Vigie was born from a very concrete problem: when Claude Code, Codex and Cursor work on the same project, each of them starts its own npm run dev. You end up with five copies of the app, ports that silently move to 3001 and 3002, and logs scattered across terminals nobody can find.
Vigie starts each server once, checks that it really answers, restarts it when it crashes, and exposes one live state (what runs, on which port, under which PID) through a JSON CLI, a loopback API and an agent skill. Agents inspect before they launch, and a desktop app for Windows shows the servers, their live output and every port of the machine. Open source, MIT license.
What it does
One process per server
Each dev server runs once. A second start on a taken port is refused and the owner is named, never moved to another port in silence.
Agents ask before they launch
An agent skill and a JSON CLI teach Claude Code, Codex and Cursor to read Vigie's state and reuse the running server.
Healthy, not just alive
TCP or HTTP health checks, automatic restarts with backoff after a crash, and the servers that were running come back at the next launch.
Who holds port 3000?
Every listening port of the machine, grouped by owner: managed by Vigie, started outside it, or protected system services.
A real desktop app
Windows window and tray menu, live logs, a terminal in one click, a built-in manual in English and French, and automatic updates.
Local and open source
The API only listens on 127.0.0.1: no account, no telemetry. The code is on GitHub under the MIT license.
Web SaaS