A floating chat copilot that just reads your screen — politely

What it does

When a message lands in a supported macOS chat app, a frosted-glass floating panel pops up with the message’s intent (8 categories), a 0–9 risk score with action advice, and candidate replies grouped by persona — including the “dog-headed strategist” trio of steady, flirty, and detached. You pick a candidate, hit fill-in, and it lands in the chat input box via accessibility APIs. Sending stays manual; the tool never presses send for you.

The interesting bit

The whole thing is read-only by architecture: no injection, no hooks, no decrypting chat databases. It literally screenshots the window, runs Apple Vision OCR, and classifies bubbles by surface, borders, and left/right alignment to figure out who said what — including handling quoted messages as background context rather than new messages. There’s also a second path that reads text directly via the system accessibility interface, no screenshots at all.

Key highlights

• Intent classification is zero-shot at 86.4% on a 22-message regression set, with the full distribution from a single local forward pass
• 14 reply personas generated concurrently, then re-ranked in place by a local model; per-persona candidate count is configurable (default 2)
• Fast on M1 Pro: ~1.5 s to intent, ~1.5–2 s to candidates (README notes these are mechanism-derived estimates; logs are the source of truth)
• Optional manual calibration mode for draggable contact lists, with a preview gate before any region is saved
• Runs fully local; a privacy doc covers data flow

Caveats

• Apple Silicon only (M-series, macOS 13+) — the torch dependency has no Intel build, and Rosetta is explicitly unsupported
• The app isn’t notarized, so first launch requires a right-click open or clearing the quarantine attribute
• Manual calibration doesn’t auto-detect internal layout changes (dragged lists, input-box height, chats with pinned announcements) — you must recalibrate yourself
• OCR-based recognition is validated on specific layouts and light/dark themes; the README is candid that this isn’t a general vision model and complex themes may still slip through

Verdict

Worth a look if you’re on an M-series Mac and want a local, privacy-respecting reply assistant for chat apps — the read-only design is genuinely thoughtful. Skip it if you’re on Intel, on another OS, or expecting a general-purpose screen-understanding model.