New Navigation and Voice APIs support on the Gateway

What You Can Do Now

Find your way around

• Switch between LLM Ops, the gateway and governance from a product switcher and an icon rail, instead of scrolling one long sidebar.
• Search the project and scope selectors instead of scrolling them, and keep the same navigation down to phone width.
• Close a toast yourself. Toasts now appear at the bottom center, and the trace drawer toolbar takes less room.

Bill realtime voice

• The LangWatch AI Gateway brokers realtime voice sessions, so voice spend is billed to a virtual key and appears in usage like every other call.
• The LangWatch AI Gateway serves ElevenLabs' own speech and text routes.
• Scenario runs every JavaScript and Python voice session through the LangWatch AI Gateway, so simulated voice is attributed the same way.

Manage gateway keys

• Set a virtual key to expire on a preset window or on a date you pick, and jump from any usage view straight to its traces.
• Route to catalog models, with a provider handle per instance.
• Revoking a key removes every cap scoped to it, and you can rotate a provider credential without restarting the gateway.

Sign in with the permissions you already have

• langwatch login creates a user-scoped key that inherits your own permissions, instead of a broader one.

Follow your coding agents

• LangWatch captures Codex conversations and repository identity by default.
• The terminal session viewer opens at the end, keeps totals anchored, and no longer jumps while you scroll.
• One cost formula across every surface, with the agent-reported figure kept as a drift alarm. Every Codex session reports what it cost and links to its own traces.

Run and judge scenarios

• The judge can fetch remote traces, with per-turn trace propagation and a local dev tunnel.
• Set maxTurns and minTurns directly in the scenario editor.
• Pass secret run parameters, and read a batch completion signal that reports the real state.
• Runs that did not pass come back as results, not errors.

Operate the platform

• Dead letters have their own page: discard them, redrive in bulk, and read why each one failed.
• The event-sourcing page has a sidebar, and the migrations page supports per-migration enrollment, human names and targeted runs.

Govern your AI estate

• Databricks warehouse compute is attributed to the Genie question that ran it, and priced by the hours it ran.
• Pick a source from a vendor menu with plain-language groups, and open a source detail page that lists every event, cursor-walked.
• The governed SQL vocabulary is now LWQL (LangWatchQL).

Notes

• Shipped across langwatch 3.15.0 and 3.16.0, with typescript-sdk 1.8.0 and Scenario 1.3.0 for both JavaScript and Python.
• SCIM fixes worth knowing about if you provision through Entra: PATCH operation names are read without regard to case so offboarding works, and a group PATCH that names no members no longer empties the group.
• Anthropic cost reports were read as USD when the amounts are cents, a 100x overstatement, now fixed along with four adjacent puller defects.
• Security: the Dependabot alert backlog was triaged and cleared, 278 open alerts.