Framework Coverage#

Supafone Labs has a native S2S harness for direct speaking-model selection and a separate Supafone Supervisor integration surface. The native catalog contains gpt-realtime-2.1, gpt-live-1, gemini-3.1-flash-live-preview, grok-voice-latest, hydra-v1.0, and hydra-v1.1, with browser previews and Supafone-managed, Twilio, Telnyx, Plivo, and SIP phone adapters. Configured platform keys supply the model by default; account BYOK overrides are optional. See the native guide for readiness and feature limits.

Supafone Supervisor exposes fourteen audited runtime integrations. Every adapter converts provider-specific events into one canonical call state. When a runtime has a supported control channel, the same abstract Supervisor directive is compiled back into that runtime's native message or developer-owned context.

This page distinguishes five different claims that should never be conflated:

Support class	Meaning
Managed native control	Supafone owns the delivery path and sends the control through the managed call runtime
Native control	The provider exposes a documented live control accepted by its active session
Developer-owned context	The application owns the LLM or framework context and applies the compiled directive locally
Observation only	Supafone can normalize events, supervise, score, and report, but the provider exposes no universal live prompt-control channel
Explicit host hook	The component exposes an event transport, but the developer's agent must decide how to apply it

Five speaking-provider families#

The shared SupafoneS2S API covers Ultravox, OpenAI, Gemini, Grok, and Hydra. Ultravox uses the default managed runtime. The six native choices use the common S2S relay with explicit capabilities, including Hydra's lack of native transcript events. These five speaking families are distinct from the five phone transports and from the fourteen Supervisor adapters below.

Shared interface and provider classes

Direct Agent Factory realtime delivery#

All five Agent Factory speaking families support hosted Supervisor coaching. Ultravox keeps its existing deferred guidance path. Native OpenAI, Gemini, Grok, and Hydra use a shared call-scoped coach and return bounded guidance through model-invoked check_guidance tool results. The same relay handles browser and the five existing phone transports.

Hydra supplies model-reported context, not transcripts; OpenAI, Gemini, and Grok can supply provider transcripts. Coaching support, the agent's enabled setting, and Supervisor credential readiness are separate. See the native coaching contract.

Standalone Supervisor adapters#

The matrix below describes Supervisor adapters for existing stacks. It does not describe direct model hosting. The Agent Factory also has a native S2S transport for OpenAI, Google/Gemini, xAI, and Smallest AI/Hydra; that path is documented in Native Realtime Agent Factory.

Shared hosted workflow#

All five speaking families share the generated/custom stage planner, durable facts and tool receipts, Manager reasoning, specialist consultation and Supervisor coaching. Native calls additionally support public widgets, opt-in recording, configured carrier controls and opt-in native model handoff. These features have separate readiness checks and do not make every provider voice or carrier capability interchangeable.

See the hosted capability matrix for execution gates, recording, DTMF, transfer and remaining Ultravox-specific voice features. The matrix below covers the separate SDK adapters for stacks that developers already operate.

Runtime matrix#

Runtime	Support class	Supervisor delivery	Acceptance criterion
Supafone Agent Factory (managed)	Managed native control	Ultravox `user_text_message` with `urgency=later`	Managed call accepts the data message
Ultravox	Native control	Deferred `user_text_message`	Send Data Message returns HTTP 204
Vapi	Native control	System `add-message` through the live call `controlUrl`	Control request succeeds and the message enters live context
Retell	Developer-owned context	System entry in the custom-LLM WebSocket context	Entry exists before the next response is emitted
Bland	Observation only	No universal prompt-injection action	Events normalize without emitting an unsupported action
OpenAI Realtime	Native control	System `conversation.item.create`	Item-created or item-done event arrives without provider error
Grok Voice Agent	Native control	`response.create.instructions`	Provider emits `response.created`, then `response.done` or an error
Gemini Developer Live	Observation only	`clientContent` is ordinary conversation history	Default declared contract observes events; raw conversation context is not a privileged system instruction
ElevenLabs Agents	Native control	`contextual_update`	Socket remains healthy and the next turn completes
Deepgram Voice Agent	Native control	`UpdatePrompt`	Provider emits `PromptUpdated`
LiveKit Agents	Developer-owned context	`ChatContext.add_message` followed by `update_chat_ctx`	Persisted context contains the system entry
Pipecat	Developer-owned context	`LLMMessagesAppendFrame` with `run_llm=false`	Context aggregator retains the developer message
Cartesia Line	Explicit host hook	Custom metadata event; no default prompt action	Host agent explicitly handles the event
Inworld Realtime	Native control	System `conversation.item.create`	Item-added or item-done event arrives without provider error

GenericWebhookAdapter is the configurable extension path for proprietary systems. It is deliberately not counted as one of the fourteen audited runtimes.

Runtime selection versus voice selection#

Ultravox supports a separate compatible TTS voice in the managed Agent Factory. The native OpenAI, Gemini, Grok, and Hydra paths use model-native voices. Standalone TTS adapters in the package are not automatically available on live managed calls; Deepgram is currently preview-only in the hosted Ultravox catalog. See Ultravox + Custom TTS and Native S2S.

What the package covers around the runtime#

The voice runtime is one layer. Supafone Labs also normalizes the infrastructure developers otherwise assemble around it.

Layer	Supported surfaces	Pain removed
Agent runtimes	The fourteen integrations above plus generic webhooks	Rewriting supervision and state for every provider
Telephony	Supafone-managed, Twilio, Telnyx, Plivo, SignalWire, SIP/custom trunks	Separate number, carrier, webhook, and media-stream implementations
TTS	Supafone hosted, Cartesia, Inworld, ElevenLabs, Deepgram Aura, custom `TTSProvider`, deterministic fake	Provider-specific synthesis APIs and incompatible voice metadata
STT	Deepgram Nova-3 live multilingual tap, provider-native transcripts, Twilio/raw audio taps	Duplicate transcripts, missing language authority, and provider-specific event parsing
Supervisor LLM	Supafone hosted, Anthropic, OpenAI, xAI, custom `LLMProvider`, deterministic fake	Hard-coding the supervisor to one model vendor
Prompt programs	DSPy, LangChain, raw templates, provider-native message arrays, `PromptProgram`	Rebuilding optimization and prompt conversion per framework
Developer access	Python, TypeScript, Node, React/browser, REST, WebSocket, MCP	Maintaining separate product APIs for every application surface

Managed delivery versus adapter support#

The fourteen-row matrix describes audited event parsing and action compilation. It does not mean Supafone hosts every provider account automatically.

Supafone's default managed Agent Factory path currently uses its managed

Ultravox transport and managed provider accounts.

BYOK deployments can use the provider-native or developer-owned controls

listed above.

Bland remains useful for observation, post-call grading, QA, and telemetry,

but its documented live API does not expose a universal hidden-instruction channel.

Cartesia Line requires an explicit handler in the host agent. A custom event

is transport, not proof that the agent applied the instruction.

Unsupported or uncertain capability always degrades to no action. Supafone

does not invent a provider control.

Transcript and language authority#

For supported transcript-producing Supervisor integrations, Supafone selects one transcript authority per call. Native Hydra supplies no transcript events and its native adapter does not add a live transcription service. Optional recording-based Deepgram transcription runs after a native call; it does not supply live Hydra evidence. The separate Supervisor integrations use:

Hosted Hydra coaching uses labeled model-reported context and server tool outcomes; that context must never be displayed as a caller transcript.

Provider transcript for a supported monolingual stream.
Deepgram live tap when multilingual language authority is required and raw

audio is available.

Supervisor language inference only where a provider supplies transcript text but no

language tags and no raw-audio tap is available.

This prevents duplicate ingestion and conflicting language decisions. See Live language and voice routing for the opt-in hosted-agent behavior.

Release gates#

The public release verifies framework support at three levels:

tests/test_provider_injection_e2e.py runs all fourteen adapters from a

provider event through canonical state, Supervisor decision, and exact action.

tests/test_live_injection_contracts.py performs credentialed acceptance

probes where the vendor exposes a live test path. Missing credentials are skips, never passes.

the generated documentation parity checks requires this page to contain

every runtime from provider_contracts.py and rejects duplicate or stale matrix entries.

The provider contract registry includes the primary vendor documentation, acceptance behavior, verification date, and probe type for every row. The technical reference is also available in Providers and frameworks.

View raw Markdown