A Grammarly-like Mac app built on Jev

It does not, and the author calls that the catch. Jev returns a probabilistic decision rather than generated text, so proposing a fix still needs an LLM. His workaround is routing only the flagged text to the slower model instead of everything typed, which is a real saving but not a Jev-only pipeline.

That is the author's own testing figure for his Mac app, not a published benchmark. He says responses were typically around 300 ms and feedback felt almost instant, with no percentile data, no measured baseline LLM comparison, and no cost breakdown.

Yes, and he flags it himself as the bigger problem. Jev is reached through TypeSafe's hosted API today, so text sent for checking leaves the Mac, which he considers a privacy issue for a writing assistant. He notes that a local or on-device version would be far more interesting and is not available.

A video demo. The post links no repository and gives no install path, so there is nothing to audit for keystroke handling or reuse as a starting point. If you build your own, the shape is simple: watch input, ask a Boolean or Score per fragment, and escalate only what trips.

It watches what you type across the whole system rather than inside one editor, which is what a Grammarly-like tool has to do. That implies system-level input monitoring, and the post describes the app watching typing across macOS without detailing the permission model or what it does with text before deciding whether to ask Jev.

Niels Mouthaan, a Dutch indie developer based in rural Sweden who lists several apps in his bio. The Mac writing assistant is a one-off experiment to feel how fast Jev responds in a real app, and he links no product page or repository for it, so it reads as personal tinkering rather than a shipping tool.

Speed and cost at high volume. He says an LLM can make these decisions too, but Jev is designed for them, which makes it much faster and cheaper for apps that make lots of small semantic decisions in real time. The LLM still comes in, but only for the text Jev flags rather than everything you type.