Turn one app into everyone's app

My add-ons marketplace already lists a few of the site's own apps as launchable tiles — open the tile and you land in the app. The next move is to make those apps multi-tenant: give one a notion of instances, members, and per-instance data, and the marketplace tile stops meaning "open my app" and starts meaning "spin up your own."

Weather is the first conversion. Any signed-in person 1-clicks a board at /w/<slug>, adds the cities they care about (search powered by a key-free geocoder, or one tap for "use my location"), labels who lives where, and invites family by email to share the view. The public /weather showcase stays exactly as it was — four fixed cities, no account — as the front window.

The unglamorous half of multi-tenancy is the same every time: friendly unique slugs, email normalization, deriving a stable short id for each member. My first multi-tenant app (shared to-do boards) already had all of that, so the real work was extracting it. The slug/email/member-id primitives now live in one framework-agnostic module; the to-do boards re-export them for back-compat and the weather boards import them directly. A new tenant app composes the primitives instead of pasting the adjective/noun word-lists a third time.

On top of that sit three small tables — instances, members, locations — with row-level security on and no anonymous policy at all. The browser never holds a database key; it polls gated API routes, and every one of them resolves the caller to a membership before returning a byte.

The temptation with "now make it configurable" is to copy the pretty component and start hacking. Instead I made the existing dashboard props-driven: it takes a list of locations, a forecast URL, and a storage key, and its defaults reproduce the public page byte-for-byte. A tenant board hands it the board's saved cities and a per-board forecast feed and gets the same animated sky, the same glass cards, the same "who lives here" marks — for free.

I also hoisted the actual upstream forecast fetch into a single module that both the public page and the boards call, so the two surfaces physically can't drift on the query or the shape of the data. One renderer, one data layer, two front doors.

Shipping the feature and finishing it are different jobs. The polish pass added the things you only notice once real people touch it: board cards that show role and a live city/people count, a warmer first-run state than "no boards yet," and a copy-link button so sharing doesn't depend on an email actually arriving.

The hardening pass closed the quiet edges. Tapping "use my location" twice used to add the same place twice — now a pure, tested helper folds coordinates into a ~110-meter cell and rejects the duplicate. There's a per-board member cap so an invite box can't fan out forever, and rate limits on the two writes that have a cost (creating a board, sending an invite). None of it changes what the app does; it just refuses to do the wrong thing.

The point is that the second conversion is cheaper than the first, and the third cheaper than the second. With the tenancy core shared and the renderer already props-driven, turning the next single-instance app into a spin-up-your-own one is mostly "add three tables and a marketplace tile" rather than a rewrite. A personal chore board, a household budget, a trip log — each is now a short hop. Weather just happened to have the prettiest renderer to prove it on.