# How it works

BlinkWrite's suggestions are good because they're grounded in **context**: the
actual conversation, thread, or document you're writing in. Here's how that
works, end to end.

## 1. Reading context

BlinkWrite reads the *text* of the active field and its surroundings: the
messages above your reply box, the email thread you're in, or the document
you're editing. Wherever possible it uses the macOS **Accessibility API**, the
same system feature assistive tools use. That surrounding thread or document is
the **context** BlinkWrite draws on to make suggestions sound right.

Importantly, BlinkWrite:

- **Reads only the text it needs** to make a suggestion.
- **Isn't monitoring your screen** or tracking your activity.
- Only looks at the app you're actively writing in.

For popular messaging and email apps, BlinkWrite understands the structure of a
conversation: who said what, whether it's a one-on-one or a group, and which
message you're replying to.

## 2. What gets changed

Reading context is broad; **applying** a suggestion is narrow. Completions,
corrections, and rephrasing only change the **current paragraph** — the block
of text where your cursor sits, bounded by hard line breaks (when you press
Return or Enter). BlinkWrite uses the wider conversation or document to decide
*what* to suggest, but when you accept one of these, only that paragraph is
updated. Everything else in your message or document stays exactly as you wrote
it.

## 3. Masking personal information on your device

Before anything leaves your Mac, an **on-device** step scans the text and masks
personal information: things like emails, phone numbers, credit card numbers,
addresses, and credentials. This happens locally, on your machine, and the
masking is reversible only on your device so the final suggestion reads
naturally. See [Privacy & security](/docs/privacy-and-security).

## 4. Generating a suggestion

Your masked **draft** and any masked **surrounding context** are sent securely
to BlinkWrite's **customized, self-hosted AI models**, running on our own
infrastructure. They never pass through a third-party AI provider. The model produces a suggestion
appropriate to what you're doing: a full reply, a completion, a correction, or a
rephrase. See [Models & languages](/docs/models-and-languages).

Your text isn't stored or logged, and it's never used to train models. The
suggestion is generated in the moment and then discarded.

## 5. Showing it in a floating popup

The suggestion comes back and appears in a small **floating popup** right next
to where you're writing. It's a preview you can accept with **Tab** or ignore by
continuing to type. Nothing is changed in your document until you accept.

## Putting it together

This is why BlinkWrite's help feels different from a basic autocomplete: it's
reading the *whole* situation (the people, the history, and the tone) and only
ever suggesting, never acting on your behalf until you say so.