Voice Agent Lab · for teams building voice agents
Put your customer-service agent on a phone call with a simulated customer. Both talk on Gemini Live, as lip-synced faces with live captions. Your agent uses its tools, a judge scores the call against the scenario, and every call is recorded and transcribed.
Try it live starts a call in your browser. Load the demo plays back a recorded call, with its transcript, tool calls and scores. The public demo has a few limits.
crm.lookup_customer call.How a simulation works
What the customer wants, the hidden facts behind it, and the criteria that count as success. The caller and the back office know the hidden facts; your agent has to find them out.
Your customer-service agent takes the call, and a customer agent with its own persona and voice makes it. Both run on Gemini Live and speak one at a time, as they would on a phone line.
The CRM, identity checks, payments and the knowledge base are function calls. A simulated back office answers them, in line with the hidden facts and with its own earlier answers.
During the call, a judge ticks off each success criterion from the transcript and the tool results. At the end it gives the outcome, RESOLVED, ESCALATED or FAILED, and a one-sentence summary.
On the call
The server bridges two Gemini Live sessions. Each turn's audio streams to the listener, and the turn ends only once it has finished playing, so the agents never talk over each other. Personas, traits, pace and warmth shape how each one speaks. If your agent says "let me check" without calling a tool, it's nudged to call one.
server/src/call.ts
Each agent has a face from Wireface's face engine, wearing one of eight stock portraits or a picture you upload. The mouth follows the call audio, worked out in the browser from loudness, formants, frication and lip closures. A slider takes each face from wireframe to skin.
client/src/face/
Call any agent from your microphone: play the customer to test your support agent, or the support agent to hear how a caller behaves. Test calls use Gemini's own voice-activity detection, so you can interrupt, as a real caller would. They work in Chromium browsers; use headphones.
server/src/testcall.ts
After the call
Every turn, timestamped, with each tool call in line: the tool, its arguments, its result and how long it took. The history lists every call and batch, filtered by scenario, outcome and mode or searched by agent, and any session exports as JSON.
00:17AvaThanks Tom. Please give me a moment while I pull up your account.
crm.lookup_customer5009 ms
{"email":"TomHartley78@gmail.com"}
→ {"customer_id":17382,"name":"Tom Hartley","email":"TomHartley78@gmail.com","phone":"07700900078","account_status":"active"}
00:24AvaI've found your account, Tom. It looks like your account is active.
client/src/pages/History.tsx
A judge model reads the transcript, tool results included, and ticks each criterion with the time it was met. Saying isn't doing: identity counts as verified only once a check succeeds. At the end, it gives one of three outcomes and a one-sentence summary.
server/src/judge.ts
Realtime calls are saved as WAV, one track per agent, on the call's own timeline. Play one back with both faces lip-synced and a waveform for each side, with the tool calls marked. Click a line of the transcript, or a met objective, to jump to it. Rerun the call as it was, or with changes.
client/src/pages/Session.tsx
The same agents and prompts in text only, on the matching Gemini text model, many calls at once: up to 48 in a batch, 1, 2, 4 or 8 at a time, cycling through every customer agent if you like. Watch the resolved, escalated and failed counts add up, then open any run's transcript and scores.
server/src/turbo.ts
Pluggable tools
Your agent's tools are grouped into MCP servers, as they would be in production, and declared to Gemini as function calls. Each agent has its own choice of servers.
Out of the box, a simulated back office answers every call: a text model that knows the
scenario's hidden facts and the conversation so far, and keeps its answers consistent for the whole call.
To test against your own systems, implement ToolProvider, one method that takes a tool's name
and arguments and returns its result, and make real MCP calls behind it.
// server/src/tools.ts
export interface ToolProvider {
call(name: string, args: Record<string, unknown>):
Promise<unknown>;
}Agents and scenarios
It comes set up as account support for an online casino and sportsbook, with three support agents, four customers and six scenarios: logins, locked accounts, withdrawals, rejected documents, deposit limits and closures. Change any of them, or add your own.
Customer service
Ava LindqvistAccount support, Tier 1
Marcus ReidPayments specialist
Noor HaddadEscalations, Tier 2
Customers
Tom HartleyFrustrated returning customer
Margaret EllisElderly, low digital confidence
Priya NairCalm and concise
Luca RomanoNon-native English speaker
Add one with New scenario on the simulation page, or double-click a scenario to edit it.
Use cases
Changed a prompt, a model or a tool? Run a Turbo batch over the same scenario and callers, and compare the outcomes with the last batch in the history. A realtime call shows what text can't: pacing, turn-taking, how it sounds.
Duplicate an agent, change its model, system prompt, temperature or voice, and run both against the same scenarios and customers. Batches give you numbers; a realtime call lets you hear the difference.
Write customers who are impatient, confused, unsure of the language, or set on something policy doesn't allow, like the caller who wants a deposit limit raised during its cooling-off period, and see whether your agent holds the line.
Every call is saved as JSON, with its turns, tool calls and results, objectives and outcome, and every realtime call with a WAV track for each side. Export sessions to review, label, or build into a test set.
Show a voice agent at work before it is anywhere near a phone line: two faces on a call, captions, tool calls as they happen, and the score at the end. Or play back a call you recorded earlier.
Run it yourself
You need Node.js and a key for Gemini: a Vertex AI Express key, or a Gemini API key (the kind that starts
with AIza). npm install also fetches MediaPipe and the face landmarker model.
git clone https://github.com/compsmart/wireface-agent-call-simulator
cd wireface-agent-call-simulator
npm install
export VERTEX_API_KEY=your-key
npm run devIn PowerShell, set the key with $env:VERTEX_API_KEY = 'your-key'. Then open
http://localhost:5173. For production, npm run build and npm start: the
server serves the client too. npx tsx scripts/try-call.ts runs one call with no browser and
prints the transcript.
It runs as a single Cloud Run service behind Identity-Aware Proxy, so people sign in with their Google account and you choose who gets in.
/data keeps agents, scenarios, sessions and
recordings through scale-to-zero and redeploys.The public demo
The demo on this site is shared by everyone who visits, so it's kept small. For longer calls, bigger batches or your own agents and data, run it yourself.
What the demo stores and for how long: the privacy notice.
Start a live call in your browser, play back the recorded demo, or clone it and point it at your own agent.