Build Mode: describe an app, get a working app.
On August 19, 2026 the thing that used to cost $300 a month became free. Grok’s in-app building surface left its Early Beta and shipped to every plan — web, iOS and Android. You type a paragraph; a working app runs in the thread. This lesson is about what that’s actually good for, and the three things people get wrong with it.
01 What changed, precisely
Precision matters here because the name is a mess. SpaceXAI calls two different products Grok Build:
Build Mode, in the Grok app (this lesson) — you describe an app in chat and Grok writes and runs it in front of you. Introduced July 2026 as an Early Beta for SuperGrok Heavy only. Since August 19, 2026: every plan, web and mobile.
Grok Build for the terminal (Lesson 14) — a command-line coding agent with its own model, for people who already write code. Different product, different audience, same name.
Along with the access change, four things arrived that the Early Beta didn’t have: apps can now be published to a real URL, they can call SpaceXAI’s own models from inside their own code, they can be exported to GitHub, and they can be remixed by other people. That combination is what turns it from a toy into something you can hand to a colleague.
02 Before you scroll: which of these fails?
Four one-paragraph prompts. Three of them are real published apps from SpaceXAI’s own launch. Guess which category is the one this tool is bad at — then read the reveal.
A. “A peaceful 3D driving game: an endless forest I can drive through, no missions, no timers.”
B. “An isometric city sim: let me zone districts, lay roads, and grow a town with a live economy.”
C. “A 16-step drum sequencer with 808 and lo-fi kits, tempo and swing, and genre presets.”
D. “A client portal where my customers log in, see their invoices, and pay them.”
The reveal: A, B and C are real — SpaceXAI published all three, each live at its own grok.me link, each from roughly the text above. D is the shape that disappoints. Not because Grok can’t render a login page — it will, cheerfully — but because “my customers,” “their invoices” and “pay them” are three separate systems of record that live outside the app, plus a payments relationship, plus consequences when it’s wrong.
The dividing line: Build Mode is excellent at things that are self-contained — the whole world of the app fits inside the app. It is unreliable at anything whose correctness depends on data and accountability that live somewhere else. Simulations, tools, toys, calculators, visualizations, prototypes: yes. Systems of record: no.
03 Your first build, done properly
Describe the experience, not the code
The failure mode for experienced people is over-specifying implementation. You are not writing a ticket. Say what it should feel like to use, what the person can do, and what the constraints are — SpaceXAI’s own example prompts read like a design brief, not a spec.
[WHAT IT IS: a one-sentence description of the thing]
[WHAT I CAN DO: the two or three actions the user takes]
[WHAT IT SHOULD FEEL LIKE: calm / fast / dense / playful]
[CONTROLS: the sliders, toggles or presets I want exposed]
Iterate in the thread, not in a rewrite
It builds a working version live in the chat, so treat the next message as a change request against something running: “the camera is too tight, pull it back”; “add a reset button”; “make the economy react slower.” Starting over with a longer prompt is almost always worse than three small corrections against a thing you can see.
Decide the audience before you publish
Publishing gives the app its own address on grok.me with three access settings: only you, anyone with the link, or the whole internet. Pick deliberately — “anyone with the link” is not privacy, it is obscurity. Grok also generates a preview cover image for the link, which you can regenerate or replace, and apps shared on X render inline in the timeline with a play affordance for games.
04 The four upgrades worth knowing
- Built-in model access. Turn on the SpaceXAI APIs for an app and its own code can call Grok for chat, images and voice — no keys to create, paste or rotate. Access is per app and revocable. A quiz app can write new questions on demand; a notes app can summarise itself.
- Secrets. Keys for third-party services get stored outside the app’s code. Use this. The alternative — a key pasted into something you published at a public URL — is the single most common way hobby projects leak credentials.
- Export to GitHub. Pushes the whole project to a repo. This is your exit ramp, and the reason to bother at all for anything you might keep.
- Connectors and remix. Connectors pull business data in to make live, filterable dashboards; remix lets other people fork a published app. Both are useful; both widen the blast radius, which is the next section.
05 The eyes-open part
Three things that are true of this and of every tool like it, and that the launch post is not going to tell you:
- A published app is a website you don’t control the hosting of. It lives on SpaceXAI’s domain, under SpaceXAI’s uptime and SpaceXAI’s product decisions. Fine for a demo, a game, an internal calculator. Not fine as the thing your business runs on — not until it’s exported and hosted somewhere you own.
- “It runs” is not “it’s right.” A generated app that renders convincingly can still compute wrong. If yours does arithmetic that matters — pricing, dosages, hours, taxes — check it against numbers you already know the answer to before anyone else touches it. Same discipline as spotting hallucinations, applied to code.
- Don’t feed it what you wouldn’t publish. Whatever you paste into the prompt to make the app real — a customer list, a pricing sheet, a spreadsheet of staff — is going through the same account and the same terms as the rest of your Grok usage. Read Grok and privacy first, and build the prototype on fake data that has the same shape as the real thing.
Before you publish anything
- Is this self-contained, or does its correctness depend on data living somewhere else?
- Have I set access deliberately — only me, link, or public — rather than accepting the default?
- Are any third-party keys in Secrets rather than in the code?
- Did I test the numbers against a case where I already know the answer?
- If I’d be annoyed to lose this, have I exported it to GitHub?
This week’s challenge
Build the smallest tool you personally keep re-doing by hand — a quoting calculator, a shift roster, a unit converter for your trade, a checklist that scores itself. One paragraph in, three corrections, then publish it as only you and use it for a week. At the end, answer one question honestly: did you keep using it, or did you go back to the spreadsheet? That answer is worth more than any benchmark.