Skip to content

Playground

Watch one request move.

Simulated in your browser. It does not reach a deployment, and the log is scripted rather than measured, so read it as a map rather than a benchmark.

Simulated. Nothing here reaches a real deployment.

Where work arrives

Four ways a conversation or a task starts. Each maps to a channel the deployment records against.

Embedded widget

channel: web

A script tag puts voice and text on any page.

Inbound calls

channel: phone

A number rings the agent instead of a person.

Proactive events

channel: agent task

Your systems report something; a trigger decides what happens.

Connected mailboxes

channel: email

Mail arrives and is worked without anyone opening it.

The deployment you own

One tenant, one database, no broker to operate. Everything above and below runs inside it.

Idle, still yours

Your deployment

single tenant

One tenant, one database, no broker to run.

What handles it

Realtime voice, background reasoning, retrieval, and tools. A request uses whichever it needs.

Realtime agents

gemini live

The voice on the line, interruptible mid-sentence.

Orchestrators

any provider

Longer work, on whichever model you prefer.

Knowledge and skills

retrieval

What it knows, and the procedures it follows.

Tools and integrations

mcp, both ways

Reaching other systems, and being reached by them.

What comes back

The result reaches the customer and the record of it stays, so the next contact starts informed.

Answered now

same session

The reply goes back on the channel it came in on.

Followed up

campaigns, triggers

Reminders and callbacks happen without a person queueing them.

Remembered

persistent

The next contact starts where the last one ended.

Event log

  1. Nothing yet. Run a scenario, or send from any card.
Honesty

What this deliberately does not show.

A simulator is convincing because it looks like the product running, which is exactly why it should not invent anything.
  • No throughput or latency figures. Real numbers depend on your hardware, your model choice, and your call volume, and a made-up millisecond count would be the least trustworthy thing on the page.
  • No conversion lifts or recovered revenue. What answering faster is worth depends on your current response time and your margins, which we cannot know.
  • Capacities are set by your licence, not by the simulation. Every card that mentions a ceiling points at the plan table rather than quoting a number the demo made up.