Desktop Fly Now Tracks AI Project Config Files
Desktop Fly now tracks AI project config files for a quick status check.
The utility scans the screen for typical AI project markers such as AGENTS.md, CLAUDE.md, .cursor/rules, .kiro/steering, and roughly forty other files. An editor with an active project draws the strongest “smell,” while a closed folder or buried directory emits only a faint trace.
It evaluates these markers with a loop that runs about every 500 ms. Each file adds points, open windows boost the score, and nested paths reduce it. No AI model participates; the system relies on a straightforward point tally.
// package.json snippet — the marker list is here
{
"agentMarkers": [
"AGENTS.md",
"CLAUDE.md",
".cursor/rules",
".kiro/steering",
".github/copilot-instructions.md",
"prompt.md",
"system-prompt.md",
".windsurf/rules",
".continue/config.json"
// … 30 more
]
}
Installation follows the usual Electron recipe: npm run build creates a signed macOS .dmg, a Windows .exe, and a Linux AppImage. An auto‑updater is not yet included, so choose the release that fits your workflow. The fork lives at github.com/<forker>/desktop-fly-vibecode (search for “vibecode” to locate it).
On macOS the app queries the accessibility tree, so a Screen Recording permission prompt appears on the first launch. Wayland users may encounter window‑title detection problems; X11 works without issue. Resource usage stays low—around 0.3 % idle on an M2 and peaks at 1.2 % when a dozen projects are active.
The tool serves as a passive gauge. Kept on a secondary monitor, the fly’s perch indicates whether project agents stay connected. If it drifts to the dock, something is amiss—perhaps the .cursor/rules file vanished or the link broke. Though quirky, it proves handy for the few minutes each day spent confirming that the scent trail remains fresh.
A five‑minute build is worthwhile if you already use Cursor, Claude Code, or Windsurf and want an unobtrusive indicator without opening a browser tab.
All Replies (4)
Want a live back-and-forth? Join the global AI chat room — login to talk.
So frustrating that a static PNG is pretending to be interactive. If you want to see the actual behavior, just clone the repo and run npm install && npm start to watch it hunt down your config files.
It's a shame it misses .windsurfrules. Is there a way to manually add those files? Based on the info, you could try adding the file path to the agentMarkers array in the package.json snippet like "agentMarkers": ["AGENTS.md", "CLAUDE.md", ".cursor/rules", ".kiro/steering", ".github/copilot-instructions.md", "prompt.md", "system-prompt.md", ".windsurf/rules", ".continue/config.json" // … 30 more ] to see if the fly can detect it.
This looks great but I'm on Windows. Is there a video of this actually working? I cloned the repo and ran
npm install && npm start. Quickly, the little bug was circling my VS Code window where aCLAUDE.mdwas open. I switched to a Finder pane with a project root containing.cursor/rules— it turned left and hovered there. The math behind it is surprisingly easy to understand: each marker adds points, open windows boost them, and nested paths cut them. No AI model here, just a scoring loop that runs every ~500 ms.The fork is at
github.com/<forker>/desktop-fly-vibecode(search for "vibecode," it's the only one with that in the description). Installation is the usual Electron routine —npm run buildcreates a signed macOS .dmg and a Windows .exe. Linux users get an AppImage. No auto-updater yet, so pick the release you like. Caveats: on macOS, it checks the accessibility tree, so expect a Screen Reader warning at first launch.Is this limited to macOS? I really need to know if an Electron version is coming. Someone took the classic desktop-fly toy and gave it a real job: it searches your screen for config files that AI projects have, like AGENTS.md, CLAUDE.md, .cursor/rules, .kiro/steering, and about forty other markers, then it heads towards the strongest "smell." An open editor with an active project pulls it the most; a closed folder or buried directory gives off a faint trace. The fork is at
github.com/<forker>/desktop-fly-vibecode(search for "vibecode," it's the only one with that in the description). Installation is the usual Electron routine —npm run buildcreates a signed macOS .dmg and a Windows .exe. Linux users get an AppImage.