Cotypist: The German Indie Behind "AI Autocomplete Everywhere"

Cotypist sells a simple idea with a deceptively hard technical punch: it wants your Mac to finish your sentences wherever you type, without turning your writing into a chatbot session. The app markets itself as "AI autocomplete for Mac," promises to work across (almost) all applications, and frames the whole experience as flow-preserving rather than prompt-driven. It also leans hard into a second promise: your text stays on your device.

That pitch lands at a very specific moment. Big vendors push "AI writing" as a feature in a suite, a sidebar, or a web app. Cotypist goes after the mundane surface area instead: email replies, support tickets, comments, forms, and the half-finished sentence you would have written anyway if your fingers moved a bit faster. It aims to remove friction, not invent prose.

Under the hood, that ambition pulls Cotypist into a tricky niche. It must watch keystrokes across apps. It must overlay suggestions reliably. It must stay fast enough that users trust it. Cotypist's most strategic choice sits right there: it claims local prediction as a feature, not just an engineering preference.

What follows reads like a familiar startup story, except it starts in a suburb near Munich rather than in a Bay Area coworking space.

One developer, one suburb, one obsessive problem

Cotypist's corporate home sits in Neuried, a small municipality just outside Munich. German corporate records show Accelerated Thought GmbH founded on March 28, 2025, with €25,000 in share capital—standard, modest GmbH scale. CEO Daniel Gräfe previously built Timing, a Mac time-tracking tool that sold privacy as product differentiation long before "on-device AI" became a talking point.

Cotypist inherits that design philosophy directly. The technical bets—local processing, Mac-specific optimization, subscription revenue without cloud costs—all echo Timing's architecture.

A Reddit post describes an 18-month build wrestling with system-wide activation across nearly any text field. Eighteen months for autocomplete sounds excessive until you understand what system-wide means on macOS. Apple designed Accessibility APIs for screen readers and assistive technology. Using them for productivity means working through a deliberately narrow interface. You monitor text fields without triggering security warnings. You inject suggestions without breaking application behavior. You handle AppKit, UIKit on Catalyst apps, Electron wrappers, web views, terminal emulators—each framework implements text input differently.

Imagine trying to paint a portrait through a mail slot. That's the constraint. You're outside the room, shouting instructions, hoping the app on the other side interprets your intent correctly.

Bootstrap economics and subscription friction

Cotypist does not behave like a capital-saturated startup. Market-intelligence profiles state it has not raised funding. The company runs its AI locally, so it does not need to pay cloud inference costs for every user action. That design choice changes the economics entirely. Many AI writing companies must either charge aggressively or subsidize usage until they reach scale. Cotypist can fund development with software-style subscriptions rather than cloud-style consumption bills.

The Terms of Service make the long-term intent explicit: Cotypist plans to sell subscriptions and does not offer a perpetual license. A 30-day trial, then 50 completed words per day without a subscription.

This answers the question users always ask about indie Mac tools: will this become a one-time purchase, a subscription, or a quietly abandoned experiment? Cotypist has chosen the subscription path. Reddit threads already show users pushing for one-time pricing. Cotypist's terms effectively close that door.

The subscription economics work differently for local-first software. Cotypist's costs scale with development time, not with usage. Each new user adds no marginal cloud expense. But each new macOS version can break compatibility across the entire user base simultaneously.

What works, what strains, what breaks

Cotypist's core service is prediction. The app watches as you type and offers inline completions in real time. You can accept a whole suggestion or accept it word by word.

Here's the hard part: Cotypist relies on macOS Accessibility APIs to monitor text input across applications. The app detects your keystrokes and sends context to a language model. The model generates predictions. Those predictions get displayed as suggestions. The entire cycle runs while you're typing the next character—maybe 100-200 milliseconds if you're moving quickly. Miss that window and the suggestion arrives late, which breaks flow worse than no suggestion at all.

Gräfe chose Qwen 2.5 1.5B as the primary model because it generates completions fast enough on Apple Silicon to stay usable. Larger models might produce better predictions but miss the latency budget.

Cotypist requires Apple Silicon and macOS 14 or later. The site recommends at least an M1 Pro or M2 with 16 GB of memory, and claims the app uses roughly 1–2 GB of system memory while active. This is not a lightweight utility. Cotypist asks users to trade RAM for speed and privacy.

Privacy stands at the center of the sales pitch, but the fine print reveals a distinction. The privacy policy states that Cotypist processes text predictions locally and does not send typed content to servers or third parties. It transmits license information to validate licenses and check updates. It collects anonymous usage statistics while explicitly excluding written content.

This is a realistic privacy posture. It is not the "air-gapped" fantasy some users imagine. Cotypist's differentiator lies in where it draws the boundary. Your words stay on your Mac.

We tested the app. Despite beta status, it works remarkably well. The inline suggestions feel responsive in email clients and browsers. The predictions match tone and context better than expected from a 1.5B parameter model. The memory footprint stays noticeable on machines with 8 GB of RAM, and occasional misfires appear in complex web applications, but the core experience delivers on the promise: faster typing without leaving your text field.

The competition question

Cotypist competes in a market that rarely defines itself cleanly. It sits between autocorrect, writing assistants, and code completion.

Apple ships predictive text and expanding writing-related features. Cotypist's advantage comes from focus and speed. But Apple's roadmap creates existential risk. This isn't just market competition. It's the specific dread of every Mac developer: waking up during WWDC to see your entire business model announced as a bullet point in macOS 16.

Cloud writing assistants like Grammarly built audiences by living in browsers and office suites. Grammarly has over 30 million daily active users. The Premium tier costs $12-15 monthly. Cotypist emphasizes speed and continuity instead. It does not want to rewrite your paragraph. It wants to finish your sentence.

System-wide AI companions for Mac try to bring LLM features into every text field via hotkeys or floating panels. Many send text to cloud APIs. These tools assume you will pause, select text, trigger the assistant, review output, then paste results back. Cotypist assumes you never want to leave the text field at all.

GitHub Copilot owns code completion at $10 per month for individuals. Cotypist explicitly avoids competing there. It defaults off in developer contexts and signals that it wants to live where Copilot does not: the messy sprawl of non-code writing.

This dynamic creates an unusual risk: Cotypist can succeed and still lose. If Apple builds a strong system-wide completion engine that runs on-device, many users will accept the default. Cotypist would then need to defend itself as "better than the built-in one," the hardest position in consumer software.

The sustainability bet

Cotypist arrives at a specific moment in the local AI model trajectory. Two years ago, running decent language model inference on consumer hardware meant either cloud APIs or unusably slow results. Apple Silicon changed that calculus.

That capability window creates opportunity. It also signals when the window might close.

Gräfe's background with Timing provides hints about likely strategy. Timing charges $80 for perpetual license updates or $30 per year for subscription. That puts it at the higher end of utility pricing but below professional creative tools. Cotypist needs to establish similar positioning to justify premium utility pricing.

The next eighteen months will determine whether "AI autocomplete everywhere" becomes a category with multiple successful players or whether it collapses into OS features and enterprise suites.

That bet looks reasonable. But reasonable bets lose all the time in consumer software. Users who pay for productivity tools expect perfection. Competing against free OS features requires flawless execution.

If Cotypist pulls that off, it will achieve the rarest outcome in productivity software: users will forget it exists—right up until they sit down at a machine without it and suddenly feel like they're typing with mittens on.

Tech Force: Government’s $200K Plan to Rehire What It Fired

The administration cut 317,000 federal workers. Now it wants 1,000 tech recruits from Palantir, Amazon, and Microsoft—who keep their stock while shaping government AI. The math is interesting. So are the conflicts of interest.

Implicator.aiRobert Brown

ChatGPT 5.2: What You Need to Know About OpenAI’s Update

Sam Altman’s “Code Red” memo triggered OpenAI’s fastest major release ever. Ten days later, GPT-5.2 arrived with doubled benchmarks and 40% higher API costs. The gains are real. So are questions about what got sacrificed for speed.

Implicator.aiMarcus Schuler

The Class Divide in Teen AI: Who Gets Which Chatbot

64% of teens use AI chatbots. But which ones? Higher-income teens cluster around ChatGPT for productivity. Lower-income teens are twice as likely to use Character.ai—the companion bot facing wrongful death lawsuits. The technology is sorting kids by class.

Implicator.aiMarcus Schuler