dashboard, cost attribution UI, prompt registry, experiment compare, policy enforcement, and the Governance Hub. SOC 2 Type II certification is described as in progress; RBAC is available today, with OIDC/SAML SSO on the enterprise roadmap.

Why It Matters for Regulated Deployments

Traditional APM tools monitor infrastructure health; Traccia is purpose-built for AI-native systems where the risks are model-level: runaway LLM costs, PII leakage in traces, restricted model usage, and regulatory audit requirements. The policy engine's hard-block capability — stopping agents mid-execution rather than logging violations after the fact — and the built-in EU AI Act and HIPAA evidence generation address compliance requirements that generic observability platforms do not cover out of the box. The platform is built and operated by Algen AI Private Limited, headquartered in Bengaluru, India.