Hotato

hotato is self-hosted conversation QA for voice agents: give it both channels of a recorded call, it measures the timing between the two voices, and locks every catch into a CI contract. MIT, offline, no account.

Get started

Point your coding agent at it. Paste this repo at Claude Code, Cursor, or any coding agent and ask it to try hotato on your voice agent. It reads AGENTS.md and drives the whole loop: run the demo, wire a CI gate, and re-check the numbers itself.

Or run it yourself:

uvx hotato demo --fail                 # catch a failure in 10 seconds, no account
pipx install hotato                    # keep it in a project
uvx --from "hotato[mcp]" hotato-mcp    # drive it over MCP, local stdio

Five dimensions

• Outcome: job done, on tool-call and state evidence.
• Policy: required disclosures, PII handling.
• Conversation: did the agent yield when the caller took the floor, and how fast.
• Speech: response latency and turn timing.
• Reliability: pass@1 / pass@k / pass^k with a Wilson interval.

A call comes in two channels (caller on one, agent on the other). A mono or bad export is marked NOT SCORABLE, so a verdict measures timing, not intent.

License

MIT (LICENSE)

mcp-name: io.github.attenlabs/hotato