hotato

Find what broke in your agent calls. Pin it so it never ships again.

pip install hotato
hotato check --demo             # a bundled failing call: what broke, and when
hotato check ./call.wav         # then your own recording
hotato vapi health              # or your last 100 Vapi calls

Zero config. Vapi, Retell, Bland, Synthflow, Millis, or local audio. Works with any recording: mono, dual-channel, or a transcript. Timing math, not a judge. Offline. Free. MIT.

hotato.dev

What it finds

• Say-do gaps: the caller interrupts to cancel (a barge-in), the agent says "canceled", the booking tool fires anyway. hotato takes the turn timing from the audio and the tool call from your OTel trace.
• Latency spikes: the pause before a reply going from 800 ms to over 2 s.
• Dead air: that pause reaching 5 s, or the line going quiet.
• Talk-over: the agent starts a fresh utterance over the caller.

Quickstart

Vapi

pip install hotato
export VAPI_API_KEY=...
hotato vapi health --last 7d --output report.html

Open report.html: every critical incident, timestamped, and your Voice Stability Score.

Retell

export RETELL_API_KEY=...
hotato retell health --call-id CALL_ID

--call-id is required and repeatable: you name the Retell calls to pull. hotato bland health, hotato synthflow health, and hotato millis health follow the Vapi shape. Those stacks mix both voices onto one channel, so they measure silence timing, dead air and latency gaps, each finding with its measured confidence; barge-in and talk-over need two channels.

Local audio

hotato autopsy ./call.wav

Writes a self-contained HTML report to hotato-output/, plus the JSON pin reads. Open the HTML in your browser.

From finding bugs to preventing them

autopsy turns one recording into timestamped incidents. scan reads a folder and tracks the trend. When you are ready, move a finding into CI: hotato pin turns one incident into a portable failure check, and hotato prove re-runs every stored check and fails the build rather than pass on evidence it cannot re-read. Every verdict carries its own evidence across five dimensions: outcome, policy, conversation, speech, reliability.

Pin a bug →

For continuous use: run hotato vapi health on a schedule, and open hotato console --production-db evidence.db to watch calls land live.

Wire it into CI

The exit code is the verdict: 0 pass, 1 fail, 2 refuse (could not tell).

# .github/workflows/voice-qa.yml
on: [pull_request]
jobs:
  hotato:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: attenlabs/hotato@v1.19.1
        with:
          contracts: contracts/
          hotato-version: 1.19.1

contracts/ holds what pin wrote. Full workflow: docs/CI.md.

Point your agent at it

Point Claude Code, Cursor, or any coding agent at this repo: it reads AGENTS.md and runs the loop end to end offline, no key. Over local stdio the MCP server adds the scorer plus read/verify/propose tools: uvx --from "hotato[mcp]" hotato-mcp (docs/MCP.md). Deploying is yours.

Nothing leaves your machine

hotato runs offline, on the machine that invokes it. The core is stdlib-only Python: no account, no key, no network call of its own. Your traces, prompts, and audio stay on your disk. Scoring with a language model you host yourself is a separate opt-in add-on, outside that core.

Go deeper

The whole loop, command by command: docs/LIFECYCLE.md. First touch to a CI gate: docs/GETTING-STARTED.md. Feed it what you already have: docs/CONNECT.md · docs/TRACE.md · docs/SIMULATE.md. What every verdict stands on: docs/EVIDENCE-CONTRACT.md. Next to the hosted alternatives: docs/COMPARE.md.

The deep toolkit -- capture, simulation, load, benchmarking, the fix ladder, the fleet control plane -- lives under hotato lab (hotato lab --help). The public commands are durable. hotato lab moves faster, and every command name that worked before 1.17 still runs unchanged.

Specifications

Property	Value
Footprint	~10 MiB installed, 0 runtime dependencies (stdlib-only)
Reproducibility	byte-for-byte: the same recording, the same report
Exit codes	0 pass · 1 fail · 2 refuse
Release integrity	OIDC Trusted Publishing + build-provenance attested
Runtime	offline, off the production data path

Verify the measurement yourself

PYTHONPATH=src python3 -m hotato.benchmark \
  --scenarios corpus/real/scenarios --audio corpus/real/audio

On 13 recorded AMI Meeting Corpus clips, the median error between measured caller-onset and the human word-alignment label is 20 ms. Provenance: corpus/real/README.md · method: METHODOLOGY.md.

Timing is measurable only when the two voices arrive on separate channels; a mono or mixed export is marked NOT SCORABLE and refused (hotato trust --stereo call.wav). The full four-tier evidence policy (what each verdict stands on, per input) is docs/EVIDENCE-CONTRACT.md.

Contribute

Issues and PRs welcome: CONTRIBUTING.md · SECURITY.md · CHANGELOG · docs/

License

MIT (LICENSE)

_{Know when to pass it on.}

mcp-name: io.github.attenlabs/hotato