What we've shipped

Blocking live lookup actions — configure a spoken hold message while your webhook runs, so callers always hear something while the agent waits for data.

Per-call latency ledger — every turn now records response-start, STT, LLM, TTS, and media queue timings for operator debugging.

Latency analytics dashboard — p50/p95 response-start charts and delay bucket breakdowns in the Analytics section.

Component-level timing badges in the browser test call transcript.

TTS session prewarm — Deepgram Aura-2 persistent session opens at call start, eliminating ~300ms setup time on the first agent turn.

Deepgram Aura-2 is now the default TTS provider — 329ms TTFB vs ElevenLabs, 90 voices with accent metadata and sample previews.

Persistent per-call TTS session — one WebSocket connection per call, dramatically reducing per-sentence latency.

Prompt caching — agent system prompt cached across turns, cutting LLM input cost and latency by ~40% after the first turn.

Live transcript chat view in the browser test call — scrolling conversation UI with per-turn timing annotations.

Provider-neutral LLM port — Claude is the default; Gemini and GPT-4o bake-offs can now be measured without changing the voice runtime.

LiveKit duck handler now non-destructive until confirmed barge-in — prevents queued audio being cleared on echo/background-noise VAD events.

Billing dashboard — current plan, usage meter, plan catalog with upgrade prompts.

Usage metering — every call is measured in billable seconds and settled against your plan's included minutes.

Plans catalog — Starter, Growth, Scale, and Enterprise plans seeded with NGN and USD pricing.

Twilio call transfers now correctly derive duration from started_at for billing accuracy.

Business intelligence dashboard — KPI cards, daily call/success trends, sentiment breakdowns, end-reason analysis, per-agent performance.

Actionable review queue — calls flagged by AI analysis for human review, with resolve/dismiss actions.

First-run onboarding panel — new workspaces see a guided checklist (create agent → test call → assign number) before the analytics view.

Call reliability — 15-minute sweep closes orphaned calls. Heartbeat tracking per turn.

Graceful end-of-call — built-in end_call tool with spoken farewell and clean session shutdown.

Per-agent knowledge base — upload PDF, DOCX, or TXT files, paste text, or add URLs. Content is chunked, embedded (Voyage AI), and retrieved per turn.

Post-call AI analysis — every completed call is automatically analysed for sentiment, summary, end-call reason, and success score.

Call transcript drawer — analysis panel with re-run button, sentiment badge, and success score in the calls list.

Conversation Layer v1 — natural turn length, backchannel discrimination, idle timer, latency filler, bounded TTS lookahead.

TTS is now genuinely streaming — ElevenLabs/Deepgram chunks forwarded immediately, no sentence-level buffering.

Actions catalog refactor — type discriminator with custom webhook, information extractor, and call transfer types.

Test-before-save gate on webhook actions — Save is disabled until a 2xx test response is confirmed.

Call Transfer action — cold transfer via Twilio with pre-transfer spoken message.

Information Extractor action — typed field extraction (string/number/date/enum) from conversation, saved as call variables.

Developer settings — API key management, server URL, webhook signing secret, egress IP display.

Phone Numbers — search, purchase, assign, and release +234 Nigerian numbers from the dashboard.

Inbound routing — calls to assigned numbers route directly to the correct agent.

Variable system — {{system.*}}, {{env.KEY}}, and action response mappings in prompts and action URLs.

Secrets vault — encrypted workspace environment for storing API keys used by webhook actions.

Agent list UX — cards are fully clickable; toggle between card view and list view.

Full dashboard — agents list, create/edit/delete, call logs, transcript drawer.

Browser test call — LiveKit WebRTC-powered in-browser test calls with real voice pipeline.

Agent settings — voice selection, call config, end-call reasons, knowledge, compliance.

Workspace auto-provisioning — Clerk auth wired to workspace creation.

STT → LLM → TTS → Twilio voice pipeline — full end-to-end voice call handling.

Deepgram Nova-3 STT with Nigerian English optimisation.

Claude Haiku 4.5 as the conversational LLM.

Twilio Media Streams WebSocket integration for phone call audio.

Agent CRUD API with workspace isolation.