Concepts

AI Providers (Chat, Embedding, TTS & STT)

A single, unified way to connect any AI provider — chat, embedding, text-to-speech and speech-to-text models.

UnderOcean uses a unified provider layer as the single integration point for every model provider — chat completion, embeddings, text-to-speech (TTS) and speech-to-text (STT). Agents, flows, and knowledge bases all configure "providers" the same way, regardless of which vendor is behind them.

Providers live on the Global AI Providers page (organization-wide) and on each project's AI Providers page, grouped in tabs by model type: Chat, Embedding, Text-to-Speech and Speech-to-Text.

Supported provider types

ProviderModel typesExtra field
OpenAIChat, Embedding, TTS, STT—
AnthropicChat—
Google GeminiChat, Embedding, TTS, STT—
MistralChat, Embedding, TTS, STT—
OllamaChat, Embedding—
Azure OpenAIChat, Embedding, TTS, STTDeployment name
AWS BedrockChat, Embedding—
Custom (OpenAI-compatible)Chat, Embedding, TTS, STTBase URL
DeepgramTTS, STT—
ElevenLabsTTS, STT—
Azure SpeechTTS, STTRegion
Google CloudTTS, STTService-account JSON instead of an API key

Choosing Custom exposes a base URL field, so any self-hosted or OpenAI-compatible endpoint can be plugged in without code changes. For voice, this covers servers that implement OpenAI's /v1/audio/speech and /v1/audio/transcriptions (for example self-hosted Whisper or Kokoro servers).

Voice providers: delivery mode

Every TTS and STT provider has a delivery mode:

ModeHow it runsProviders
StreamingThe provider's realtime API; audio/transcripts flow while the user or the assistant is still speaking. Lowest latency.STT: Deepgram, Azure Speech, Google Cloud. TTS: Deepgram, ElevenLabs, Azure Speech, OpenAI, Google Cloud
BatchOne request per utterance (STT) or per sentence (TTS). Simpler, roughly 1–1.5 s slower per turn.STT: OpenAI, Azure OpenAI, Deepgram, ElevenLabs, Gemini, Mistral, Custom. TTS: OpenAI, Azure OpenAI, ElevenLabs, Gemini, Mistral, Custom

A voice provider can also store a default language, a default voice (TTS; picked from the provider's own voice list where it offers one, or typed in) and, for Azure Speech, the region.

Platform models

Providers are yours: you configure them per organization or project and pay for their usage. Separately, a platform administrator can operate platform models that organizations never configure themselves — today they power the AI assistant. A platform model is assigned to an organization with a monthly token quota, its credentials are encrypted and never exposed, and it does not appear in your provider lists or agent model dropdowns.

Security

  • API keys are encrypted at rest and never stored or logged in plaintext.
  • API keys are never included in any API response — the encrypted value never leaves the backend.
  • Each provider configuration can be tested without exposing the underlying key to the client.

On this page