Grok Build through the web interface

Grok · chat interface

Grok Build through the web interface

One-prompt apps with live preview, parallel subagents, Imagine assets, and publish — no local toolchain required.

Web Build Mode packages the coding agent into the Grok chat experience (web, iOS, Android). You describe a product; the agent scaffolds, runs a preview, and iterates while you watch. Heavy subscribers get the full builder + higher ceilings for long multi-agent jobs.

What you are actually driving

  • A remote sandbox with a real project tree, package manager, and browser tools.
  • A live preview that streams the running app into the chat UI as files change.
  • The same agent family as CLI (grok-build lineage) with skills for UI, games, images, and video.
  • Optional publish to a grok.me subdomain or custom domain, plus source export paths (e.g. GitHub) when enabled.

How to prompt for excellent web builds

Web Build rewards product language, not shell scripts.

Strong one-shot prompt
Build a polished client portal for a landscaping company.
Pages: home, services, quote request, admin dashboard of leads.
Tone: earthy, professional, mobile-first.
Leads stored per signed-in admin. Email/password optional; Google/X sign-in ok.
No placeholder lorem — real sample content. Demo-ready.
Iteration prompt (after first build)
Keep the layout. Make the hero calmer (less gradient),
increase spacing on mobile, and add an empty state to the leads table
when there are zero submissions.

Recommended web workflow

  1. State the product in one paragraph (who, what, done looks like).
  2. Name constraints (auth, data, offline, brand colors).
  3. Let the first build finish; use the preview as truth.
  4. Iterate in small product requests (“make filters sticky”, “add dark mode”) rather than pasting huge rewrites.
  5. When happy, publish or export for Hostinger / other hosting.
Hidden web superpower

You can ask the agent to verify in a real browser (screenshots, console errors, mobile viewport). Demand: “Confirm the preview shows real content with a clean console on phone width.” That single instruction prevents blank-page deploys.

Skills the web agent quietly uses

Depending on your request, Build Mode loads specialized playbooks:

  • design-ui — tokens, anti-slop, shadcn patterns
  • building-games — game loop, controls, engines
  • generate2dsprite / generate2dmap — real 2D art pipelines
  • imagine — image/video prompt craft
  • Auth, database, multiplayer P2P when you ask for those features

You rarely name them. Saying “make a twin-stick shooter with custom sprites” is enough. Naming them can still help for power users: “use the design-ui rules strictly.”

Parallel subagents on the web

Heavy builds can spawn multiple subagents (research, UI, game logic, assets) — often in isolated worktrees. You experience this as faster multi-surface delivery rather than as raw shell control. For maximum control over worktrees and hooks, use the CLI.

Publish & handoff

  • Hosted preview / grok.me — fastest share path.
  • Export — pull source into your own git for Hostinger FTP/SSH, CI, or long-term ownership.
  • Production builds must pass typecheck/build; blank deploys are usually asset path / SSR issues — ask the agent to verify the production build, not only dev preview.

When web is the right choice

  • You want a demo or MVP without installing Node
  • You are co-creating with non-engineers in the loop
  • You need Imagine + coding in one conversation
  • You value live preview over local IDE muscle memory

Full contrast with PowerShell/CLI: CLI vs Web.