Short answer: yes. Claude can write application code, run it, and help you improve it. The route you choose determines where the app runs, how people access it, and who handles deployment. We build VibeKit, one of the platforms discussed below.
Watch an app being used and updated
Five More is a trivia game we made for a VibeKit team demonstration. This 47-second recording shows the playable quiz, score sharing and history, and a new round published by its agent after we closed VibeKit.
Read the video transcript
I made a daily trivia game for my friends with VibeKit. I picked Game, chose friends and family, and typed in my idea. It started with a space quiz. Then I asked its agent to find fresh questions and post the next round later. I closed VibeKit, and came back to a new animal quiz. My friends can just open the link and play. No download, no sign-up. I can send them my score and see if they can beat it. My scores are saved online, so I can look back at how I did each round. And I can keep changing the game. Movie questions, harder rounds, a different daily challenge? I just ask the agent. I made the game. Now I get to play too.
The recording demonstrates the hosted-app workflow, not a timed build benchmark or a model comparison. It shows one scheduled run; the demo's schedule is now off. Original TikTok post.
The useful question is which route fits your project. There are three, and they suit different people.
Route 1: Claude artifacts, for interactive tools inside Claude
Ask Claude in the Claude app to make a tool, game, or mockup and it can open an interactive artifact beside the conversation. You can edit and share the result.
Artifacts can retain data between sessions. Anthropic documents personal and shared storage on Pro, Max, Team, and Enterprise plans, with a 20 MB text-storage limit per artifact. That makes trackers and collaborative tools possible, not just disposable prototypes. Check the current Artifacts documentation for storage, sharing, and plan limits.
Choose this route when Claude's hosting and sharing model fits your tool. If you need an independently deployed service, check its requirements for accounts, integrations, data, and domains against the other routes.
Route 2: Claude Code, for working in a real codebase
Claude Code is Anthropic's coding agent. It reads and edits a real repository, runs shell commands, executes tests, and builds multi-file applications. Developers use it to build production software.
You do not always need a local terminal: Claude Code cloud sessions run on cloud infrastructure and can be started from a browser or phone on supported plans. Local sessions are another option. In either case, the project still needs a suitable environment, dependencies, and a deployment destination. Running the coding agent in the cloud is different from hosting the finished app for its users.
Route 3: an app platform that runs Claude for you
Platforms like VibeKit, Lovable, Bolt, and Replit put a coding agent behind a chat box: you describe the app, the agent writes the code, and the platform hosts the result at a real URL. This is the route for "build an app for me": the infrastructure decisions are made for you.
Differences that actually matter between them:
- Which model does the building. Some platforms let you pick or bring your own Claude access; others hide the model entirely. If you have your own Anthropic key, check whether the builder accepts it, because that changes the economics.
- Where the app lives afterward. A real URL you can share, a custom domain, and an export path for the code are the difference between owning an app and renting a demo.
- What iteration costs. The first version is the cheap part everywhere. The real product emerges over dozens of follow-up requests, so per-message economics matter more than the free first build. Real numbers in how much an AI-built app costs.
"Can Claude build an app for iPhone?"
This exact question deserves a precise answer. Claude can absolutely build a mobile web app: built mobile-first, it looks and feels native, and people add it to their home screen where it launches full-screen with its own icon.
A native App Store app also needs a build, signing, and Apple's review. Claude Code can help write native code, and some platforms provide cloud builds and submission tooling. In VibeKit, native creation and App Store shipping depend on what's enabled for your account; publishing uses your own Apple developer account. Check that access before choosing this route. A mobile web app is another option when your customers just need to open a link on their phones.
What Claude builds well, and where projects really fail
Forms, trackers, dashboards, and browser games are useful starting projects because you can define a small first version and test its behavior. A generated screen is only the beginning: submit a record, reload, check permissions, and verify the deployed version before calling it working.
Where projects actually stall is rarely code generation. The common failure points are:
- Pixel-perfect replication. "Make it look exactly like this screenshot" is the hardest request in the category, for every tool.
- Third-party services. Requirements vary by platform. VibeKit provides transactional email without requiring your own email-provider key; taking payments still requires connecting your Stripe account. Check the specific integration before assuming every service needs a separate key, or that the agent can authorize an account for you.
- Vague iteration. "Make it better" burns messages; "move the booking button above the gallery" lands first try.
These are useful checks to make during a build, before inviting real users.
Build your own small app, then improve it
Start in VibeKit, or open the iOS app. Choose a new web app and describe a first version you can test. For example:
Build a five-question trivia game about space. Show an explanation after each answer, save my scores between visits, and let me share a result with a friend. Use sourced facts and give me the deployed URL.
Open the result, play a round, reload, and check that your score is still there. Then ask the same app's agent to add an animal round while keeping your existing scores. Verify the new round and the history before sharing it.
Select Claude in the model picker if that is the model you want to test. Model choice and usage determine the AI charge; check your credits before starting. The goal is a working first version you can keep improving.