NEW Adversarial QA — test your agent A/B →
Real-time supervision for voice AI agents

Your voice agent needs a second mind during live calls.

Supafone Labs watches live calls and sends silent corrections to your agent so it can recover, clarify, escalate, and stay on-script — before the call goes wrong.

Voice agent mistakes are usually discovered after the call. Supafone Labs catches them during the call and silently nudges the agent before the next response.

Add Supafone Labs to your agent See provider integrations
new accounts get 5 minutes free — no card
Live on a call — watch the whisper land

One supervisor for every voice agent you run.

The problem

Testing catches known failures. Real callers create new ones.

The moment that matters is the one your prompt and your test suite didn't anticipate — live, mid-call.

"Our fee is usually around 30%…"
It made that up.
"Perfect — you're booked for Tuesday!"
The booking failed.
"Hola… ¿hablan español?"
It kept going in English.
Who it's for

Built for people whose agents talk to customers.

You run an agency

AI receptionists for law firms, clinics, home services. One invented fee or phantom booking costs you the client. Supafone Labs is the supervisor you can't hire for every call — with an audit log you can show them.

You build on Vapi, Retell, Ultravox…

Prompt-tweaking is whack-a-mole: fix Tuesday's failure, cause Thursday's. One line adds a second model that corrects live, QA that attacks your agent before customers do, and prompts that improve from graded calls.

You answer to compliance

Legal, medical, finance. "What stops it from saying the wrong thing?" finally has an answer — every correction timestamped, attributable, auditable.

Where it fits

A different moment than your QA stack.

Hamming, Cekura, and Coval are good products. They solve before and after the call. Supafone Labs solves during it.

Test-and-monitor toolsSupafone Labs
Pre-call simulationLive-call correction
Post-call monitoringIn-call intervention
QA reportsSilent agent coaching
Regression testingRuntime supervision
Find issuesFix the next turn

Keep your QA stack. Supafone Labs handles the moment QA cannot: the live call itself.

Use cases

The silent directive, by vertical.

Legal intake
"Do not imply legal advice. Collect facts and offer attorney follow-up."
Healthcare front desk
"Verify patient identity before discussing appointment details."
Sales calls
"Caller raised a price concern. Acknowledge budget before pitching."
Support
"User is frustrated. Apologize first, then summarize the next step."

Testing finds bugs. Monitoring flags failures. Supafone Labs improves the next sentence.

How it works

Listen. Reason. Whisper. Recover.

ListenWatches every transcript turn as the call happens.
ReasonDetects confusion, risk, missed context, or bad tone.
WhisperSends a silent directive through Vapi, Ultravox, ElevenLabs, OpenAI Realtime, and more.
RecoverThe agent uses the guidance — the caller never hears it.
Self-improving prompt
standing directive v1

Try it live in the Agent Builder, or put any agent through a graded call in the Agent Tester.

The one class

Ears in. Mouth out. One reasoning core.

You describe the shape of events coming in and the platform to whisper out to. Everything between is identical, on every stack.

import supafone_labs brain = supafone_labs.SupafoneLabs( provider="deepgram", # EARS · events IN inject_via="ultravox", # MOUTH · whispers OUT scenario="legal_intake", # prompt profile oracle_instructions="Never quote fees.", mode="return", # you deliver the action ) result = await brain.observe(raw_event) # result.actions[0] → the native whisper (or [])
Raw eventTwilio, Vapi, Retell…
EARS · provider=
Ingest adapterparses that provider's format
Reasoning coreidentical for every provider
belief_prompt
perception state
directive_prompt
coaching decision
MOUTH · inject_via=
Egress adapternative channel for your platform
result.actions[0]ready to send — never spoken

Same code runs three ways: hosted on one Supafone Labs key, on your own vendor keys, or fully offline on deterministic fakes. Full diagram (SVG) · the class, documented.

Under the hood

What it actually does.

One contract in, one whisper out — compiled to your platform.

raw event from Vapi

      

      

      

        
the brain — live belief state

        

          
emotiondistressed

          
intentfee question, fresh injury

          
languageen

          
risk caughtabout to quote fees

          
confidence0.86

        

      

      

      

        
native whisper to Vapi

        

      

    

  






  

    
One line.

    

      
shell

      
$ pip install supafone-labs[all]

    

    
import supafone_labs
brain = supafone_labs.supercharge(my_agent)

    
5 free minutes, no card. Every whisper lands in your console — timestamped, auditable.

  






  

    
Compatibility
Works with your whole stack.

    
Model providers

    
TTS providers

    
STT providers

    
Telephony providers

  






  

    
Pricing
Free to start. Minutes, not seats.

    

      

        
Open source

        
$0 forever

        
    
          
  • The whole thing, MIT
    • 
          
  • Bring your own AI keys
    • 
        

        GitHub
      

      

        
Cloud

        
5 min free then pay as you go

        
    
          
  • We host the models & voices — one key
    • 
          
  • Console with every correction
    • 
          
  • $10 → 400 min · $49/mo → 2,000 min
    • 
        

        Start free — no card
      

    

  







Powered by Supafone Labs