Configure an AI provider

Before anyone can chat or run an agent, a project needs at least one configured AI provider. Elitea supports seven provider kinds, and you configure each one the same way: pick a type, fill in its connection details, save.

1. Open AI providers

   Go to Settings → AI Providers (/settings/model-configuration). Configurations are grouped by kind: LLM, embedding, image generation, speech-to-text and text-to-speech.

2. Pick a provider type

   Click Create and choose a type, or go directly to /settings/create-configuration/<type> (for example open_ai, azure_open_ai, ai_dial, anthropic, ollama, amazon_bedrock, vertex_ai, or vllm).

3. Fill in the connection fields

   Fields vary by provider — see the table below. Give the configuration a display name for identification, then save.

4. Add a model

   Once the provider connection exists, add the specific model it serves (display name, the exact model identifier the provider expects, context window, max output tokens). The model then appears in the model selector across chat, agents and pipelines. The most recently added model becomes the default for new conversations.

Provider fields

OpenAI, Anthropic, Ollama, vLLM

Field	Notes
API key	Required for OpenAI and Anthropic; optional for Ollama and vLLM, which are commonly self-hosted with no key
Endpoint (api_base)	Defaults to the provider’s public API; override it for a self-hosted or proxied endpoint

Azure OpenAI / EPAM DIAL

Field	Notes
API key	Required
Endpoint (api_base)	Your Azure resource URL, or your DIAL host
API version	Optional for DIAL, required for Azure (for example 2024-02-01)

Amazon Bedrock

Field	Notes
AWS access key ID	Required
AWS secret access key	Required
AWS region	Required, for example us-east-1

Google Vertex AI

Field	Notes
Google Cloud project	Required
Location	Required, for example us-central1
Service account JSON	Required — paste the whole service-account credentials document

Note

The field names above are the LLM gateway’s own credential keys, not display labels this form invents — they’re what actually gets sent through when a model call is made, which matters if you’re troubleshooting a configuration that saves but doesn’t resolve.

Warning

Never paste a real API key or service-account document into a screenshot or shared document. Treat every field above marked as a key or secret the same way you’d treat a password.

An administrator can also configure providers globally, shared across every project, from the admin console’s LLM Proxy section — useful for a team-wide default model that individual projects don’t need to configure themselves.