Cotypist's Local AI Autocomplete: The Anti-Chatbot Writing Tool

Cotypist, a Mac app that predicts your next words in any text field using a local on-device model, launched on Product Hunt and is generating the kind of word-of-mouth buzz that founders dream about. John Gruber at Daring Fireball called it “eerily good.” Quinn Nelson at Snazzy Labs said “it’s bloody fantastic.” Reddit users report it saving their wrists from carpal tunnel.

The app is not a chatbot. It does not open a separate window. It does not require a prompt. It sits in the background, runs entirely on Apple Silicon, and suggests the next word or phrase as you type. Press Tab to accept. Keep typing to ignore. That is the entire interaction model.

Cotypist’s rise is a signal about where a segment of the AI writing market is heading. After two years of everyone building chatbot wrappers, a small but vocal group of users is voting for something much narrower: an autocomplete that augments typing speed without replacing the writer’s voice. The app claims to cut typing by up to 50 percent. It works offline. It never sends text to a cloud server.

The product is the creation of an independent developer, not a well-funded startup. The website is a single-page site with a download link, a pricing table, and testimonials. The pricing is straightforward: a free tier that caps at 100 accepted words per day, a Plus plan at $6 per month for unlimited completions on one Mac, and a Pro plan at $9 per month for up to three Macs plus the largest local models. Every new install gets a 30-day Pro trial.

What makes Cotypist interesting is not the technology alone. On-device language models are not new. Apple has been shipping on-device ML for years. Google’s Gboard has had on-device next-word prediction for even longer. What is new is the product sensibility. Cotypist is designed for people who write for a living and who have grown tired of the chatbot workflow: stop writing, open a separate app, craft a prompt, wait for a wall of text, edit it back into your voice, then paste it in. The Cotypist website calls this “dancing with the AI” versus “delegating to it.”

The app runs a curated set of open-weight models. The free and Plus tiers use Gemma 4 E2B, Qwen 3 1.7B, and Gemma 3 1B for lightweight suggestions, plus three mid-size models for sharper completions on capable hardware. The Pro tier adds Gemma 4 26B A4B, Qwen 3 30B-A3B, and Qwen 3 8B. All run locally. The developer notes that larger is not always better: if a model cannot keep up with typing speed, the experience degrades. Cotypist recommends a model based on the Mac’s hardware.

Memory usage runs 1 to 2.5 GB. The app requires macOS 14 or later and an Apple Silicon Mac. Intel Macs are not supported. The developer recommends at least 16 GB of system memory for the best experience.

The business model is notable for what it is not. Cotypist does not charge per token. It does not have a usage-based API. It does not monetize user data. The privacy pitch is central: all processing happens on device, no data leaves the Mac, and macOS automatically protects password fields. The developer explicitly states that Cotypist does not claim copyright over anything written with its assistance.

The app’s traction on Product Hunt and the quality of its early endorsements suggest genuine product-market fit for a specific use case. Gruber’s review, Nelson’s video, and the Reddit threads all describe the same feeling: the suggestions are accurate enough that they feel uncanny, and the friction of the interaction is near zero. That is a hard combination to achieve. Most AI writing tools optimize for capability — longer outputs, more creative suggestions, better reasoning. Cotypist optimizes for latency and restraint. It tries to guess the next word you were about to type, not the next paragraph you should write.

This is a bet that a meaningful number of knowledge workers want AI that makes them faster at their existing process, not AI that replaces the process entirely. The chatbot paradigm assumes the user wants to outsource the writing. Cotypist assumes the user wants to keep writing, just with fewer keystrokes.

The limits are real. Cotypist is Mac-only. It does not work in code editors for main editing — the developer explicitly says tools like GitHub Copilot are better suited there. The free tier’s 100-word daily cap is tight for anyone who writes more than a few emails. The app’s value is cumulative; saving 50 percent of keystrokes matters most to people who type tens of thousands of words per week.

But the signal is worth watching. Cotypist’s early success suggests that the pendulum may be swinging back from “AI writes for you” toward “AI helps you write.” The developer’s decision to run everything locally, charge a flat subscription, and design for the existing typing flow rather than a new chat interface is a deliberate counterpoint to the industry’s dominant paradigm. If the app sustains its momentum, it will have demonstrated that there is a real market for AI that gets out of the way.