You take the laptop on the train to show a club friend the running-club site, still running from last night. The home page loads at the localhost address, then Trails, and your friend scrolls the list. You add an RSVP for the weekend run, and the list grows from six names to seven. Then you paste a messy group-chat message into the tidy box, "run at 7 tmrw hendricks park 8 miles easy coffee after who's in", and press the button. The box spins, retries, and fails while every other page keeps working. The train has had no signal since the last station. The pages ran on the laptop. The model that turns that message into a clean post runs on a computer somewhere else, and that is where your message needed to go.

Local and remote: the two places software runs

Local means the software runs on the device in front of you, like the site on the laptop on the train. The files, your data, and the computer doing the work are all with you, so losing the signal changes nothing. Remote means the work happens on a computer somewhere else, in a building you will never see. The app on your screen is only a window onto that computer, so without a signal it goes blank.

A data center is a building full of computers that a company owns and runs, a warehouse of computers. Vercel rents you space in buildings like that, and so does the company behind the tidy step. The cloud is those same remote computers under a friendlier name.

Sorting the running-club site

The home page, the trails page, and the RSVP list run on the laptop. Their files sit in the running-club folder, and the terminal you left open runs them. The tidy page differs in one step: the piece that writes the clean post is not in your folder. A model is the part of an AI product that produces the answer, and here it is the part behind the tidy button. A company built it by training on huge amounts of text, and the result is far too big for a laptop, so it stays in a data center. A hybrid is an app that runs partly on the device and partly somewhere else, and the running-club site is one.

The tidy page's box and button sit on the laptop like every other page. Only the last step reaches out, and the message goes with it. How the laptop reaches that computer, and what travels, is the next chapter.

A model is the part of an AI product that produces the answer. It is built by training on huge amounts of text, it is far too big for a phone, so it almost always runs on a computer somewhere else.

Client and server, the builders' words for the two halves

Builders name the two halves. The client is the local half, the phone or laptop in front of the user. The server is the remote half, the computer doing work for you somewhere else. On the train your laptop was the client for the pages, and the terminal you left open did the server's job. Once the site moves to Vercel, the rented computer is the server and your members' phones are the clients. The model stays on its own computer either way.

The Wi-Fi test sorts any app in a minute

The Wi-Fi test means putting the phone in airplane mode, asking the app for its main job and then for something new, and watching what survives. Apple's Notes app passes in full: every note opens, edits save, search works. The ChatGPT app shows old chats, but a new question spins and fails, so it is remote: a text box in front of a model on OpenAI's computers. Granola, the AI meeting notetaker, keeps recording and keeps your typed notes while the polished notes wait for a signal. That is a hybrid, like the running-club site.

No app carries a label saying which half is remote, so you take the network away and watch. Small helpers on the phone now proofread a paragraph or summarize a short note with no signal, but open-ended drafting still goes out to a model.

Put the device in airplane mode and watch what survives: what keeps working runs on the device, what goes quiet runs somewhere else, and an app that half-works is both.

A cache is a saved local copy of remote data, like yesterday's newspaper on the table.

The obvious objection: you are almost never offline

Phones have signal almost everywhere, so the split can read like trivia. The lost signal only made the split visible. Even on a perfect connection, what you can promise a customer depends on which half is remote.

What leaves the device. Work done remotely takes its inputs with it every time; work done on the device keeps them there.
What each use costs. A phone is paid for once; a remote computer is rented and metered, so every press of the tidy button costs you a little. In Hosting, renting a computer on the internet, you learn to read that meter and when to check it.
What survives an outage. When the company running a model has a bad hour, the remote half of every product built on it goes quiet.

Where a feature runs is where your users' data goes

The model cannot tidy text it does not have, so the text has to leave the laptop. The same holds for every remote half: whatever it needs travels out of the device, signal or not. For a meeting tool that is the audio, for a writing helper the document, and for a support bot the customer's question.

When a builder says a step is server-side, that means it runs on the server. Ask about that step: what leaves the device when it runs, where it lands, and how long it is kept.

Asking where a feature runs is really asking where your users' data goes: whatever the remote half needs has to leave the device to get there.

Data, where information lives covers what the server keeps, and for how long. The pre-flight check: secrets, money, data, and the law returns to protecting what travels.

Try it now

This drill takes about ten minutes. The first path costs nothing, and the second a few cents of usage. You arrive with the idea file and the running-club site running on your laptop.

No setup: Put your phone in airplane mode and run the Wi-Fi test on three apps you used today. Ask each for its main job, then for something new, and write one line per app with its verdict and which half went quiet. The list might read:

Notes: local. Every note opened, edits saved, search still worked.
ChatGPT: remote. Old chats showed; a new question spun and failed.
Granola: hybrid. Recording kept going; the polished notes waited for signal.

Those lines are the verdict list; the next chapter opens the app you marked remote.

With your tools: Put the laptop in airplane mode with the site still running. Click through the pages, then paste any message into tidy and watch it fail. Add a line to the verdict list: running-club site, hybrid, the model behind tidy went quiet. Now turn airplane mode off. Claude Code, the AI coding tool we use in this course, fails the test the same way tidy does. It edits files on your laptop, and its answers come from a model somewhere else. Type into Claude Code: "Which parts of this site run on my laptop, and which need a computer somewhere else? What leaves my laptop when someone presses the tidy button?" The answer should mark the page files local, the model remote, and the pasted message as what travels. In Codex or Cursor, the same question goes into the chat box inside the app once airplane mode is off. If nothing is installed yet, The Setup Clinic walks you through it.

Either way, you leave with the idea file one line longer and the verdict list in hand. Add the line: "My idea's remote half is [the AI step]; what leaves the device is [the text the user types]."

Chapter Summary

Local means the software runs on the device in front of you; remote means the work happens on a computer somewhere else, usually in a data center, and the cloud is the same computers under a friendlier name.
The running-club site is a hybrid: home, trails, and RSVP run on your laptop, and the model behind the tidy button runs in a data center.
A model is the part of an AI product that produces the answer; it is built by training on huge amounts of text and is far too big for a phone, so it almost always runs somewhere else.
The client is the local half in front of the user, the server is the remote half doing the work, and server-side means a step runs on the server.
The Wi-Fi test: airplane mode on, ask the app for its main job, then for something new; what survives is local, what goes quiet is remote, and half and half is hybrid.
A cache is a saved local copy of remote data, and it can fool the test, so always ask for something new before you judge an app.
Even on a perfect connection, what leaves the device, what each use costs, and what survives an outage all depend on which half is remote.
Where a feature runs is where your users' data goes: whatever the remote half needs has to leave the device to get there.
Your artifacts are the idea file, now with its remote-half line, and the verdict list from your own phone.
Next, The journey of a request opens the app you marked remote on the verdict list and follows one tap to the far computer and back.

Sources

Granola product and security documentation on device audio capture and cloud-based note generation (last verified July 2026).
Apple support documentation on using the Notes app offline with iCloud sync, and on Writing Tools with Apple Intelligence and on-device processing (last verified July 2026).
Android developer documentation on Gemini Nano on-device summarization (last verified July 2026).
OpenAI help center documentation for the ChatGPT mobile app (last verified July 2026).
Vercel documentation (last verified July 2026).

Where software lives