The built-in assistant

The Assistant app is a chat in the dashboard. Ask about the NAS in plain language (“how much space is left?”, “are all my apps running?”, “clean up the duplicates in my Downloads folder”) and the assistant finds the answer with the same tools as an MCP assistant: files in Places, apps, storage, accounts, shares, backups, updates and the audit log. An administrator’s assistant can also create, rename and remove Places (removing one keeps its folder and files). It acts as the account that chats, never with more rights.

Set it up

An administrator sets it up once, in Settings > AI access (or from the Assistant app, the first time):

• AI provider: OpenRouter, OpenAI, Anthropic, Mistral AI, Ollama, or any other service that speaks the OpenAI-compatible Chat Completions API. OpenRouter gives access to most models with one key. Ollama runs a model on your own hardware, so nothing leaves your network.
• API address: filled in for each provider. For Ollama on the NAS itself, http://127.0.0.1:11434/v1.
• API key: the provider’s key. It is encrypted on the server (with HSI_SECRETS_KEY) and never shown again; leave the field empty to keep it.
• Model: Check lists the provider’s models. Choose one that can call tools: the assistant does nothing without them.
• Who may use it: administrators only (the default), or every account.
• Read-only: the assistant looks at everything the account sees but changes nothing: no file, app or setting.
• Daily limit per account: the tokens an account may use per day (UTC), to keep the provider’s bill in check. Empty means no limit. Tokens are counted as the provider reports them (OpenRouter, OpenAI), or estimated from the length of what is sent and received.

Chat

The assistant starts with only a few tools, and loads the families a question needs (files, apps, storage, users, system) as it goes: sending every tool’s description with each call would cost thousands of tokens. A family loaded in a conversation stays loaded.

Conversations are kept per account (the 50 latest) and can be deleted. While the assistant works, each step it takes shows under its answer. Stop ends the answer at once and keeps what it had written. The assistant works on the server: leaving the chat, opening another app or closing the page does not stop it. Coming back to the chat opens the last conversation again and follows it where it is; a conversation still running is marked in the list. Once a request is approved or rejected, wherever it was decided, the assistant goes on with the decision by itself. Restarting HSI ends the turns under way. Below the message field, the tokens used by the conversation and by your account today (with the limit, if there is one). A busy or unreachable provider is tried again twice before the assistant gives up.

Creating folders and files, renaming, moving and copying happen at once. Deletions, file overwrites and every storage, app, account, share, Place, permission or update change wait for your approval: the request shows in the chat with Approve and Reject, and the assistant goes on once you decided. The same requests also show in the requests panel at the bottom right of the dashboard.

Privacy and safety

• Places closed to AI assistants (Settings > AI access, administrators): no assistant, built-in or over MCP, can list, read or change what they hold, whatever the account’s rights, so their files never reach an AI provider.

• What the assistant reads, such as file contents and logs, is sent to the AI provider. Choose a local model to keep it on your network.

• The assistant never goes beyond the rights of the account that chats, and it can neither change its own settings nor approve its own requests.

• File contents and logs are treated as data: the assistant is told never to follow instructions found in them, and what deletes or reconfigures waits for you anyway.

• Its actions are in the audit log (Monitor > Audit log), marked as the built-in assistant’s, and can be shown alone.