Skip to main content
Tuner is where voice agents get tested, watched, and improved. You send it your calls; it scores them against criteria you define, surfaces the ones that went wrong, and places test calls against your agent so you find failures before your customers do.

​
What problem it solves

Running voice agents in production without the right tools means:
  • Scattered logs — parsing endless unstructured output to find one issue
  • Manual call reviews — listening to calls by hand to judge quality
  • Manual self-testing — calling your own agent over and over, never sure you covered the edge cases
  • No real-time visibility — hearing about problems from angry users instead of your own systems
  • Guesswork — changing your agent without knowing whether it got better

​
The loop

Tuner is built around a continuous loop. You don’t just monitor production and react — you simulate, ship, observe, fix, and simulate again.
1

Simulate

Stress-test your agent with Call Simulation before launch and after every change. The same Evals that score production calls score every simulated call.
2

Observe

Every call arrives as structured data — transcript, per-turn timing, tool calls, eval results. Dig into individual calls in Call Logs to find what went wrong.
3

Monitor

Track performance in real time on the Overview dashboard, with real-time alerts for the things you can’t afford to miss.
4

Optimize

Fix the agent, confirm the fix against production data, and re-simulate to catch regressions. Diagnose with MCP finds root causes across every affected call at once.
Overview Dashboard

​
Three ways to connect

No-code platforms

Retell, Vapi, Dograh. Add a key, create an agent, done. Calls sync automatically.~5 minutes

SDKs

LiveKit, Pipecat, or your own Python stack. The SDK captures per-turn timing and latency your platform never exposes.~2–10 minutes

Direct API

Any language. You build the call timeline and POST it after the call ends.~30 minutes
SDK-level capture is the difference between knowing a call went badly and knowing which pipeline stage was slow. Where a platform hands you a finished transcript, the SDKs record what happened turn by turn — end-of-utterance delay, time to first token, time to first audio byte.

​
Vocabulary

The six things Tuner is built from. Each links to the page that covers it properly. Every other term used across Tuner is in the Glossary.

​
Ready to start?

Quickstart

Connect an agent and get your first call analyzed — about 5 minutes.