AI chat
Chat with your documents using on-device models. No account, no cloud, no API key required.
Press *⌘* to open the chat pane (⌘L focuses its input from anywhere). Ask questions about the open document — “summarize this”, “does this doc cover X” — and get answers grounded in the actual text.
Everything runs on your Mac. The models run inside the app’s own process — no background daemon, no account, nothing sent anywhere.
First run
Chat needs a one-time model download:
| Model | Size | Used for |
|---|---|---|
| Fast | 3.4 GB | Most questions, inline completions |
| Deep | 4.9 GB | Rewrites and edit-heavy requests |
| Embedder | 340 MB | Workspace search and retrieval |
A setup card walks you through it. There is no model picker to fiddle with: a router classifies each query and picks fast or deep automatically. When the deep model reasons through a <think> block, it renders as a collapsible “Thinking…” section.
Editing with the model
- Apply edits with a diff. When the model proposes a rewrite, an “Apply edit +N −N” pill appears. Click it to review a color-coded diff before anything touches your document. ⌘Z works afterwards.
- Inline AI actions — select text in the editor, right-click → AI: Rephrase, Expand, Summarize, Fix Grammar, Translate, Explain Code. Replacements go through the same diff-preview sheet.
- Quick actions — one-click generation of a table of contents, YAML frontmatter, an outline, or suggested tags for the current document.
Working with the conversation
- Right-click any message: copy as markdown or plain text, regenerate, edit & resend, delete from here.
- ↑ in an empty input recalls your last message. Enter sends; Shift+Enter inserts a newline.
/opens slash commands —/clearresets the conversation,/reportfiles a bug report.- Export the whole conversation as a markdown file from the chat menu.
Dictation
Hold the fn key to dictate into whichever editor or input has focus; release to stop. ⌘⌥D toggles dictation from the keyboard. Speech recognition is Apple’s on-device framework — same privacy story as the models.
Tuning
Settings → AI exposes the system prompt, temperature, context-size cap, and chat retention (forever or session-only).