Caret

Caret's pitch is narrow and coherent, which is rarer than it sounds in this category. The product is a Tab key. The homepage states the design position bluntly — "Caret is not another chat app" — and the whole feature set follows from it: inline grey completions, screen-wide context, persistent memories of how you write, and completions delivered in your own voice. Where this genuinely helps is text you already half-know how to write. Finishing the last clause of an email in Apple Mail, populating a CRM note from the customer screen you just closed, replying to a Zendesk ticket using phrasing carried over from earlier replies, or completing a line in Xcode. Those are all seeds the vendor itself describes, and they are the right shape for the interaction: you supply intent, Caret supplies the rest. Because the tool reads your screen rather than a pasted prompt, the context transfer is the real differentiator — you don't have to explain what you're looking at. The trade-offs are structural. First, the interaction model is one-directional: Caret extends text you are already writing. It is not something you can ask a question of, so research, brainstorming and reasoning tasks still send you to a chat assistant. Second, the platform floor is steep — macOS 26 or later — and there is no Windows, Linux, web or mobile version documented on the scraped site. Third, it is a personal desktop utility rather than team infrastructure: the scraped site documents no admin controls, no centralised billing and no API. If you need to roll this out to a support floor of thirty agents with policy controls, that is not a story Caret tells today. The memory angle is the part worth judging over time. "Gets smarter the longer you work" is the vendor's claim, and it cuts both ways — a tool that learns your voice is more useful in month three than week one, but it also means the first week is the least impressive one. Budget for that ramp rather than judging it in an afternoon. Who should look elsewhere: cross-platform teams, anyone who wants a conversational assistant they can interrogate, and organisations that need SSO, seat management or audit trails before they'll install a screen-reading agent. For solo Mac users who type a lot and hate context-switching, the Tab-only model is the point, not the compromise.