Magic Cursor is a system-tray AI overlay that captures what you're pointing at and lets a local model read it, explain it, transform it, or act on it — across any Windows app. No cloud. No telemetry. No login.
Magic Cursor adapts to the moment. A fast factual question gets a 220-pixel bubble. A complex task opens the full agent. You don't pick — it does.
Type a question. The cursor's context — selected text, screenshot, active window — comes along for the ride.
The model uses tools to do: insert at cursor, open URLs, read files, drive the UI. Every risky step asks first.
A compact 220×52 pill for short, factual answers. Auto-dismisses. Tap to expand if there's more to say.
One-tap quick actions match what you're pointing at — Explain, Fix, Summarize, Translate, depending on content type.
The model is wired to seventeen sandboxed tools. Safe ones run silently. UI control, file writes, and shell commands always require your approval — every call is logged.
Capture a region, read the active window, pick up cursor position.
Pull from your clipboard for context; push results back when done.
Read text files and list directories — strictly inside an allowlist you control.
Create, overwrite, or append. A confirmation dialog fires for every write.
Type, click, scroll, hit a hotkey — in whichever app is focused.
Hand off URLs to your default browser — no embedded webviews, no headless tabs running in the background.
Magic Cursor speaks to Ollama on localhost:11434. There is no cloud fallback, no analytics endpoint, no "anonymized" telemetry. Your screenshots, selections, and conversations stay on your disk — and you control what they touch.
Four phases between the wiggle of your wrist and a result you can use. Each one is observable, each one is interruptible.
A global mouse listener watches for ≥3 direction reversals inside a 600ms window. Wiggle the cursor like you'd swat a fly.
Clipboard, a 350-pixel screenshot around the cursor, and the active window's title — all snapped before the panel finishes its fade-in.
The local model picks a tool and calls it, gets the result, picks the next one. Up to six iterations, an emergency stop on every step.
Stream the answer back, insert it at the cursor, copy it, or run the side-effects you approved. Esc dismisses everything.
The last 50 sessions are saved locally with their context, query, response, and the tools that ran. Slide the panel open, scroll, replay anything — or wipe the lot.
Every session stores a thumbnail of what was on screen, the detected content type, the timestamp, and a transcript with the tool calls inline. Click any entry to load it back into the overlay — read-only, no re-run.
History lives on disk as plain JSON. Search it, grep it, delete it. FIFO at 50 entries by default; bump the cap in settings.
We're shipping signed Windows installers in waves of 200. No marketing, no fluff — one email when your slot opens, with the .exe attached.