Release Notes
Month-by-month release notes for the Traccia platform: traces, cost, evaluation, runtime policies, the Gateway, and compliance.
What shipped on the Traccia platform each month, newest first. These notes cover the hosted app, the dashboard API, trace ingestion, and the Gateway. SDK releases that change what the platform can do are called out here too.
Looking For SDK Changes?
The Python and TypeScript SDKs keep their own version-by-version changelogs: Python SDK and TypeScript SDK.
Month In Progress
Cursor joins Claude Code as a coding agent you can trace, and cost totals hold up when exporters retry.
Observe
- Cursor Sessions As Traces. Cursor sessions show up as coding-agent traces. Each turn is priced from the model catalog and labeled as list-price cost, and a failed tool call no longer marks the whole turn as failed.
- No Double-Counted Cost On Retries. Ingest drops duplicate spans and billing events per tenant, so an exporter retry no longer inflates cost or call counts.
Policies moved from flagging problems after the fact to stopping them on the live call, and the new Gateway brings that to apps without the SDK.
Govern
- Preventive Policies On The Live Call. Spend Cap, Model Boundary, and Loop Cap are checked before each LLM or tool call under
govern(). Observe logs, Warn alerts, and Block stops the call. Detective policies still run after ingest. Requires Python SDK 0.1.29 or TypeScript SDK 0.1.14 or later. - Traccia Gateway. Point OpenAI, Anthropic, or Gemini clients at the Gateway with two Traccia headers, and Spend Cap and Model Boundary apply without the SDK. See the Gateway docs.
- Tool Guards. Refund Guard, Purchase Guard, High-Risk Tool, and Dangerous Shell Commands can deny a single tool call from the SDK.
- Four More Real-Time Policies. Freshness Guard, Retrieval Completeness, Unique Customer Cap, and Prompt Pin, all SDK only. Requires Python SDK 0.1.30 or TypeScript SDK 0.1.17 or later.
- Preventive Policy Wizard And Decision Log. Build a preventive policy step by step, declare which agents it covers, and see the decision for each call. The Decision Log shows whether the SDK or the Gateway made the call.
Evaluate
- Jev Decision Scorer. Score offline experiments with a decision rubric, using your own judge provider key.
Prove
- Richer Audit Context. Audit entries capture what was requested, not only who made the request and the status code.
Platform
- Delete Your Data Or Account. Delete workspace data or your whole account from settings.
- Card Payments With Tax. Pay for a plan in the app, with tax applied at checkout.
Evaluation got real scoring and comparison, policies got a lifecycle and alerts, and the app was reorganized around Observe, Evaluate, Govern, and Prove.
Evaluate
- LLM-As-Judge And Code Scorers. Grade experiment rows with an LLM judge using your own provider key, or with restricted custom Python. Batches run in the background with live progress.
- Experiment Compare. Put a baseline and a candidate side by side on the same dataset, with score, cost, and latency deltas.
- Datasets From Traces. Add real production spans to a dataset, and edit rows in place.
- Prompt Metrics. Prompt detail pages show usage by version and recent calls, joined from generation spans.
- Evaluate From Code.
evaluate()in the Python and TypeScript SDKs saves experiments to the platform, with eval traces marked so they stay out of production views. Requires Python SDK 0.1.27 or TypeScript SDK 0.1.10 or later.
Govern
- Violation Lifecycle And Email Alerts. Violations move through Open, Acknowledged, Resolved, and Auto Resolved, with a cooldown and opt-in email alerts.
- Draft Or Activate Policies. A policy wizard lets you save a draft, simulate the decision it would make, or activate it.
Platform
- New Navigation. The app is grouped into Observe, Evaluate, Govern, and Prove. Governance Hub is now Compliance Hub. The dashboard and agent pages were redesigned to put cost and activity first.
- Simpler SDK Setup. Prompt and eval runtimes are served through
api.traccia.ai, so the SDK only needs an API key. - Email Sign-Up And Coupon Codes. Register with email and password, and redeem a coupon from billing.
- TypeScript And Gemini On Integrations. The Integrations page includes TypeScript SDK snippets and Gemini setup.
Prompt management arrived end to end, from registry to playground to governed promotes, alongside a HIPAA module and AI trace summaries.
Evaluate
- Prompt Registry. Named prompts with immutable versions, a protected production label, role-based access, and runtime fetch from the SDKs.
- Prompt Playground. Compare up to three panels and models side by side, with latency, tokens, and estimated cost.
- Datasets And Experiments. Curate test cases, run a prompt across them in batch, and save results as evidence for a promote.
Govern
- Blocking With Context. Policies can block an agent, and blocked runs show which rule fired and why.
- Governed Promotes. Warn-first evidence checks before promoting a prompt, links between prompts, agents, and AI systems, a policy sandbox for Playground runs, and redaction when saving a prompt from a trace.
Prove
- HIPAA Hub. An opt-in module for PHI-capable agents, safeguard drafts, vendor BAA tracking, and CFR-labeled exports. Traccia does not sign a BAA yet.
- Auditor Packets. Export a prompt with its version history and evidence for a reviewer.
Observe
- AI Trace Summaries. Get a plain-language summary of a trace, with keyword search in the traces list.
- Token Counts Fixed. LLM token usage is no longer mistaken for a secret and scrubbed at ingest.
The compliance foundation landed, Claude Code became traceable, and the app got a unified design system.
Prove
- Governance Hub. AI system registry, reviews, incidents, evidence exports, audit logging, and retention controls.
- EU AI Act Module. An opt-in overlay with a FRIA wizard and EU-labeled evidence.
Observe
- Claude Code Tracing. Claude Code sends telemetry over OTLP, and sessions are grouped with billing-aware cost.
- Clearer Trace Status. A trace is marked failed based on its final LLM call, and duplicate agent rows are merged.
- Cost Filters. Filter the Costs page by time range. Local models no longer show a misleading cost.
Platform
- Unified Design System. Consistent components and cost formatting across the dashboard, agents, and the demo portal.
- AWS Marketplace. Subscribe to Traccia through AWS Marketplace.
- Join An Organization. New users can join an existing organization during sign-up.
A live demo portal so you can see Traccia working before you instrument anything.
Platform
- Demo Portal. Run sample agents from the app and watch their traces arrive, streamed as they run.
- Integrations Page Refresh. Updated setup steps for each supported framework.
Observe
- LLM Call And Token Totals. Trace summaries lead with LLM calls and tokens instead of raw span counts.
Costs became platform-computed and consistent everywhere, and guardrail coverage became visible.
Observe
- Platform Pricing Engine. Costs are recomputed on the platform from a maintained price list, with org-level price overrides, and shown as the primary cost in every view.
Govern
- Guardrail Posture. See which guardrail detectors cover your traces at the org and agent level, separate from policy violations, with the option to suppress noisy detectors.
Platform
- Self-Serve Sign-Up. Create an account without talking to us first.
Multi-agent traces got clearer, and you could start running agents from the platform.
Observe
- Multi-Agent Traces. Orchestrators are identified as such, and trace details show how many agents and tools took part.
- Accurate Agent Costs. Agent views no longer double-count span and metric costs, and date ranges are respected.
- Agent Playground. Run an agent from the app and jump straight to its trace.
Govern
- LLM Model Policy. Flag runs that use models outside an allowed list. Trace-level violations appear on the trace itself.
Platform
- Microsoft Sign-In. Log in with a Microsoft account.
The Traccia platform launched: hosted traces, agent metrics, workspaces, and the first cost and tool policies.
Observe
- Hosted Traces And Dashboard. Send OpenTelemetry traces from the SDK and explore them with trends, rich trace details, data export, and dark mode.
- Agent Identity And Metrics. Traces are attributed to agents set in the SDK, and OTLP metrics are ingested and charted per agent.
Govern
- First Policies. Cost policies and tool-call policies flag violations on ingested traces.
Platform
- Workspaces, Teams, And API Keys. Separate workspaces with roles, access control, and per-environment API keys.
- Onboarding For Solo And Enterprise. Guided setup for individual developers and for organizations.
- Plan-Based Retention. Trace retention follows your subscription.
Tell Us What To Build Next
Feature requests and feedback are welcome at support@traccia.ai or as a discussion on GitHub.
© 2026 Traccia.