Most AI apps
- Leave your work, open another window
- Every switch breaks your focus
- Paste your context in every time
- Every prompt waits on a cloud round-trip
- Sign up, verify email, add a card
- Pay every month, forever
Local · Private · Open Source
Thuki, a floating AI assistant for macOS. Hiding in plain sight.
Your AI assistant, summoned by double-tapping ⌃
Floats above any app. Your model, your machine.
The problem with AI tools todayEvery other AI tool makes you stop, switch,
Every other AI tool makes you stop, switch,
and break your flow.
Thuki
- Summon it over any app. No tab-switching.
- Floats above everything you do
- Your selection is already the context
- Answers from your own local model
- No account, no card, no email
- Set up once: your own local model
How it worksOne key.
One key.
Two taps.
Anywhere.
- 01
Press Control twice.
From your browser. Your terminal. A fullscreen Figma board. Thuki listens at the HID level, so it always answers.
- 02
Ask, paste, drag, or capture.
Type a question, drop in an image, or use
/screento attach what you're looking at right now. - 03
Get an answer. Dismiss.
Your local model answers. No cloud round-trip. No cost. Press esc and it disappears as if it was never there.
Built for the way you actually workA pocketable assistant
A pocketable assistant
that never gets in the way.
Always one
keystroke away.
Double-tap Control from any app, including fullscreen. Thuki floats on top. No window-switching. No tab-juggling.
Highlight,
then ask.
Select text in any app, double-tap, and the selection is pre-filled as a quote. Ask what it means. Thuki already has the context.
Capture
your screen.
Type /screen and Thuki snaps your full desktop, then attaches it as visual context. Vision-capable models see exactly what you see.
Agentic
web search.
/search runs a multi-step pipeline on your machine: SearXNG, Trafilatura, judgement steps, fetching the live web with a trace of every fetch and decision.
Extended
reasoning.
Type /think and the model reasons step-by-step before answering. You see the thought process; you keep the conclusion.
Just a slash.
Zero menus.
Type a slash and Thuki completes it. Every task has a verb. No dropdowns. No settings.
thuki · ask
›/translate"Bạn có khoẻ không?"
›/tldr[selected: 3,000-word article]
›/rewrite"per my last email…"
›/screen→ capture full desktop
›/extract[receipt.png]
›/search"latest llama.cpp flags"
›/think"why is this O(n²)?"
›/explain"what is a closure?"
›/bullets, /todos, /refine
›/
Lock the model
in a sandbox.
Optional: run your local model in a hardened Docker container. Capabilities dropped, read-only weights, localhost-only. You opt in; it is not on by default.
SECURED
$ bun run sandbox:start # model runtime, sandboxed: # cap_drop: ALL # read-only weights # 127.0.0.1 only
✓
Capabilities droppedNo privilege escalation or raw socket access
✓
Read-only filesystemAll volumes mounted with :ro
✓
Localhost-only networkNo external egress. Your data stays local.
Privacy isn't a feature. It's the architecture.No Thuki server.
No Thuki server.
Your model runs on your machine.
0
Servers we run
0
API keys to manage
0
Telemetry in the app
$0
Per-token cost, ever
Your data: one SQLite file on disk.
No backend, no server. Inference runs on your own machine. Delete the file, the data is gone.
Your conversations live in one place: a SQLite file on your disk.
That's not a marketing claim. It's the entire data architecture. There is no backend. There is no server. There is no log file you can't cat. Inference runs on your own machine, not a server we operate. If you delete the file, the data is gone.
Frequently asked questions
Straight answers.
Stop going to the AI.
Make it come to you.
No account. No card. Open source. Yours to keep.
Get Started for Free→Apache 2.0 · Open source · macOS 13+