# AI providers

URL: https://docs.vaultpanel.net/guides/ai-assistant/providers

> Choose which AI service answers in VaultPanel's assistant — Claude, an OpenAI-compatible API such as Ollama or vLLM, or a free model on the panel's server.

The assistant needs an AI provider to write its answers. Choose one under **System > Assistant Settings** (the `settings.manage` permission, which owners and admins have). People with the assistant permission can then use it.

| Provider | API key | Where your data goes |
| --- | --- | --- |
| **Claude (Anthropic)** | Required | The Anthropic API |
| **OpenAI-compatible** | Optional | OpenAI, or the service at the address you enter |
| **Free model on this server** | None | Nowhere: the model runs on the panel's own server |

## Set up a provider

### Choose the provider

Open **System > Assistant Settings** and pick a provider under **AI provider**. The one in use is marked **In use**; a provider with a saved key shows **Key saved**.

### Fill in its settings

The model, address and key, as described for each provider below.

### Test it

**Test connection** sends a tiny request with the settings in the form (saved or not) and shows which model replied and how long it took.

### Turn the assistant on

Tick **Turn the assistant on** and click **Save settings** (or **Save and use** the provider when switching). It cannot be turned on while a required setting is missing. When off, nobody can ask questions; past conversations stay readable.

## Claude (Anthropic)

The default and the most capable choice. It needs an API key from [console.anthropic.com](https://console.anthropic.com).

| Setting | Description |
| --- | --- |
| **Model** | **Claude Opus 5.5 (recommended)** (`claude-opus-5-5`), **Claude Sonnet 5.5 (faster, lower cost)** (`claude-sonnet-5-5`) or **Claude Haiku 4.5 (fastest, lowest cost)** (`claude-haiku-4-5`). **Another model…** lets you type any model id. |
| **API address (optional)** | Leave empty for `https://api.anthropic.com`. |
| **API key** | Required. |

Opus answers hard questions best; Sonnet and Haiku are faster and cost less. Requests time out after 2 minutes and are retried twice on network errors.

## OpenAI-compatible

OpenAI, or any service with an OpenAI-compatible chat API, such as Ollama or vLLM on your own server. The model must support tool (function) calling: the assistant works entirely through tools.

| Setting | Description |
| --- | --- |
| **Model** | The exact model name your service uses. |
| **API address** | Leave empty for `https://api.openai.com/v1`. For your own Ollama elsewhere, its address ending in `/v1`. The panel calls `/chat/completions` under it. |
| **API key** | Optional; sent when set. |

## Free model on this server

A small open model that the panel installs on its own server with Ollama. It needs no API key and nothing leaves the server; answers are slower and simpler than Claude's. Install a model on the same page first, then choose it as the **Model**. See [Free model on this server](https://docs.vaultpanel.net/guides/ai-assistant/local-model).

## Switching providers

Each provider keeps its own settings (since VaultPanel 0.3.0). Switching between Claude, an OpenAI-compatible service and the free model keeps each one's model, address and API key, so switching back needs nothing again.

## API keys

- API keys are encrypted in the panel's database with a key derived from `VAULTPANEL_SECRET_KEY`, and never shown again: the form shows only their last four characters.
- Leave **API key** empty to keep the saved key, or tick **Remove the saved key** to delete it.
- The audit log records only that a key was changed or removed, never the key.

> **Note:**
If `VAULTPANEL_SECRET_KEY` changes, stored API keys can no longer be decrypted. Enter them again.

## API

The same settings are available through `GET` and `PUT /api/v1/assistant/settings` and `POST /api/v1/assistant/settings/test`. See the [assistant API reference](https://docs.vaultpanel.net/api-reference/assistant).
