Home/Docs/Overview

Hotato docs

Hotato scores voice-agent turn-taking (barge-in, talk-over, backchannel, endpointing latency) from any call, offline, one command.

What it is

MIT, zero-install CLI plus one MCP tool. Per event: yielded or not, how fast, how long talked over. Pass or fail, non-zero exit for CI.

Measures energy over time only, not transcription, speaker ID, or intent. How it works covers the method.

Two ways to run it

The self-test proves the plumbing. Real verdicts come only from scoring your own recording.

bash

# 1) the built-in self-test (synthetic battery, zero input): proves the plumbing
uvx hotato run --suite barge-in

# 2) score your OWN call: capture auto-pulls the dual-channel recording, then scores it offline
uvx hotato capture --stack vapi --call-id <id>

Synthetic is a floor

Synthetic audio: a floor and regression guard, not production accuracy. Trust it only after running your own calls.

Where to go next

See it fail (demo)

A bad agent, caught talking over.

Watch →

Getting started

Sixty seconds to a real signal.

Start →

Capture your calls

Pull your recording, then score it.

Capture →

How it works

The energy VAD, signals, and honest ceiling.

Method →

MCP tool

An agent or IDE runs the eval.

MCP →

CI integration

Fail the build on a turn-taking regression.

CI →

Contributing

The best contribution: real, labeled call fixtures.

Contribute →

The honest scope, once

No accuracy %, benchmark, or leaderboard: reproducibility is the edge. Energy isn’t intent: no speaker ID, diarization, transcription, or emotion detection. MIT open core, never relicensed.