Supafone Labs#

Production infrastructure for voice agents that need to work after the demo.

Python SDK · TypeScript SDK · GitHub · Developer console · API reference

Why we built it#

A voice demo can be assembled quickly. A dependable voice product cannot. The production agent is split across a realtime model, telephony, TTS, STT, tools, retrieval, state, recordings, compliance, monitoring, and post-call workflows. Every vendor exposes a different event format, and the speaking model is still expected to notice and correct its own mistakes while talking.

Supafone Labs was built around the failures that appear at those boundaries:

Production problem	Supafone innovation	What changes for the developer
The speaking agent must supervise itself	Voice Watcher and SecondMind run beside the call and issue one bounded directive only when evidence is strong	Add supervision without replacing the agent or extending the audio hot path
Every voice platform has different events and controls	Canonical runtime plus 14 audited adapters normalize call events and compile guidance into the control each platform actually supports	Keep the current provider and reuse the same supervision, QA, and telemetry
Prompts make operational claims that tools never confirmed	Truth state and guardrail policies track verified bookings, transfers, deliveries, consent, and failures separately from model language	Prevent the agent from claiming an action succeeded before a tool proves it
Every new agent starts as another prompt-engineering project	Agent Factory turns a job description into editable stages, tools, routing, numbers, voices, and artifacts	Provision complete inbound, outbound, browser, and campaign agents through one API
Testing is manual role-play	Adversarial QA and SSR grading generate scenarios from the agent objective and compare supervised with unsupervised behavior	Measure regressions and supervision lift before deployment
Calls disappear into provider dashboards	Durable activity APIs retain agents, plans, calls, recordings, transcripts, watcher events, and post-call outcomes	Build one operational console instead of reconciling vendor logs
Multilingual calls lose context or use the wrong voice	Language-aware transcription and opt-in language/voice profiles preserve the active workflow while the language changes	Configure multilingual behavior without rewriting the agent graph
Phone, WebRTC, SMS, campaigns, and signing become separate systems	One SDK and one account model connect managed delivery, messaging, campaigns, artifacts, and writebacks	Stop rebuilding the surrounding product for every customer

The architecture#

mermaid

flowchart LR
    caller[Caller] --> agent[Speaking agent]
    agent --> tools[Tools and business systems]
    agent -. events .-> runtime[Canonical call state]
    tools -. verified outcomes .-> runtime
    runtime --> watcher[Voice Watcher / SecondMind]
    watcher --> gate[Confidence and policy gate]
    gate --> adapter[Provider adapter]
    adapter -. silent guidance .-> agent
    runtime --> artifacts[Transcripts, recordings, QA, analytics]

The call never waits for the Watcher. If supervision is unavailable, late, or uncertain, the gate emits no directive and the original agent continues.

Two ways to use the package#

Supervise an agent you already run#

python

import supafone_labs

supervisor = supafone_labs.supercharge(my_agent)
result = await supervisor.observe(provider_event)

The package auto-detects supported agents when possible, normalizes their events, and returns the provider-appropriate action. Start with Voice Watcher, then check the framework coverage matrix.

Provision the complete agent#

import { Supafone } from "supafone-labs";

const supafone = new Supafone({
  apiKey: process.env.SUPAFONE_TOKEN!,
  voiceWatcher: true,
});

const agent = await supafone.labs.agents.createInboundWithNumber({
  agentKey: "northline-intake",
  name: "Northline intake",
  description: "Understand the request and book the right next step.",
  number: { search: { areaCode: "415" } },
});

Agent Factory adds the plan, number, voice, stages, tools, call artifacts, and Watcher. Developers can inspect and edit the generated plan before creation.

Framework coverage#

The release gate covers fourteen runtime integrations, not a marketing-only logo list:

Native control: Supafone Agent Factory, Ultravox, Vapi, OpenAI Realtime,

Grok Voice Agent, Gemini Live, ElevenLabs Agents, Deepgram Voice Agent, and Inworld Realtime.

Context owned by the developer: Retell custom LLM, LiveKit Agents, and

Pipecat.

Observation or explicit hook: Bland and Cartesia Line.
Extension path: GenericWebhookAdapter for proprietary systems.

Integration depth is different for each provider. The complete framework matrix shows the exact control, acceptance criterion, and managed-delivery status for every runtime.

Package surfaces#

Surface	Use it for
Python	Local runtime, adapters, replay, supervision, STT/TTS components, and backend automation
TypeScript	Node, React, browser, Agent Factory, campaigns, activity, and product integrations
REST and WebSocket	Hosted agents, realtime services, events, recordings, transcripts, and custom clients
MCP	Agent creation, calls, QA, logs, and operational workflows from AI development tools

Start here#

Read the production problems.
Follow the quickstart.
Understand Voice Watcher and SecondMind.
Review all supported frameworks.
Install the Python or TypeScript SDK.
Choose managed delivery or BYOK.
Run the voice-agent QA workflow.

Supafone Labs exists so developers can define the caller experience, tools, and safety policy while one framework handles the infrastructure around them.

View raw Markdown