Assistant and Apple Intelligence

Choose your assistant in Settings › General: Apple Intelligence, a built-in model, a model server, or a cloud API provider.

Apple Intelligence is the default when the Mac can run it: Apple’s Foundation Models framework, the model behind Writing Tools. It needs macOS 26 with Apple Intelligence turned on, on a Mac that supports it.

The built-in model is the default everywhere else. It is a small language model that runs inside Cleat through MLX on the GPU; see below.

A local model server is for people who already run Ollama, LM Studio or llama.cpp and want Cleat to use whatever model they have loaded there.

Cloud / API model supports OpenAI, Anthropic Claude, xAI Grok and custom HTTPS Chat Completions endpoints. Supply a model ID and your own API key. Your provider bills API usage separately from Cleat.

When the palette offers it

Sums, unit conversions, dates, time zones, colours, definitions and the developer tools answer on their own and never involve a model; Instant answers lists what they accept. The assistant is only offered for what is left, and by default only when the phrase reads as a question: it opens with what, how, why, who, when, where, which, explain and so on, or ends in a question mark. open safari please never gets an Ask row.

Settings › Search › The assistant widens that to any phrase with no answer, or switches the row off entirely. Typing ask and a space always works, whatever the setting.

The built-in model

Cleat ships the engine, not the weights. Pick a model in Settings and click Download; the files come from Hugging Face once, into ~/Library/Application Support/Cleat/models, and after that the assistant works with the network cable out. Remove the model from the same place and the space comes back.

Three sizes, all 4-bit:

Model	Download	Good for
Qwen 2.5 0.5B	290 MB	Quick rewrites and short answers. Rough on facts.
Qwen 2.5 1.5B	880 MB	The default. Noticeably better writing.
Llama 3.2 1B	713 MB	Similar to Qwen 1.5B, a little more casual.

These are small on purpose. A launcher’s answer has to arrive before you look away, and a 1.5B model answers in a second or two on any Apple silicon Mac. For long documents or careful reasoning, use a real chat app; for “fix this paragraph” and “what is a monad”, this is enough.

The first answer after launch takes a few seconds longer while the weights load. They stay loaded until you quit.

A local server instead

Settings › General › Assistant › Local model server. Enter the address (Ollama’s is http://localhost:11434, LM Studio’s is http://localhost:1234) and click Refresh Models; Cleat lists what the server has and you pick one. Anything that speaks the OpenAI chat API works. If the address is not localhost, the Ask row says where your text is going.

Ask

Type ask (or ai) and a question. Return sends it to the model and the answer streams into the palette. Continue with a follow-up in the chat, copy an answer, or start a fresh conversation.

The empty chat offers starters so the first message is one click: Summarize the clipboard, Explain this error, Draft a reply, Rewrite professionally (each sends the clipboard’s text with the prompt) and Give me three ideas for…, which fills the field and waits. New Chat archives the conversation; the last ten sit under Recent chats in the empty state, and reopening one puts the words back on screen with a fresh model context.

Clipboard tools

These appear in the root palette and work on whatever is on the clipboard:

Command	What it does
Fix Grammar in Clipboard	Spelling, grammar and punctuation, same meaning and tone
Summarize Clipboard	Two or three sentences
Key Points from Clipboard	A short dashed list
Make Clipboard Professional	A clear, polite rewrite
Make Clipboard Friendly	A warm, casual rewrite
Explain Clipboard	A short plain explanation, code included

The result replaces the clipboard and is pasted where you were.

What is sent where

With Apple Intelligence or the built-in model, nothing. With a local server, the prompt goes to the address you typed, which is this Mac unless you changed it.

With a cloud model, Cleat sends prompts, text selected for clipboard actions and conversation context to your chosen provider. API keys stay in macOS Keychain and are excluded from backups. See API model setup.

Shortcuts

Ask Apple Intelligence is one of Cleat’s App Intents, so a Shortcut can send a prompt and use the answer.