Bring the transcriber that fits your language. Bring the brain your agent actually needs. Bring the voice that holds the persona. Hyponema orchestrates them — and the conversation, the memory, and the persona stay yours, swappable per session.
ONE PLATFORM, YOUR PROVIDERS
Transcriber, model, and voice — all swappable, mid-conversation if you need to. We own the orchestration and the memory. That’s what makes the agent yours, not your provider’s.
If your main provider has a bad minute, the next one in line takes the turn. Your user never feels it.
You bring your provider keys, encrypted with your name on them. We never see them in plaintext.
Fastest, highest-quality, or cheapest combos out of the box. Tune later, per agent or per session.
VENDOR-NEUTRAL
MOVE AS FAST AS THE MODELS DO
Voice models leapfrog every quarter. Our customers don’t rebuild their agent when a competitor ships something better. They change one line and the next call uses it.
Save the change in the dashboard. The next session picks it up. No redeploy, no downtime.
Different agents in the same workspace can use different stacks; the same agent can pick a different combo per session via voice_stack overrides. No redeploy, no downtime.
Anything that speaks an OpenAI-compatible chat-completion API plugs in as a custom LLM. Custom LLMs do not participate in the cascade chain — they run as the single primary, no automatic fallback.
Architecture
| stage | provider | status | latency |
|---|---|---|---|
| stt | deepgram | 502 | 2400ms |
| stt | azure | 200 | 380ms |
When the main provider stumbles, the backup catches the turn. Set a main and up to two backups per layer. We record every attempt — which provider, how long it took, and why.
New provider in one save. Edit the stack, save it. The next session uses the new one. No deploy, no downtime, no break in the conversation.
STT, LLM, and TTS each have their own primary → fallback → fallback chain (max 3 retries per layer, independent). TTS fallbacks carry their own voice_id because voice catalogs aren't portable across providers; STT fallbacks may override the language for multilingual recovery.
Every step, every layer, on the record. How long the transcription took, the model took, the voice took. Every call, broken down. Open it in the dashboard or pull it through the SDK.
EVERY CALL ON THE RECORD
THE NUMBERS
The customer who gets the new model first doesn’t rebuild their agent. They change one line of voice_stack JSON.
Deepgram, AssemblyAI, Whisper for STT. Anthropic, OpenAI, Gemini for LLM. Cartesia, ElevenLabs, OpenAI TTS for voice. Custom OpenAI-compatible LLMs plug in as a single primary.
Persona, memory, and orchestration stay with you. Providers are interchangeable; the relationship doesn’t break when you swap one.
Each of STT, LLM, and TTS has its own independent primary → fallback → fallback chain. Every failover lands in error_event with provider, model, error_kind, fallback_overhead_ms.
The memory engine and the voice stack don’t depend on each other. Swap providers freely. The relationship survives.
AES-256-GCM envelope encryption. Per-credential DEK, KMS-managed KEK. Decrypted in-memory only for the seconds a call lasts.
Bring your own keys. Join the waitlist for early access.
