Yaksh calls your voice agent over a real phone line with simulated callers, judges each call against your criteria, and shows what went wrong.
Find out how your voice agent behaves on real calls before your customers do.
Yaksh calls your AI voice agent over a real phone line, plays a realistic caller, and tells you what went wrong, quoting the transcript lines that prove it.
When to use it
- You test your agent by calling it yourself a few times before each release.
- You need to catch failures that only show up in volume: a missed confirmation one call in five, talking over a caller in a noisy street, long silences after hard questions.
- You want to test every change to your agent’s prompt, model or voice stack.
What you get
- Real phone calls through Plivo, so you test what callers hear: speech recognition, latency, interruptions and hang-ups.
- Personas with their own voice, language and background noise. Callers speak English (US, UK, India, Australia), Spanish, Hindi, Bengali, Gujarati, Kannada, Marathi, Tamil and Telugu.
- Judges that check each call against your criteria and quote the transcript. If the simulated caller makes a mistake, the test is marked blocked, not failed.
- Metrics with confidence intervals: pass rate, response time p50 and p90, talk-overs and dead air.
- Agent import from Plivo CX, Vapi, Retell or ElevenLabs, or any agent with a phone number. Inbound and outbound agents.
- A CLI and skill so coding agents such as Claude Code and Codex can write test cases and run tests.
Get started
You need Docker, a Plivo account with a phone number, and API keys for an LLM provider (OpenAI or Azure OpenAI), Deepgram and ElevenLabs.http://localhost:8000, add your provider keys and Plivo credentials under Settings, add your agent and a scenario, and start a run. Plivo needs a public HTTPS address for the worker: for local tests, run scripts/tunnel.sh.
Read the Yaksh docs on GitHub →