# Assistant

> A product chat assistant: on the site it explains what Buff can do, in the app it tells you what is happening with your tasks.

The **assistant** is the chat in the bottom right corner. Ask in your own words — it answers in plain language and gives a link to where to look next.

It works in two places and knows different things in each.

## On the site

On the site and in the documentation the assistant answers questions about the product: what Buff can do, how to set your first task, how much it costs, how to connect a repository behind a firewall. Its knowledge comes from this documentation and the site’s pages, so its answers include links to the relevant sections.

No registration is needed.

## In the app

In the app the assistant also sees **your work** — read-only, it changes nothing:

- which tasks are in progress, which one awaits your decision, what is done;
- what is happening with a particular task and where its review is;
- whether the connected repositories are in sync with their sources;
- the organisation’s balance and plan.

It understands which screen you are on: “what’s up with this task?” on a task page means that very task. Links in its answers lead inside the app — to the task, its execution scheme, its review, the repositories.

A conversation in the app is billed like [“Ask”](/en/docs/ask) — by tokens, like any question to the model. On the site it is free.

## What it does not do

- **It does not answer off-topic questions.** Homework, other people’s code, texts “just because” — it politely declines and brings you back to the subject. Tokens cost money, and they are spent on questions about the product.
- **It changes nothing** — it does not start tasks, edit code or write to a repository. It is only a conversation.
- **It does not see the code of the Buff app** — only the documentation and, in the app, the state of your projects.

## How to use it

- **Start over** (the circular arrow in the header) opens a new conversation from scratch. The previous one stays in the history, but the assistant no longer takes it into account.
- The conversation survives a page reload: come back and you continue where you left off.
- An answer usually takes 15–30 seconds; in the app, where your data is checked, up to a minute. While the assistant is thinking, you can see what it is doing.

## What we keep

The conversation is stored: in the app — together with who asked; on the site — without being tied to a person. We read these conversations to understand what the product explains poorly and to improve the documentation. More in the [Personal Data Processing Policy](/en/legal/privacy).