Build voice agents.
Observe humans.
Two products, one platform. Vox Workflow ships production voice agents from a visual canvas in minutes. Vox Observe scores every call you run and turns each finding into a coaching moment, across humans, AI agents, and the hybrid in between.
Five agents. One platform.
Pick an agent, glance at the sample script, and call it from your browser. All five are running on Vox Workflow — same canvas, same observability, same deploy pipeline.
Build it. Observe it.
The agent you ship and the QA that watches it speak the same language. That’s the whole pitch.
Build voice agents, visually.
From idea to phone call in three steps.
A drag-and-drop canvas where every node is a real runtime primitive. Pluggable LLMs, voices, and tools. One-click deploy to a phone number.
- 9 typed node types · LLM, agentic, HTTP, RAG…
- Sub-200ms turn latency on managed cloud
- LangGraph under the hood · self-host with Docker
Score every call. Coach the right ones.
QA, compliance, and coaching — unified.
Hybrid AutoQA (rules + LLM + golden examples) scores 100% of your calls. Compliance flags carry evidence and audit trails. Findings auto-become coaching clips and lessons.
- Human, AI, and hybrid calls — same rubric
- Real-time TCPA / FDCPA / custom rules
- Closed-loop coaching · no third-party LMS
A canvas designed for voice, not chat.
Branching, hand-offs, parallel tool calls, mid-conversation state. Modeled as a graph, compiled to LangGraph, served with sub-200ms turn latency. The canvas is the source of truth, and the source of truth is the thing that runs the call.
- Visual + code, both directionsExport to JSON, hand-edit, re-import. Round-trips cleanly. The canvas is the source of truth.
- 9 typed node typesInput · Output · LLM · Agentic · Code · HTTP · Hand-off · State Store · Knowledge Base. Tools are nodes too — no hidden panels.
- Conditional edges with state typingBranch on intent, on tool output, on any state field. Type errors surface in the panel, not at runtime.
- Versioned by defaultEvery save is a snapshot. Deploy snapshots. Roll back in one click.
Score every call. Coach the right ones.
Hybrid AutoQA does the scoring. Compliance carries the evidence. Coaching closes the loop. All of it lives in one place.
Universal coverage
Score human, AI, and hybrid calls on the same rubric, in the same dashboard. The only QA platform that treats AI voice agents as first-class citizens.
Hybrid AutoQA
Deterministic rules handle the objective items. A GenAI evaluator handles the judgment calls. Golden-example retrieval keeps it grounded. Confidence scores route the rest.
Compliance, evidenced
Real-time TCPA, FDCPA, state, and custom rule detection. Every flag carries a transcript excerpt, severity, confidence, and a full audit trail.
Closed-loop coaching
Every QA finding becomes a clip. Every clip becomes a lesson. Every lesson becomes an assignment with completion tracking. No third-party LMS.
The agent you ship on Tuesday is the agent QA reviews on Wednesday. Same graph. Same evidence. Same conversation, end to end.
Most QA tools weren’t built for what you actually run.
Bring your stack. Or use ours.
Swap any layer (STT, LLM, TTS, telephony) without touching your workflow. Pin model versions per node when you need to.
Powerful for engineers. Accessible for everyone else.
All the control of code. None of the infrastructure overhead. Built for developers, ready for business.
Sub-200ms turn latency
Streaming end-to-end on the managed cloud. Cartesia + Cerebras + Deepgram path tested under live traffic.
Natural conversation
Turn-taking, interruption sensitivity, filler words, and pronunciation tuning per node — the cadence reads human.
Live tool & data access
Ground responses in your authenticated APIs. Trigger tools mid-call. MCP support out of the box.
Custom model support
Bring your own LLM, TTS, or STT. Or pick from 30+ managed model presets across 9 providers.
Telephony control
BYO SIP, warm transfers, voicemail detection, phone-number management, DTMF capture — all wired.
Enterprise-grade uptime
99.95% SLA with automatic failover across model providers. Recovers cleanly from network blips.
Multilingual
English, Spanish, Mandarin, plus 100+ other supported languages — auto-detected or pinned.
Automated testing
Eval suites + fixtures catch regressions before they ship. Replay a saved call after every edit.
Tool calling
Intelligently fetch data and perform actions across CRMs, calendars, ticketing, and your own APIs.
Reporting & analytics
Dashboards and exports for performance, usage, latency, cost, and conversation quality.
Bring your own keys
Plug in your own provider keys for transcription, LLM, or TTS — pay providers directly at cost.
A/B experiments
Split traffic across prompt versions, voices, or full workflows. Win-rate dashboards built in.
Deep control. No lock-in.
Multi-agent orchestration
Chain agents or run squads with intelligent handoffs and routing. Specialist agents stay focused; the graph stays clear.
Modular platform
Swap models, voices, and logic independently — no vendor lock-in at any layer.
Model fallbacks
Automatic failover and latency balancing across providers. Your agent stays on the call even when GPT-4 has a bad minute.
API-first architecture
Every capability is reachable via REST, SDK, or webhook. Your team owns the integration end to end.
Making voice AI simple and accessible.
Every capability you see in the dashboard is one REST call away. Spin up a workflow, trigger an outbound call, stream live transcripts, tail compliance flags — all from the SDK of your choice.
- TypeScript, Python, Go, and Java SDKs — generated from one OpenAPI spec
- WebSocket stream for real-time call events; webhooks for everything else
- CLI (
vox deploy,vox logs --follow) for the terminal-friendly - MIT-licensed runtime — fork, audit, self-host
// npm install @vox-builder/sdk
import { Vox } from "@vox-builder/sdk";
const vox = new Vox({ apiKey: process.env.VOX_API_KEY! });
const call = await vox.calls.create({
workflowId: "wf_support_triage",
to: "+15802881314",
metadata: { customerId: "cust_001" },
});
call.on("transcript", (turn) => console.log(turn));
call.on("compliance", (flag) => alert(flag.rule));
call.on("end", (summary) => console.log("score:", summary.qa.score));Live in days. Coaching in weeks.
Four moves get you from idea to a measurable, auditable agent picking up the phone.
Build
Drag, connect, configure. Visual canvas with 9 node types. Type-checked end-to-end. No glue code.
Deploy
Ship to a phone number in one click. Managed cloud or self-host with Docker. Bring your own SIP trunk.
Observe
Every call — human, AI, hybrid — auto-transcribed, scored, and flagged for compliance in real time.
Coach
Findings auto-generate clips, lessons, and assignments. Agents learn what mattered. Auditors see proof.
Built for everyone who owns the conversation.
Filter by your role, your industry, or by the kind of agent you’re building.
Score every call.
Coach the right ones.
Stop sampling. AutoQA grades every call against your rubric, flags the ones that need a human reviewer, and routes the rest to auto-pass. Analysts only see what matters.
- 100% call coverage via hybrid rule + LLM scoring
- Cohen's Kappa calibration across reviewers
- Golden-example retrieval keeps the model on-rubric
- AI QA Reviews dashboard with agent trends and confidence bands
Pricing that scales with usage, not seats.
One subscription covers building and observing. No separate QA contract, no sales cycle to score your calls.
- 30 voice min / mo
- 1 deployed agent
- 100 calls auto-scored / mo
- All node types
- Self-host with Docker
- Community support
- 1,000 voice min included
- Unlimited agents & workflows
- AutoQA on every call
- Eval suites + call replay
- $0.07 / min overage
- Email support
- 5,000 voice min included
- 25,000 calls auto-scored / mo
- TCPA / FDCPA / custom rules
- Coaching: clip → lesson → assignment
- Priority support · 99.9% SLA
- SSO-ready, audit log
- Volume voice pricing
- Unlimited AutoQA & retention
- Dedicated infra · data residency
- SOC 2 · HIPAA · SSO · RBAC
- Custom integrations & SLAs
- Dedicated CSM
AutoQA is included on every minute. No separate quality contract, no add-on, no per-agent seat fee. Self-host stays free forever. Bring your own LLM / TTS keys to pay providers directly at cost.
Vox is the visual builder layer most teams end up wanting on top of those platforms — plus the QA layer none of them ship. We give you the canvas, the node library, the test harness, the deployment runtime, and the AutoQA engine — and you can self-host the whole thing.
Stop wiring SDKs.
Start building agents.
Free tier, no credit card. Bring a recording and we’ll score one of your real calls live in the demo.
Book a demo. We’ll score one of your real calls live.
Tell us a little about your team and we’ll set up a 30-minute call. Bring a recording if you have one — we’ll run it through AutoQA on the demo so you see what your current tool missed.
- 30-minute call with an engineer who built the product, not a sales rep.
- Live AutoQA scoring against your own rubric, on your own call.
- If you can’t make a meeting, we’ll send the AutoQA whitepaper instead.