Langy for Everyone, and Voice Agent Testing

What You Can Do Now

Langy, your in-product assistant

• Langy now mounts for every team member of a project. No staff flag, no waitlist; if you can see the project, Langy is there.

• Moving between pages inside a project keeps the panel open, preserves your composer draft, and lets an in-flight response finish instead of resetting on every click.

• Each project gets its own dedicated Langy API key, provisioned automatically.

• Langy's own worker activity is traced into your project (tagged "langy"), so you can see exactly what it did.

Test voice agents (Scenario)

• The TypeScript SDK can now test voice agents end to end on a consolidated, clean stack.

• The OpenAI Realtime adapter moved to the GA Realtime wire protocol, and empty-content turns no longer trip the snapshot emitter.

• A new voice agents banner on Home points you straight at it.

Author trigger automations

• One staged authoring drawer replaces the old three-form split for building triggers, with a dispatch-health view so you can see what actually fired.

• Trigger notifications are now Liquid-templated, so you can customize the email and Slack output your automations send.

Map datasets faster

• Mapping dropdowns now offer every span, evaluator, and event-type name seen across the project in the last 30 days, not just whatever is in the trace in front of you.

• The span mapping dropdown is searchable, with clearer one-row-per-span labels.

Configure data retention

• Set per-category retention (traces, scenarios, experiments) at the organization, team, or project level, enforced with ClickHouse-native TTL. Narrower scopes override broader ones.

And a lot of polish

• Deleting an experiment now archives it (recoverable) instead of hard-deleting it across every store.

• The workflow code editor gained a light and dark Monaco theme, Python standard-library intellisense, and secrets autocomplete.

• Traces v2 drive-through pass: a facet-manager popover to choose your sidebar facets, near-live facet updates, and a smoother trace drawer.

• REST trace ingestion no longer rejects traces with more than 200 spans.

• Evaluator result popovers show the full reasoning and details again, for both scorers and LLM-as-judge.

• A unified langwatch claude command (plus codex, gemini, and opencode) gives one-step gateway and tracing onboarding.

• Workflow secrets resolve as secrets.NAME in code blocks and {{ secrets.NAME }} in HTTP-block URLs, headers, and auth.

Screenshots

A new voice agents banner on Home.

Workflow code editor: secrets autocomplete in the Monaco editor.

Dataset mapping now offers every project span from the last 30 days.

Per-tenant data retention: add a policy at org, team, or project scope.

Notes

• Under the hood we removed the Python langwatch_nlp service entirely; nlpgo is now the single execution engine. Fewer moving parts.

• A consolidated security sweep closed more than 40 dependency alerts across the monorepo this week (including 9 CRITICAL).

• More ClickHouse performance work: time-predicate fixes that stop cold scans and topic-clustering out-of-memory errors.