Setting up a provider

The AI assistant is off until you add a connection. Open Tools → Settings → AI, or click Go to AI settings from the assistant panel.

You can configure several AI connections and switch between them per query.

Fields

Field	Meaning
Connection name	How it appears in the picker
API URL	The provider's base URL
API type	openai, anthropic or gemini — this selects the request format
Username (Basic Auth)	Optional, for gateways that require basic authentication
API key / Token	Your key
Model	The model identifier
System prompt	Instructions prepended to every request
Send table schema	Include the structure of the selected tables
Send data sample (TOP 5)	Include up to 5 rows from each selected table

Providers

OpenAI-compatible

API type: openai
API URL:  https://api.openai.com/v1

The request goes to {URL}/chat/completions with a Bearer token.

Anthropic

API type: anthropic
API URL:  https://api.anthropic.com/v1

The request goes to {URL}/messages with an x-api-key header and API version 2023-06-01. Responses are capped at 4096 tokens.

Google Gemini

API type: gemini
API URL:  https://generativelanguage.googleapis.com/v1beta

The key is passed as a query parameter and the system prompt as systemInstruction.

Local models

Any server that speaks the OpenAI chat API works — set API type to openai and point the URL at it. For Ollama:

API type: openai
API URL:  http://localhost:11434/v1
Model:    llama3.1        # or whatever you have pulled
API key:  ollama          # required by the field, ignored by the server

The same applies to LM Studio, vLLM, llama.cpp's server and similar. Nothing leaves your machine in this configuration — which is the point, if your schema is confidential.

What gets sent

Read this before enabling schema and samples

• Send table schema transmits table and column names and their types.
• Send data sample (TOP 5) transmits real rows from your database — up to five per selected table, personal data included.

Both are off unless you turn them on, and they apply only to the tables you pick for a given question. With a cloud provider, that data leaves your machine and is subject to that provider's terms. With a local model it does not leave the machine at all.

Request preview in the assistant panel shows the exact JSON before you send it. When in doubt, look.

Checking it works

Open a SQL editor, click AI, pick the connection, ask something simple like "select the ten newest rows from users" and send. An authentication or model-name error comes back verbatim from the provider, which usually says exactly what is wrong.