Debug voice agents like software
Stop scrubbing recordings to find what broke. Every production call lands structured, scored, and searchable — transcript, per-stage latency, tool calls, Intents, and Outcomes. Then point your coding agent at it and ask why.
Call Records
Every call is a record you can query and debug.
Intent, outcome, extracted fields, cost, and eval verdicts are resolved on ingest. Open any call and see the same data laid out turn by turn — transcript, audio, per-stage latency, tool calls, and state transitions. Everything in the UI is available through the API.
Voice Metrics
Every silence has a source.
Tuner breaks every turn into the stages your voice pipeline actually runs — endpointing, transcription, model, and speech — so you can see where the silence came from.
Stage-level breakdown
Break every turn into endpointing, STT, LLM, and TTS and see exactly how long each stage took.
Endpointing measured
See when turn detection waits too long or cuts the caller off. Endpointing delay gets its own number instead of disappearing into overall latency.
See typical and worst-case latency
Compare p50 and p90 to tell whether latency is consistently high or a smaller number of calls are dragging the experience down.
Dead Air and Longest Monologue
See how long callers wait for a response and how long the agent speaks before giving the caller a turn.
Diagnose
Ask why. Get an answer.
Tuner exposes an MCP server. Point Claude Code or Cursor at it and ask what's breaking. It can inspect the affected calls, look for patterns across the full window, and write the evals and alerts that catch the problem next time.
Read the affected calls
Search logs, red flags, eval results, and stage latency across the full time window instead of relying on a sample.
Trace the problem to its source
Find the calls behind a pattern, understand what's going wrong, and see the evidence supporting the diagnosis.
Create the checks that catch it
Turn what you found into evals, call outcomes, and alert rules directly from the same conversation.
Dashboards
The patterns you'll never hear on a single call.
One call tells you something went wrong. Hundreds show you what's actually broken, what it's costing you, and whether the fix worked.
See success by intent
Break down outcomes across intents and see which types of calls consistently resolve and which ones fail.
Know the cost of every call
See what successful and failed calls cost, and understand how agent behavior affects your spend.
Compare versions side by side
Compare success rate, latency, cost, and other call metrics across agent versions before and after a deployment.
See what's driving failures
Rank red flags and failed evals by frequency to see which failure modes are having the biggest impact.
Traces
The same timeline, as OpenTelemetry spans.
Everything on the call timeline is emitted as standard OTel spans — turns, pipeline stages, tool calls, workflow steps — so voice appears in the same waterfall as the services around it.
Export standard OTel spans
Voice spans export to Datadog, Grafana, or any OTLP compatible backend, no custom shim.
Trace every turn
Capture the arguments, response, and duration for each turn as a span in the trace.
Keep workflow context
For graph-based agents, see state changes and workflow steps directly in the conversation trace.
Data Control
Keep the insight. Drop the PII.
Redaction runs before analysis and before storage. The unredacted transcript is never written.
Choose what gets redacted
Redact names, phone numbers, email and street addresses, card numbers, and other sensitive fields according to your policy.
Set your retention policy
Choose what data Tuner keeps and how long it stays available.
Apply it across the workspace
Set the policy once and apply it across every agent in the workspace.
Keep data where you need it
Store voice data in the EU, UK, or on-prem.