DesktopFly simulates 668 neurons behind one desktop pet
DesktopFly puts a procedural fruit fly on a transparent macOS overlay and lets a neural simulation influence what it does. The modeled circuit contains 668 neurons and roughly 19,000 connections derived from FlyWire data. Cursor approaches feed looming-detector cells, while named descending neurons affect escape, steering, walking, grooming, and backward movement. A separate brain window plots 23,210 neuron positions and flashes active circuit locations. The result is part toy, part data visualization, and part argument for making scientific datasets tangible.
The strongest design choice is also the easiest to miss: the README labels what came from measurement and what the author supplied. FlyWire contributes the connection graph and strengths downstream of the sensory cells. Leaky integrate-and-fire dynamics, neurotransmitter signs, synaptic delays, electrical-coupling boosts, and conversion of cursor motion into a looming value are model choices. The 3D body is procedural because the connectome has no body geometry. DesktopFly therefore demonstrates behavior generated through a biologically grounded graph; it does not reproduce a whole animal.
Our 4-second build covered Python data tooling, not the Swift app
Our sandbox cloned commit 7014d37, identified the Python data-generation path, and installed 35 packages in 14 seconds. Those dependencies occupied 37 MB. The configured build step succeeded in 4 seconds, while pip-audit reported 0 known vulnerabilities. No test script or target was available, so the lab skipped tests. That is a clean result for the detected Python side, but it does not establish that the Cocoa application compiles or renders correctly.
The actual app has a different boundary. Its shell script calls Swift with Cocoa and SceneKit, and the README requires macOS 13 or newer plus Xcode Command Line Tools with Swift 5.9 or newer. After compilation, the user launches DesktopFly directly and quits through a menu-bar item. There is no latest GitHub release, signed application bundle, or installer described. For a personal experiment this is manageable. For distribution inside a company, packaging and provenance work remains with the adopter.
The 1.4 MB checkout uses macOS surfaces without extra permissions
The repository contains only 16 files, about 2,564 source lines, and occupied 1.4 MB in our checkout. Within that small codebase, the fly can walk along window edges, ride a dragged window, react when its ledge disappears, and move between displays. Cursor position, window frames, clicks, idle time, and thermal state provide environmental signals. The README says those inputs are available without macOS permissions and that the overlay is click-through, so the pet does not intercept ordinary mouse or keyboard input.
That permission-free approach has clear limits. Typing is inferred from idle-time changes rather than captured characters, which is privacy-friendly but coarse. Window enumeration and overlays are operating-system features, so portability is not a recompile. Issue #8 describes a community X11 and GNOME port whose author says native Wayland windows are invisible to the pet because normal clients cannot enumerate them or watch their input. Issue #7 points to another independent Wayland effort. Neither issue turns those ports into official support.
The interactive brain is informative if you keep the model boundary visible
The brain view makes the project more useful than a decorative animation. Hovering stops rotation, and clicking a region stimulates nearby circuit neurons for 400 ms. Selecting the Giant Fiber can provoke escape, while other named groups affect grooming or turning. The body also feeds gait rhythm back into ascending neurons, and fast cursor movement stimulates wind-related partners. These couplings give a viewer something to inspect rather than a canned list of animation states.
DesktopFly also ships diagnostic commands for circuit invariants, behavior reactions, and offscreen snapshots. Our lab did not run them because no automated test target was exposed, so their presence should not be confused with a measured pass. A useful next step for the project would be wiring those diagnostics into CI on a macOS runner. The repository currently showed 0 CI workflow files and no tests directory, leaving changes to simulation or rendering without visible automated enforcement on GitHub.
No release and mixed licenses keep this in experiment territory
GitHub showed 726 stars, 10 open issues and pull requests, and a last source push on August 18, 2026. Issue discussion continued on August 21 through community Linux-port reports, so the project drew attention after the last commit. GitHub had no latest release to return. That activity fits a new, compact experiment better than an established desktop utility with packaged updates and a long maintenance record.
Licensing needs equal care. Source code is MIT, while the derived connectome files in data/ use CC BY-NC 4.0 and require the cited FlyWire papers. Regenerating those files needs Python and roughly 60 MB of raw downloads. For personal use, education, or an art-and-science demo, the boundary is clear enough. A commercial fork should review the data terms before assuming the code license covers the shipped experience. DesktopFly is memorable because its constraints are visible, not because it erases them.