Traccia spans local SDK instrumentation and a hosted platform for agent observability, cost control, prompt management and evaluation, and governance. This page reflects what is available today and what we are building next.

Released

SDK (Python & TypeScript)

Python SDK (pip install traccia) and TypeScript SDK (npm install @traccia/sdk), both running on Node.js 16+
@observe(), manual spans, and span_scope() / spanScope() for streaming workflows
OpenAI and Anthropic auto-instrumentation with aligned llm.* attributes
LangChain, CrewAI, OpenAI Agents SDK, and Claude Code integrations
Token counting, local cost estimation (llm.cost.usd), and traccia pricing refresh
OTLP, console, and file exporters — including composite export in Node.js
Guardrail detection, governance enrichment, optional PII redaction, and EU AI Act trace overlay
govern() / @govern policy enforcement at the agent entry point, raising AgentBlockedError on a hard block
Works locally without an account; sends to Traccia platform or any OTLP backend

Traccia Platform

One hosted platform with three surfaces: see what ran, prove it is good, and keep it accountable.

Observability & Cost

Trace Explorer — Search, filter, and inspect agent runs with LLM cost context
Cost Attribution — Platform-recomputed costs by agent, model, team, and span
Agent & Dashboard Views — Per-agent health, spend, and volume across your workspace

Evaluation

Prompt Registry — Named prompts with immutable versions, version notes, tags, and a protected production label
Runtime Fetch In The SDKs — load_prompt() / loadPrompt() with caching, fallback, and prompt identity on traces
Prompt Playground — Compare up to three panels and models side by side with latency, tokens, and estimated cost; replay or save prompts straight from a trace
Datasets & Experiments — Curate test cases, run a prompt across them, and save graded runs as promote evidence
Scorers — Built-in checks, LLM-as-judge with your own provider keys, and restricted custom Python, reusable across experiments
Experiment Compare — Baseline versus candidate on the same dataset with panel pairing plus score, cost, and latency deltas
Datasets From Traces — Select From Traces on a dataset, or Add To Dataset from a span, with optional PII redaction
Prompt Metrics — On prompt detail: Editor / Metrics / Connections tabs; By Version and Recent Calls (paginated), joined by prompt id on generation spans
Policy Sandbox — Offline, detect-only policy check on a Playground run before you ship the change

Governance & Compliance

Governance Hub — Registry, reviews, incidents, and evidence exports
Guardrail Posture & Policies — Org- and agent-level guardrail visibility, policy rules, violations, and admin email alerts
Runtime Policy Enforcement — Wrap an agent entry point with govern() in the Python or TypeScript SDK; a hard-block policy with an open violation raises AgentBlockedError so the run never starts. Fail-open by default, and every block is recorded in the Policies UI
Governed Promotes — Warn-first evidence checks, override reasons in the audit log, linked agents and AI systems, prompt export, and auditor packets
EU AI Act Module — Opt-in compliance overlay, FRIA workflows, and EU-labeled evidence
HIPAA Controls Module — Opt-in PHI-capable agent registry, safeguard drafts, vendor BAA tracking, and CFR-labeled exports (a signed Traccia BAA is not offered yet)
Teams & Org Pricing — Multi-tenant access control and org-level pricing overrides

In Progress: Deeper Evaluation

Scoring, compare, datasets from traces, and prompt metrics are shipped. Next we are deepening score trends across experiments and SDK-native evaluate helpers.

Score Trends

Track pass rate per scorer across experiments so regressions surface early

SDK Evaluate

Run task plus scorers from Python or Node and optionally persist an experiment

In Progress: Active Governance

Enforcement today happens at the agent entry point: you opt in with govern() and a hard-block policy stops the next run. Everything else is observe and alert. We are extending enforcement into the platform itself, so controls apply without wrapping code and can act on spend and behavior mid-run.

Planned Capabilities

Halt Executions

Stop runaway agents in real time, without an SDK wrapper

Graduated Escalation

Move an agent from flag to soft block to hard block on your own timeline

Throttle Retries

Prevent retry storms automatically

Switch Models

Fallback to cheaper models when appropriate

Enforce Budgets

Hard limits on agent spend in real time

Gateway Enforcement

Apply policy at the proxy for agents you cannot instrument

Planned: Optimization Recommendations

Moving from detection → recommendation. Metadata-driven insights that help teams optimize without manual analysis.

Example Recommendations

"This agent is 2× more expensive than peers"

"Retries account for 38% of spend"

"Switch model for this tool-heavy step"

Privacy by Design

Recommendations are based on metadata only — no prompt inspection. We analyze cost, duration, retries, and tool usage patterns.

Why This Matters

For Executives

Savings, not charts. Actionable recommendations that directly reduce LLM costs and improve agent efficiency.

For Platform Teams

Justification, not guesswork. Data-driven insights to explain costs and defend optimization decisions.

Future Vision

Online Evaluation — Score live production runs on a sample, not just offline datasets
Evaluation In CI — Run experiments from your pipeline and gate prompt changes on results
Human Review — Queue outputs for reviewers and fold their labels back into scores
Alert Notifications — Slack, email, and webhook integrations for policy and cost events
Custom Dashboards — Build org-specific views and reports
Additional Framework Integrations — LlamaIndex, AutoGen, and more first-class adapters

Have Ideas?

We'd love to hear from you! Share your feature requests at support@traccia.ai or open a discussion on GitHub.

Roadmap