A 3D fruit fly that lives on your macOS desktop — driven by a live spiking simulation of the real FlyWire connectome. It walks across your windows, grooms, sleeps, and decides to flee your cursor with the same neurons a real fly uses.

Fork of DenisSergeevitch/desktop-fly — this one gives the fly a sense of smell for vibecode.

It scans your disk for agent markers (AGENTS.md, CLAUDE.md, .cursor/rules, .kiro/steering and ~40 more) and turns anything on screen that leads to them into an odour source: an editor or terminal window with the project open, a row in the front Finder window, a folder icon on the desktop. An open project smells strongest, a closed icon weakest, and the reach of each grows with how much vibecode it holds — a hub of six marked repos is smelled across the whole screen, a single weak folder only from nearby. The steering neurons then walk the fly there, and when the smell is far the population wakes up enough to make it fly.

_{The fly's brain window: 23,210 real neuron soma positions from FlyWire v783, with live spikes flashing at real neuron locations. The two glowing yellow markers are the Giant Fibers — the escape command neurons. Click any region to stimulate it.}

• 23,210 neuron soma positions (of 139,255 in FlyWire v783) render the rotating brain window, colored by super-class (FlyWire's coarse cell-type grouping).
• A 668-neuron circuit with ~19,000 real synaptic connections (synapse counts, signed by neurotransmitter prediction) runs as a 1 kHz leaky-integrate-and-fire (LIF) simulation:
  • LC4 (104) + LPLC2 (210) looming-detector visual neurons
  • DNp01 / Giant Fiber (GF) (2) — the escape command neuron
  • DNa01 + DNa02 (4) steering neurons · DNp09 (2) forward walking
  • DNg11 (6) grooming · MDN (4) backward walking ("moonwalker")
  • DNp02/DNp04/DNp11 (6) escape-maneuver (wing) neurons
  • their 330 strongest partners, including ascending (proprioceptive) and sensory (wind) neurons
• Escape is not scripted. Your cursor's approach becomes looming input to the real LC4/LPLC2 cells; the fly takes off only when the Giant Fiber actually spikes through its real synapses — ~1,200 synapses of feedforward inhibition push back, which is why slow approaches are tolerated and fast lunges trigger escape in ~4 ms, just like the real animal.

The body itself is procedural (FlyWire is a brain connectome — no body geometry exists), with a tripod gait, visible wing-beat, altitude-scaled flight, grooming, and sleep postures.

Requirements: macOS 13+, Xcode Command Line Tools (Swift 5.9+). No permissions or entitlements needed — everything it senses (cursor, window frames, clicks-as-taps, thermal state) is permission-free.

git clone https://github.com/DenisSergeevitch/desktop-fly.git
cd desktop-fly
./build.sh
./DesktopFly

A 🪰 item appears in the menu bar; quit from there. The fly wanders your desktop on a transparent, click-through overlay — it never intercepts your mouse or keyboard.

The brain window is interactive: hovering pauses the rotation; clicking a region "optogenetically" stimulates the ~60 nearest circuit neurons for 400 ms. The fly's reaction is whatever the real network does downstream — click the Giant Fiber and it escapes; click DNg11 and it grooms; click one side's DNa01/02 and it turns.

The loop also closes body→brain: the gait rhythm feeds the circuit's real ascending (proprioceptive) neurons in phase with the legs, and fast cursor motion stimulates its sensory (wind) partners.

• Window terrain: window top edges are ledges — the fly lands on them, walks along them, rides a window you drag, and startles when one closes under its feet.
• Window looms: a window appearing near the fly feeds the looming pathway; the circuit decides whether to flee your dialogs.
• Clicks are substrate taps; clicking next to the fly startles it through the wind→GF pathway. Typing is vibration (idle-time API — knows when keys were pressed, never which).
• Circadian rhythm: dawn/dusk activity peaks, midday siesta, night quiescence. Sleep: idle at night → it sleeps, breathing slowly, with raised arousal threshold; it grooms after waking.
• Temperature: flies are ectotherms — a hot Mac is a faster fly.

data/ ships with compact derived files. To rebuild them from the raw FlyWire Codex dumps (~60 MB download):

mkdir -p /tmp/flywire && cd /tmp/flywire
B=https://storage.googleapis.com/flywire-data/codex/data/fafb/783
curl -O "$B/classification.csv.gz" -O "$B/coordinates.csv.gz" \
     -O "$B/connections.csv.gz" -O "$B/consolidated_cell_types.csv.gz"
cd - && python3 etl.py /tmp/flywire

./DesktopFly --simtest        # circuit invariants: GF silent at rest, 4 ms loom latency, ...
./DesktopFly --behaviortest   # 17 end-to-end checks: stimulate neurons -> body reacts
./DesktopFly --snapshot f.png  # offscreen fly render
./DesktopFly --brainshot b.png # offscreen brain render

Honesty section: the connectome gives wiring, not physiology. The LIF dynamics, neurotransmitter signs (ACh+, GABA−, Glu−), the gap-junction boost on LC→GF and wind→GF (documented electrical coupling), synaptic delays, and the sensory transduction (cursor → looming value) are standard modeling choices layered on the real graph. Everything downstream of the sensory neurons — who connects to whom, and how strongly — is FlyWire data.

Code is MIT. The files in data/ are derived from FlyWire (FAFB v783) and are CC BY-NC 4.0 — see data/DATA_LICENSE.md. If you use this, cite: