The AI that's
already there.
Double-tap Shift anywhere. Caret reads your screen, remembers you, and answers instantly — and nothing ever leaves your device.
Want to help build it? Become an early beta tester →
Rewrite anything,
in place.
thanks for the email, ill get back to u soon
One keystroke
Quick. Chat. Actions.
Summon Caret the same way every time — it works out whether you want a quick edit, an answer, or something done.
Caret Quick
Transform text, in place.
Rephrase, fix, translate, reply, refactor — Caret edits what you selected and pastes it right back, in about a second.
Caret Chat
Ask about your screen.
It reads exactly what's in front of you — code, errors, tables, images — and answers, with a memory of you.
Caret ActionsSoon
Tell it what to do.
Caret acts across your apps — open, click, type, send — with a confirm step before anything that matters.
Ask anything on your screen.
Point at anything — a cryptic error, a dense table, a chart you don't follow — with Ctrl+Shift+C, then ask. Caret reads the exact text on your screen, not a blurry guess, and tells you what it means. Nothing gets uploaded.
TypeError: Cannot read properties
of undefined (reading 'map')
at Cart.render (cart.tsx:42:18)It remembers you.
Tell it once — “I prefer short answers” — and it sticks. A private memory that compounds over time, fully visible in /memories and yours to edit or wipe. Nothing hidden, nothing in the cloud.
Caret Memory/memoriesOr just say it.
Prefer to talk? Speak your instruction and Caret transcribes it on-device — then rewrites, answers, or drafts, right where you're working. Your voice never leaves your machine.
Caret QuickSoon, it won't just answer.
It'll act.
Tell Caret what to do, in plain words. It reads context from one app and acts in another — with a confirm step before anything that matters.
“log this invoice total in my expenses sheet”
The hard part
How a small, local model behaves like a big one.
The trick isn't calling an LLM — it's making a tiny, on-device one reliable. That's the part competitors don't have.
Reconciled by geometry
Most “AI sees your screen” tools send a screenshot and hope. Caret fuses the accessibility tree, a dedicated OCR engine, and vision — then merges them by position, so the most trusted source wins. Prices, code, and tables come back verbatim, not a blurry guess.
Code decides, the model suggests
A small on-device model is weak if you trust it blindly. Caret wraps it in deterministic guardrails that gate what gets remembered, which source to trust, and when to act — so the system stays reliable even when the model wavers.
The right tool for each job
Exact text comes from accurate engines, never the model's eyes. The LLM only does what it's genuinely good at — phrasing, intent, judgement — on top of context that's already rich and correct.
Built for speed
Some answers don't need a model.
Just start typing. Math, conversions, dates, colors, encodings — Caret answers the instant you type, in about 50 milliseconds, without ever waking the GPU.
Private by design
Runs entirely on
your machine.
Privacy isn't a setting in Caret — it's the architecture. Everything happens locally, so your work stays yours by default. That's exactly what makes it safe for an AI to see your screen and remember what matters.
FAQ
Questions, answered
A local-first AI assistant for Windows. Double-tap Shift from anywhere and Caret reads what you've selected, takes a typed or spoken instruction, and answers — often typing the result straight back into the app you were in.
Windows 10 or 11, the free Ollama runtime, and a recent GPU (around 8 GB of VRAM is a comfortable baseline for capable models, though Caret runs on less with smaller ones).
Yes. Caret runs free, open models on your own machine — no subscription and no per-query cost. You only ever pay for the hardware you already own.
Yes. Once you've downloaded a model, Caret works with no internet connection at all. The only things that ever need the network are optional live lookups like currency rates.
Completely. Inference runs locally through Ollama, nothing is sent to the cloud, there's no telemetry, and your memory lives in a local store you can view, edit, or delete at any time.
Any model available through Ollama — Gemma, Qwen, Llama, Mistral, and more. You can switch between them in seconds, and Caret adapts to each model's capabilities.
Caret is in active development. Join the waitlist and you'll be among the first to get access when early builds go out.
Windows comes first — that's where the deep OS integration lives today. Other platforms are on the roadmap.
Be first to try Caret.
It's almost here. Leave your email and we'll send your invite the moment early access opens — nothing else in between.
Windows · Free · Works offline