Cotypist

System-wide predictive autocomplete for Apple Silicon Macs that finishes your sentences inline, runs a local Gemma model, and never sends text to the cloud.

Cotypist is a Mac app that predicts your next words in almost every text field and completes them inline. You type, a grey suggestion appears, you press Tab to accept the whole line or just the next word. It runs a local language model on Apple Silicon, so nothing you type leaves the machine. Think GitHub Copilot's ghost-text completions, but everywhere you write instead of only inside a code editor.

Why it exists

The developer is a solo maker in Munich (also behind the long-running time-tracker Timing). His origin story: he kept pasting text into VS Code just to get Copilot's inline completions, then pasting it back into Mail or Slack. Autocomplete shouldn't live in one editor. It should work wherever you write, and it should stay on your Mac because what you type is some of the most personal data you generate.

The positioning is deliberately narrow. Most AI writing tools want to write the whole thing for you, then you spend ten minutes editing it back into your own voice. Cotypist does less on purpose: it finishes the next word or two and hands the keyboard back. The output still sounds like you because they are the words you would have typed anyway. No prompt to write, no chatbot detour.

How it works

  • Local LLM, no cloud. Cotypist runs a roughly 3 GB Gemma model on-device via llama.cpp. No API calls, no server in the loop. Works fully offline.
  • Inline, system-wide. Suggestions render as ghost text in the app you are already in. No per-app setup for most apps.
  • Tab to accept. Press Tab for the whole suggestion, or press it repeatedly to take one word at a time. Keep typing to reject. It re-snaps to what you actually meant within a letter or two.
  • Learns your voice. A Personalize Word Choice slider nudges completions toward your vocabulary, names, and phrasing. The more you write, the better it fits.
  • Model picker. It recommends a model that suits your Mac. Bigger is not automatically better: a heavy model that can't keep up with your typing makes the experience worse, not sharper.
  • Extras. Inline typo autocorrect (no red squiggles), emoji shortcodes (type : then filter), and screen-aware context so suggestions match the email thread or chat you're in.

Privacy

All processing happens locally. Words and data never leave the device. macOS automatically blocks Cotypist from reading password fields. On-device personalization data is stored encrypted via Keychain. There is no cloud logging of what you type. This is the strongest reason to pick it over any cloud-backed writing assistant for sensitive work (client emails, contracts, internal docs).

Where it works

Works out of the box in most standard Mac text fields: Safari, Chrome, Apple Mail, Word, Apple Notes, Notion, Obsidian, Messages, Slack, ChatGPT, and many more.

A few apps need a setup step:

  • Google Docs requires enabling Accessibility mode.
  • Arc / Dia need a browser setting toggled for inline suggestions.
  • Code editors (VS Code, Cursor): works in the AI sidebar chats, not the main editor, where Copilot is the better fit.
  • Terminals: activates when it detects you are typing into an AI agent's prompt.
  • Warp terminal is not supported (it doesn't expose the access Cotypist needs).

English works best. Other languages are supported with weaker results. Good for multilingual / code-switching writers per several user reviews.

Requirements

  • Apple Silicon only (M1 or later). Intel Macs are not supported.
  • macOS 14 or later.
  • Recommended: M1 Pro or M2 and up with 16 GB RAM.
  • Uses about 1 to 2.5 GB RAM while active. The local model is energy-intensive, so expect some battery cost on laptops.

Typical use cases

  • Email. Repetitive replies and follow-ups, where keystrokes stack up fastest.
  • Slack / chat / support tickets. Same phrasing over and over is exactly where prediction lands hardest.
  • Prompts for other AI tools. Faster prompts for ChatGPT, Claude, coding agents.
  • Repetitive technical writing and docs. Predictable by design.
  • Accessibility. Helps non-native English writers, dyslexic typists, one-handed typing, hand arthritis. The site claims up to 50% less typing.

Pricing

Every install starts with a 30-day Pro trial, no credit card. After that you drop to Free automatically and your settings are kept. Free is not a hard cutoff: past the daily limit, suggestions fade out gradually rather than stopping. Annual billing saves 25%.

Tier Price Daily limit Models Macs Notable extras
Free $0 100 completed words Curated set (light + mid) 1 Word-by-word accept, typo indicator
Plus $8/mo or $72/yr Unlimited Same curated set 1 Full autocorrect, custom global writing instructions
Pro $12/mo or $108/yr Unlimited Full catalog (largest models) 3 Per-app instructions, clipboard awareness, Labs early access

Only words you actually accept with Tab count toward the free limit. The three Pro Macs are for one person across machines, not for sharing between people.

How it compares

Tool Approach Scope Privacy Cost
Cotypist Local LLM, inline next-word prediction System-wide (most apps) On-device, offline Free / $8 / $12 mo
macOS native prediction OS text suggestions, small heuristic model System-wide On-device Free (built in)
TextExpander Static snippet expansion from abbreviations System-wide Cloud-synced snippets ~$5/mo and up
GitHub Copilot Cloud LLM ghost-text completion Code editors only Cloud ~$10/mo

The key difference from TextExpander: that tool expands fixed snippets you predefine (;addr becomes your address). Cotypist generates novel completions from context, no setup of canned text. They solve different problems and could coexist. Versus macOS native prediction, Cotypist completes whole phrases with real context, not single dictionary words. Versus Copilot, it is the same idea pushed out of the editor and into every app, running locally instead of in the cloud.

Honest take

Good for: anyone who writes a lot of email, Slack, or repetitive prose on an Apple Silicon Mac and wants Copilot-style flow everywhere, with full privacy. The on-device model is the standout, no client text in someone's training set.

The main limitations: Apple Silicon and macOS 14+ only (no Intel, no Windows/Linux yet), it does not work in the main code editor (Copilot still wins there), the local model costs RAM and battery, and the Free tier's 100 words/day is tight for heavy writers so daily use effectively means $8/mo. It is also a solo, non-venture project, so weigh bus-factor risk for anything you depend on.