Aywa RuntimeDocumentation
Website Create account

Providers

Bring your own model, speech, telephony, and data plane.

Aywa Runtime routes the voice-to-voice loop through provider adapters while keeping credentials in the deployed runtime environment and preserving one call/session contract. Native adapters, compatible APIs, and customer adapters can coexist when a deployment needs them.

Recommended runtime profile

The runtime ships with production-oriented defaults for fast setup. Operators can preserve imported values, override model routes, or attach custom adapters when the deployment requires a different stack.

LLMNative and compatible model routes for main reasoning, tool decisions, warm-transfer assistants, and post-call analysis.
STTStreaming partial and final transcript events with endpointing tuned for phone conversations.
TTSLow-latency voice routes, fallback voices, progressive playout, and telephony-safe output.
Voice-to-voiceStreaming STT, model routing, TTS playout, barge-in, and media mixing inside the private runtime boundary.
Knowledge toolsDedicated model route for query-style knowledge workflows when the agent uses documents or KB tools.
TelephonyBYO SIP trunk credentials and BYO phone number resources.

LLM providers

The runtime accepts native and OpenAI-compatible model routes. Configure only what the deployment actually uses.

OpenAI

OPENAI_API_KEY, OPENAI_BASE_URL

Azure OpenAI

AZURE_OPENAI_API_KEY, AZURE_OPENAI_BASE_URL

Anthropic

ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL

Google Gemini

GOOGLE_API_KEY, GOOGLE_BASE_URL

OpenAI-compatible

Groq, Mistral, Together, OpenRouter, DeepInfra, Perplexity, Cerebras, xAI, Anyscale, Inflection, or custom routes.

Custom adapters

Use the same call/session contract for internal gateways or providers that expose compatible streaming semantics.

Speech providers

Speech-to-text

Deepgram, AssemblyAI, Gladia, Speechmatics, Soniox, and custom STT routes can be wired through the runtime config.

Text-to-speech

Primary runtime paths cover ElevenLabs, Cartesia, OpenAI, Deepgram Aura, Rime, LMNT, Azure Speech, and custom TTS hooks when validated for telephony output.

Speech env 
VOICE_PIPELINE_ENABLED=true
VOICE_PIPELINE_STT_ENABLED=true
VOICE_PIPELINE_TTS_ENABLED=true
DEEPGRAM_MODEL=nova-3
DEEPGRAM_ENDPOINTING_MS=250
ELEVENLABS_API_KEY=...
CARTESIA_API_KEY=...

Telephony and web media

Telephony is intentionally BYO. The runtime owns the voice session above the edge, while FreeSWITCH, SIP trunks, or WebRTC adapters handle transport.

BYO SIPSIP REGISTER/INVITE, inbound SIP edge, in-dialog controls, DTMF, RTP ingress, and RTP playout.
FreeSWITCHInbound hooks, lifecycle event hooks, transfer bridge, and lab originate adapter.
Twilio / Telnyx / VonageAdapter-path and webhook coverage for customer telephony integrations; native completeness depends on the deployment.
WebSocketBidirectional web call transport for low-level browser or service audio clients.
WebRTCLiveKit adapter path with bridge boundary for assistant audio, microphone ingress, and browser call experiences.

Data plane providers

Postgres

Documents, resources, logs, credentials, webhook jobs, and backup metadata.

Redis / Valkey

Active-call leases and job coordination for multi-node deployments.

S3 / R2 / MinIO

Recordings, stereo recordings, PCAP files, and durable artifact access.

ClickHouse

Append-only event mirror for p95/p99 latency, cost dashboards, and long-retention analytics.

Cost tracking

Calls can expose cost, costBreakdown, and costs[]. Configure the rate card to audit provider spend without adding an Aywa platform markup line to call minutes.

Rate card 
RUNTIME_COST_PLATFORM_PER_MINUTE_USD=0
RUNTIME_COST_RATE_CARD_JSON='{
  "stt": {"deepgram:nova-3": {"perMinute": 0}},
  "llm": {"openai:gpt-4.1": {"promptPer1M": 0, "completionPer1M": 0}},
  "tts": {"11labs:eleven_turbo_v2_5": {"per1KCharacters": 0}}
}'