AI Garage

Mission control for your localhost — services, ports, tunnels, and the processes your AI agents left running. No terminal: see everything that's up, kill a stuck port or a zombie process in one click, expose a service to your phone with a public link, keep services alive — on macOS (and VPS).

▶︎ Run it — one command

npx ai-garage

Then open http://localhost:7777 in your browser. That's it — no install, no config, no dependencies.

Why

You run a dozen local servers, dev environments and side processes. Then your AI coding agents (Claude Code, Cursor, …) spawn even more — and forget to clean them up. Soon a port is "already in use" and you don't know by what.

AI Garage is a single pane of glass for all of it:

• See what's actually running — saved services and every listening port auto-discovered, including the ones AI agents spawned.
• Free a stuck port in one click — graceful SIGTERM, then SIGKILL, with a guard on system/database ports.
• Start / stop / restart / keep-alive any service — without touching the terminal.
• Share to your phone — a public link (cloudflared) in one click.

Features

• 🟢🔴 Live status of every service (port check), auto-refreshing every 3s.
• ▶ ■ ⟳ Start / Stop / Restart with one click; Keep-alive that the panel itself supervises (restarts a crashed service).
• 🧹 Free the port — kill whatever holds it (the classic "port already in use"), with data-loss protection on DB/system ports.
• 🔭 Auto-discovery — see every listening process you didn't add, classified (system / database / app / dev / AI-agent) so you don't kill the wrong thing.
• 📱 Public link — expose a local port via cloudflared (ngrok & custom domains coming).
• 👁 Live preview of a service right inside its card.
• 🤖 Bots & agents view — group Telegram bots/agents by the agent backend they run on, with one-click "open in Telegram".
• 🗂 Sections for This Mac / VPS / Bots, drag-to-reorder, rename, change device.
• 🌍 19 languages, switchable in the header.
• 🎨 Clean glass UI with a live cursor highlight.

🔌 Connections — one map of every API access on your machine

Your AI stack sprawls across tools: API keys in .env, OAuth accounts inside hubs like Composio, and MCP servers wired into each AI client separately. Any single agent only sees its own config. Nobody has the full picture — and things silently expire.

The Connections tab fixes that. It reads the live configs of all your AI clients at once and shows:

• Access map — who can reach what. Which AI clients (Claude Code, Cursor, Claude Desktop, Antigravity, VS Code, Zed, Windsurf, LM Studio) have access to which MCP service — deduped across clients, auto-discovered from their real config files. Flip it: pick a client, see everything it can reach.
• One-click grant / revoke. Give or take away a client's access to a service right from a card — every config is backed up before any write, edits are atomic, Claude Code goes through its own claude mcp add/remove.
• Health at a glance. Every API key checked live — works / expired / out of credits in plain words. Keys never leave your Mac.
• Composio, honest. Per-service connection status (connected / needs reconnect), one-click OAuth connect, and a catalog of popular apps to add — plus it flags a half-finished CLI login that makes agents throw 401.
• Honest about blind spots. It tells you what it can't see (a client's UI toggle, a cloud-only hub) instead of faking a green check.

Agents get it too: the MCP server exposes connections_overview, connections_health, grant_access, composio_connect — so you can say "give Cursor the same access Claude Code has" and it just happens. Security-reviewed adversarially; no secret values ever leave your machine.

Use with AI agents (MCP)

AI Garage ships an MCP server so coding agents can drive it directly — list what's running, free a port, open a tunnel — instead of guessing with lsof and kill -9.

Add it to your agent (Claude Code, Cursor, …):

{
  "mcpServers": {
    "ai-garage": { "command": "npx", "args": ["-y", "-p", "ai-garage", "ai-garage-mcp"] }
  }
}

Tools exposed: list_services, free_port, open_tunnel, close_tunnel, register_service, and the Connections tools connections_overview, connections_health, grant_access, composio_connect. The panel must be running (npx ai-garage); the MCP server talks to it locally.

How it works

• server.mjs — a zero-dependency Node server (built-in modules only). Listens on 127.0.0.1:7777 only.
• public/index.html — the whole UI (vanilla JS, i18n).
• mcp.mjs — the zero-dependency MCP server.
• Service registry: ~/.config/localhost-control/services.json.
• Optional config: ~/.config/localhost-control/config.json.

Requirements

Node ≥ 18, macOS (Linux/Windows partially work; keep-alive & some features are macOS-first). cloudflared is optional, only for public links.

services.json

{
  "name": "My App",
  "type": "local",
  "port": 3000,
  "url": "http://localhost:3000",
  "cwd": "~/projects/app",
  "startCmd": "npm run dev",
  "stopCmd": "lsof -ti:3000 | xargs kill",
  "host": "Mac",
  "note": "what it is"
}

• type: "local" — has Start/Stop/Restart buttons (needs startCmd/stopCmd/cwd).
• type: "link" — status + link only (managed elsewhere: a VPS, a launchd tunnel, …).

API

| Method | Path | Body / purpose | |---|---|---| | GET | /api/status | all services + discovered ports + tunnel URLs | | POST | /api/start | /stop | /restart | { name } | | POST | /api/kill-port | { port } — free a port | | POST | /api/tunnel-start | /tunnel-stop | { port } — public link | | POST | /api/keepalive | { name, enable } | | POST | /api/service-add | /service-remove | /service-rename | manage the registry |

Security

• The server binds to loopback only (127.0.0.1) — not reachable from your network by default.
• Anti-CSRF / anti-DNS-rebinding: every mutating request is checked for a same-origin Origin and an allowed Host. "localhost-only" is not enough on its own — any open browser tab can POST to localhost; most localhost tools ignore this. AI Garage doesn't.
• Optional token for sharing. Create ~/.config/localhost-control/config.json:

  { "token": "a-long-random-string" }

  Then every action requires it (status view stays open).
• ⚠️ The panel can run processes — don't expose port 7777 to the internet without a token.
• The services.json registry stores start commands; whoever can write to it can run code. Don't give untrusted processes write access.

Details that other tools skip

• Zero data leaves your machine. No cloud, nothing collected — nothing to leak.
• Freeing a port won't nuke your database. Soft SIGTERM first (lets Postgres/MySQL flush), SIGKILL only after 1.2s, plus a warning on system/DB ports. Naïve "kill-port" tools SIGKILL instantly.
• Keep-alive won't thrash your scripts. The panel supervises and restarts a service only when its port is actually down.
• It won't kill itself (port 7777 is guarded) and cleans up after itself.
• It won't lose your services. The registry is written atomically (+ backup).
• Zero dependencies = no npm supply-chain surface.

Roadmap

• Link-provider picker — ngrok (stable URL) & custom domain, alongside cloudflared.
• "Mission control for agents" — read sessions from Cursor / Codex / Claude Code / Cline locally; connect task trackers (GitHub Issues, Notion).
• Cross-machine view (VPS sessions) with a zero-backend reporter.
• Menubar app (Tauri), drag-to-reorder polish, favicons, "start all".

Contributing

Issues and PRs welcome — especially new UI languages (one object in public/index.html) and process-classification rules. MIT licensed: free to use, fork and ship, including commercially.

License

MIT © Az