Pretype

System-wide AI autocomplete for macOS.
Pretype shows a suggestion at the caret in each text field. A local model makes it. Your text stays on your Mac, unless you connect an API provider.

Features · Quick Start · Models · Privacy · FAQ

_{You type. The local model makes a suggestion. Tab accepts one word. ⇧Tab accepts all of it.}

---

Why Pretype?

Most autocomplete tools operate in one editor, and they send your text to a server. Pretype operates in all text fields on your Mac, and a local model makes the suggestions. You do not need an account, a subscription, or an API key. Pretype is a free and independent alternative to Cotypist.

Features

_{Typo fix. Tab applies the fix. Esc removes it.}

_{Rewrite. ⏎ accepts it. Esc keeps your original text.}

_{Settings show their cost. Hover over an option to see its effect on accuracy, speed, and memory.}

_{A style for each app.}

Keys

Key	Action
Tab	Accept the next word. With no suggestion on the screen, Tab operates as usual.
⇧Tab	Accept all of the suggestion.
⌥Tab	Fix the selected text. With no selection, fix the last word. You see a preview first.
⌥⇧Tab	Write a draft of your next message from the conversation on the screen. In the prompt of an AI agent, from the agent's session.
⏎ / Esc	Accept or reject a rewrite or a typo fix.
Hold ⌥	Dictate while you hold the key.
⌃⌥D	Start or stop hands-free dictation.

To reject a suggestion, continue to type. You can change the accept key to ⌘Space, ⌥Space, or ⌃Space in Settings.

What Pretype does

Feature	What it does
All text fields	Native apps, Electron apps (VS Code, Slack, Claude Desktop), and web pages. In Terminal, iTerm2, and the VS Code terminal, only the prompts that you type to an AI agent, for example Claude Code. There, the file names and the symbols that the agent shows complete at once.
Typo fixes	A pill shows above a misspelled word. The macOS spell checker finds the error in the language of the text.
Rewrites (⌥Tab)	The local model fixes grammar, typos, and phrasing. It keeps your tone. The best model fixes 69% of noisy lines exactly.
Reply drafts (⌥⇧Tab)	The local model writes your next message from the conversation on the screen. With the best model, 90% of the replies are on topic. An independent model rates them. In the prompt of an AI agent, the draft answers the agent's session. It reads the terminal text and needs no Screen Recording. With Offer replies in chats on, the draft comes when the agent finishes its turn.
Emoji and symbols	:shrug: gives 🤷, :rocket: gives 🚀, :cmd: gives ⌘.
Arithmetic	12*7= gives 84.
Speed	A suggestion takes 17–81 ms on the local models.
Your style	A persona, and a style for each app and each website. Pretype also learns from a local journal. You can clear the journal or disable it.
Shared app styles	Styles from the community. One click adds the styles for the apps on your Mac. To share your style, add one text file.
Hosted models (optional, off)	OpenAI, Anthropic, Google Gemini, Groq, Mistral, OpenRouter, DeepSeek, or a local server such as Ollama or LM Studio. Keys stay in your keychain. See Hosted models.
One switch for each feature	Suggestions, fixes, replies and dictation each have a switch in Settings → General. Typo fixes, emoji and arithmetic also have their own switches. With suggestions off, Pretype loads a model only for a fix or a reply.
Words (optional)	Pretype completes only the word that you type, from your own words and the macOS dictionary of the language that you type. No model loads. Select Words in Settings → General → Suggestions. On a Mac with 8 GB, Words is the default. When memory pressure unloads the model, suggestions use Words for two minutes.
18 interface languages	The interface uses the language of macOS. You can change it in Settings → General.

Choosing a model has the measurements for each model.

Where Pretype stops

• You can disable Pretype for one app or one website from the menu.
• To keep only the fixes, disable Suggestions in Settings → General.
• Pretype does not operate in password managers, in the shell, or while macOS reports secure input.
• Pause on Battery stops suggestions while the Mac is not connected to power.
• If you almost never accept the suggestions in an app, Pretype stops in that app. The menu shows a Resume item.

Dictation (off by default)

• Hold ⌥, say a sentence, and release the key. For longer text, push ⌃⌥D to start and push it again to stop.
• The speech models of macOS transcribe on your Mac. Audio does not go to the disk.
• A tidy-up pass adds the punctuation and the capital letters.
• Dictation uses the language of your keyboard layout.
• Names. Names and terms in the field, and in the session of an AI agent, come out as they are written there: "mlx" becomes "MLX", "suggestion controller" becomes "SuggestionController". This also operates with Parakeet. Add your own names and terms in Settings → Dictation → My words. Dictation always writes them as you wrote them there.
• Voice commands. Select text and dictate an instruction, for example "make it shorter" or "translate to English". Enable this in Settings → Dictation → Advanced.
• Send. Pretype can push Return after a dictation, for example in a chat. To enable this for an app, select Press Return after Dictation in the menu.
• If the focus changes during a capture, Pretype puts the transcript on the clipboard.
• History. To get a lost dictation back, select Copy Recent Dictation in the menu. The menu keeps the last five transcripts until you quit Pretype.

Languages	Speech model
English, German, Spanish, French, Italian, Japanese, Korean, Portuguese, Chinese	Apple SpeechTranscriber. It gives punctuation.
All other languages, Russian included	Apple DictationTranscriber. It is older and gives no punctuation. You can select Parakeet (one download of ~610 MB) instead, in Settings → Dictation when Pretype shows its interface in that language. On Russian, Parakeet makes less than half the word errors: 5.4% WER against 12.6% on FLEURS.
All languages	A hosted speech model, if you select an API provider in Settings → API.

Architecture → Dictation gives the details.

Privacy

Pretype must have the Accessibility permission. It uses it to read the focused text field, to catch the accept key, and to type the suggestion.

By default, your text stays on your Mac. Pretype uses the network only for model downloads and one version check each day. If you turn on an API provider in Settings → API, your text goes to that provider.

Do not trust these statements without proof. Three files read the field, watch the keys and type text back: AXText.swift, KeyTap.swift, and TextInjector.swift. Privacy & permissions lists the optional reads (screen, clipboard, microphone), each connection and all data on disk.

Quick Start

You must have macOS 26 or later and a Mac with Apple Silicon (M1 or newer). 8 GB of memory is sufficient.

brew install nikiomori/tap/pretype

Or download Pretype.app.zip, unzip it, and move Pretype.app to /Applications. Apple notarizes all releases since 2.5.0.

1. Open Pretype. It has no Dock icon. It operates from the menu bar.
2. Grant the Accessibility permission. If you grant it after the launch, restart Pretype.
3. Wait for the model download to complete. The menu-bar icon shows the progress.
4. Type in a text field.

The download of the default model is ≈1.5–5.3 GB. Pretype selects the default model from a fixed table of Mac memory sizes. The default uses less than a quarter of the memory. To use a different model, go to Settings → Model. Choosing a model compares the models.

To build from source, see Development setup.

How It Works

sequenceDiagram
    participant App as Focused app
    participant FT as FocusTracker
    participant Eng as CompletionEngine
    participant Win as SuggestionWindow
    participant Tap as KeyTap
    participant Inj as TextInjector
    App->>FT: Text at the caret (AX API)
    Note over FT,Eng: Wait 60 ms after the last key
    FT->>Eng: Prompt
    Eng->>Eng: Local model, then output gates
    Eng->>Win: Suggestion
    Win->>App: Panel at the caret
    Note over App,Tap: You push Tab
    Tap->>Tap: Catch Tab before the app gets it
    Tap->>Inj: Accept the next word
    Inj->>App: Synthetic key events

Loading

Architecture gives the details.

Documentation

You can read the documentation at pretype.app/docs. The source is in docs/.

Page	Contents
Choosing a model	The eval for the defaults, the RAM and latency of each tier, and the hosted-model presets
Privacy & permissions	Each permission, each connection, the data on disk, and how to remove all of it
Troubleshooting & FAQ	How to read Diagnostics, the usual problems, and the frequent questions
Architecture	The engines, the model catalog, the KV cache, and how the app decides to show no suggestion

Contributing

We accept bug reports, ideas, and pull requests.

• Open an issue for bugs and feature requests.
• Read the Contributing Guide and the Code of Conduct.
• Report security issues privately, as SECURITY.md tells you.

If Pretype is useful to you, give it a star (⭐). Stars help other people find the project.

Acknowledgements

Project	Use in Pretype
MLX and mlx-swift-lm	The on-device ML stack from Apple
Gemma	The default model on Macs with 16 GB or more
Qwen	The default model on 8 GB Macs
MiniCPM	One of the two fastest models in the catalog
swift-transformers	The tokenizers
swift-huggingface	The Hugging Face hub client and cache
Cotypist	The original idea

License

MIT. You can use Pretype free of charge for personal and commercial work. The bundled libraries and the downloadable model weights have their own licenses. See THIRD_PARTY_NOTICES.md.