Guides

Connect a provider

Add a chat, embedding or voice (TTS/STT) provider configuration, including custom OpenAI-compatible endpoints.

Providers are configured once per scope (organization or project) and then reused by any number of agents, flows or knowledge bases.

1. Open the providers screen

Organization-wide providers live under Global AI Providers; project providers under Project → AI Providers. Pick the tab for the model type (Chat, Embedding, Text-to-Speech, Speech-to-Text) and click Add provider.

2. Choose a provider type

ProviderModel field example
OpenAIgpt-4o, gpt-4o-mini-tts, gpt-4o-mini-transcribe
Anthropicclaude-sonnet-4-5
Google Geminigemini-2.0-flash
Mistralmistral-large-latest
Ollamallama3.1
Azure OpenAIgpt-4o (+ deployment name)
Deepgramnova-3 (STT), aura-2-thalia-en (TTS)
ElevenLabseleven_flash_v2_5
Azure Speechyour Speech resource model (+ region)
Google Cloudchirp_3 (+ service-account JSON)
Custom (OpenAI-compatible)any model name your endpoint serves

The Model name field is a dropdown of the models known for the chosen provider and model type. If your model isn't listed — a new release, an Azure deployment name, or a custom endpoint — pick Other and type the name instead. Custom providers have no catalog, so the field is free text.

For Text-to-Speech and Speech-to-Text also choose the delivery mode (streaming or batch — see AI Providers) and, optionally, a default language and voice. The voice list is loaded from the provider after the configuration is saved.

3. Enter the API key

The key is submitted once and immediately encrypted. It is never returned by any API response afterward — if you need to verify a key is correct, use the test endpoint rather than trying to view it.

4. For custom/self-hosted endpoints

Selecting Custom reveals a Base URL field. This is how you connect a self-hosted or any other OpenAI-compatible endpoint — set the base URL to your endpoint and choose the model name it serves.

The address must be reachable from the platform and use a full host name (for example https://llm.example.com/v1); internal service names and loopback or link-local addresses are rejected. When a call fails, the connection test explains the likely cause (credentials, model name or address) without showing the provider's raw response.

5. Test the connection

POST /v1/providers/{id}/test

This performs a lightweight call to confirm the provider is reachable and the key is valid, without exposing the key itself: a short completion (chat), an embedding, a one-word speech synthesis (TTS) or the transcription of half a second of silence (STT). Azure Speech and Google Cloud are checked against their own token endpoints.

6. Use it

The provider now appears in the selector when creating or editing agents (for chat models) and knowledge bases (for embedding models) — the same consistent selection experience is used for both.

On this page