Skip to main content

​
General questions

Tuner is an observability, evaluation and simulation platform for voice AI agents. It scores every call against the evals you define, flags the calls that went wrong, and places test calls against your agent. See Welcome to Tuner.
Developers who integrate voice agents through the SDKs or API, and product and operations teams who own an agent’s quality and work in the dashboard.
Tuner integrates with your voice AI platform to automatically ingest call data.
  • For Retell and Vapi, you connect your platform key in the setup wizard and Tuner receives calls automatically.
  • For LiveKit, Pipecat and custom stacks, you add the Tuner SDK or send calls to the API.
Tuner then classifies each call, runs your evals and voice metrics, and shows the results in Call Logs and on the Overview dashboard. See the Quickstart.
Any of them. Tuner does not run your agent or plug into your platform’s internals. The only thing an integration does is send the call to Tuner, and once Tuner has the call it handles everything else.Which route you take depends only on what you build on:
  • Retell, Vapi, or Dograh: a guided setup in the dashboard, no code.
  • LiveKit or Pipecat: install that framework’s Tuner SDK.
  • Anything else: the custom stack route, using either the custom SDK or the Custom API.
What each route involves is under Technical Questions below.
It depends which side of the product you mean.Observability and evals work in whatever language your agent handles. Tuner analyzes the transcript you send it, and the eval judges read that transcript in the language the call actually happened in. An Arabic call is scored by the same evals as an English one, with no per-language configuration and no separate agent for each language. Eval prompts themselves can be written in English or in the language of the call.Call Simulation has a fixed list, because the Simulated Caller has to speak. It supports these languages and accents:See Languages and accents.So an agent that takes calls in a language outside the simulation list can still be fully monitored and evaluated. Only the generation of test calls is limited.
  • Overview metrics: total calls, success rate, failed calls, calls with red flags and red flag rate.
  • Call classification: call intent and call outcome for each call, set up on Agent Settings > Analysis Setup.
  • Issue detection: red flags from rules on evals, voice metrics, data fields and call metadata, such as a failed hallucination eval or high latency.
  • Call details: the transcript, eval results, voice metrics and extracted data for every call.
Yes. Build a Dataset from real calls: pick them from production, or upload recordings you already have, even on your first day. Tuner replays the caller’s side against your agent exactly as it was said, so you can reproduce a call that failed, prove your fix, and compare models or providers on your own callers.
On a replay of a Dataset call, Tuner compares the reference transcript of each caller turn with what your agent’s speech-to-text heard, and scores the difference. The result is a percentage with a Good, Fair or Poor rating on the call’s Voice Metrics tab. See Transcription accuracy.

​
Technical questions

Yes. Where your agent runs makes no difference, because Tuner never runs your agent and needs no access to your infrastructure.The whole integration is one idea: get the call to Tuner. Once Tuner receives it, everything after that is handled for you, including the transcript, per-turn timing, classification, evals, red flags, and alerts. There is nothing else to wire up.How you send the call depends only on what you build on:
  • Retell, Vapi, ElevenLabs, or Dograh: a guided setup in the dashboard. Add your platform key and pick your agent, or drop the Tuner node into your Dograh workflow. No code, and Tuner configures the receiving side for you.
  • LiveKit or Pipecat: install that framework’s Tuner SDK, LiveKit or Pipecat, and attach it to your session. It captures the call as it happens and sends it when the call ends.
  • Any other platform, or a stack you built yourself, for example Plivo, Bland, or Twilio: use the custom stack route. On Python, the custom SDK attaches observers to your own STT, LLM, and TTS and does the capture for you. In any other language, build the call payload yourself and POST it to the Custom API. Both arrive in Tuner as the same thing.
Self-hosted or managed changes nothing on any of these routes. Calls travel one way, out from your stack to Tuner over HTTPS, so Tuner needs no route into your network. Call Simulation is the one exception: to place a test call, your agent either has to be reachable at a SIP URI you give Tuner, or able to dial a SIP URI Tuner gives it.
In the API Reference tab. To authenticate, create a key on API Keys and send it as Authorization: Bearer <key>. See API keys and IDs.
  • Workspace ID: identifies your workspace. API calls send it as workspace_id.
  • Agent Remote ID: links an agent in Tuner to the calls you send for it. API calls send it as agent_remote_identifier. How it’s set depends on your provider; see below.
  • Workspace ID: on Workspace Settings > General Settings.
  • Agent Remote ID: on Agent Settings > Agent Setup, in the Agent Connection card.
    • Retell, Vapi and ElevenLabs: it’s the agent’s ID on that platform, taken from the agent you pick when creating the agent in Tuner.
    • LiveKit, Pipecat, Dograh and custom stacks: Tuner generates it when you create the agent.
See API keys and IDs.

​
Security and privacy

This is a different question from where your agent runs, and the two get mixed up often.
  • Your voice agent can run anywhere, including entirely on your own infrastructure. See the question above on connecting a self-hosted voice agent.
  • Tuner runs as a hosted cloud service on the Pay as You Go, Builder, and Growth plans. On-premise deployment is available on Enterprise, alongside SOC 2 and HIPAA compliance, custom integrations, and a DPA on request. It is set up with the Tuner team, so contact hi@usetuner.ai to talk through it.
If your concern is what Tuner stores, PII redaction replaces sensitive details in transcripts and data retention sets how long call data is kept before it is erased.
On the Growth plan, you can turn on PII redaction. Tuner then detects sensitive details in each call transcript (names, phone numbers, email and physical addresses, SSN / national ID, card numbers, and spoken PINs or passwords) and replaces them with labels like [NAME]. Recordings are not redacted. Each call shows its redaction status: Redacted, Partial or Failed.The setting is workspace-wide, so it applies to every agent in the workspace. See Data Redaction (PII) for the full walkthrough.

​
Billing and credits

Tuner charges in credits. Each analysis a call goes through uses credits, so you pay for what you run. Your plan sets your monthly credit allowance and which features you can use. Plans, prices and credit rates are on usetuner.ai/pricing. For what uses credits and how each is counted, see Credits and usage.
Top up at any time on Workspace Settings > Billing. See Credits and usage.
At zero balance, Tuner stops processing new calls and you can’t start simulation runs or replays until you add credits. See Credits and usage.
At the bottom of the sidebar and on Workspace Settings > Billing.

​
Next steps

Troubleshooting

Fixes for common problems.

Glossary

Definitions of the terms used across Tuner.