Short answer: test it in two layers. First, make Claude Code check its own work: ask it to write and run tests, and to open the app in a browser and click through it. Then test it yourself the way a user would: do the main task from start to finish, reload the page, enter bad input, try it on a phone, and use the live version, not the copy on your laptop.
We build VibeKit, an app builder whose agent also tests its own pages. Everything said about Claude Code here was checked against Anthropic's documentation in November 2026.
Why can't I just trust Claude when it says the app works?
Because Claude only knows what it checked. Claude Code "reads your codebase, edits files, runs commands, and integrates with your development tools". If it never ran the sign-up form, it does not know whether the sign-up form works.
Anthropic's own advice is to give it something to check against. Under the heading "Give Claude a check to run", the Claude Code best practices say: "If you give it something that comes back pass or fail, Claude can close that loop itself." The examples there include a test file, a build's exit code, a linter, and a browser screenshot next to the mockup.
Step 1: ask Claude Code to test its own work
Tests are small programs that use your app and check the result. You don't need to write them. Ask for them:
Write tests for sign-up, log-in and creating a new item. Run them and fix any failures. If a test is wrong, tell me why before you change it.
Anthropic's docs use the same pattern: "run the auth tests after, and keep going until they pass." When you find a bug, ask for a test first: the docs suggest you "write a failing test that reproduces it, then fix it." That way the bug can't quietly come back later.
Also ask it to run the build. A web app that works in development can still fail to build for the live server, and the build's exit code (pass or fail) is one of the checks Anthropic lists.
Step 2: let Claude click through it in a real browser
Tests check the code, not the screen. A button can pass every test and still sit behind another element where nobody can tap it. Two ways to let Claude see the page:
- Claude in Chrome. Start Claude Code with
claude --chromeand it can open tabs, fill in forms, read console errors and test a web app in your own Chrome. It needs the Claude in Chrome extension and "A direct Anthropic plan (Pro, Max, Ultra, Team, or Enterprise)". It does not work when Claude Code runs on an API key. It also "shares your browser's login state", so test with a test account, not your real ones. - Playwright MCP. Microsoft's Playwright MCP server gives an AI a browser it can drive. It "enables LLMs to interact with web pages through structured accessibility snapshots, bypassing the need for screenshots". It works with any Claude Code sign-in.
Then ask for a walk-through:
Open the app at localhost:3000, sign up as a new user, create three items, delete one, log out and log back in. Tell me anything that looks wrong or shows an error in the console.
Step 3: test it yourself like a real user
Claude's checks catch a lot, but you know what the app is for. Spend fifteen minutes on this list:
- The main job, start to finish. Whatever the app is for (booking, tracking, selling), do the whole thing once, as a new user.
- Reload and come back. Add something, reload the page, close the tab, open it again. If your data is gone, the app isn't saving it where it should. Often it was only held in the page's memory; sometimes a save failed. Tell Claude what you did and what disappeared.
- Two people at once. Open a private window, sign up as a second person, and check you can't see the first person's data.
- Bad input. Submit an empty form, a very long name, an emoji, a date in the past. The app should say what's wrong, not crash or save junk.
- A real phone. Open it on your phone. Can you reach every button with your thumb? Does anything run off the side of the screen?
- The browser console. In Chrome, right-click the page, choose Inspect, then open Console. Red errors there are worth pasting to Claude, even when the page looks fine.
- The live version. Your laptop and the live server differ: different settings, a different database, different secrets. Repeat steps 1 and 2 on the real address after every deploy.
- Money, email and sign-in. Use each service's test mode, and check at least one real email arrives in a real inbox.
What if Claude Code built a phone app?
For an app made with Expo (a popular way to build iPhone and Android apps with JavaScript), run npx expo start. Expo's docs say: "you will see a QR code in your terminal. Scan this QR code to open the app on your device."
That usually opens it in Expo Go, a free app for trying a project on your phone. Expo calls it "a playground app, so it's limited and not useful for building production-grade projects". So features like push notifications or in-app purchases need a real build before you can call them tested.
How do I tell Claude what broke?
A good bug report gets a fix on the first try. Include:
- what you did, step by step
- what you expected to happen
- what actually happened
- the exact error text, copied rather than retyped, and a screenshot if it's visual
Then ask for a failing test before the fix, as in Step 1.
How does VibeKit handle testing?
VibeKit is a different route to the same goal: you describe the app in a chat, from your phone or the web, and the agent builds it. The platform puts each change live at a real address right after the agent replies, unless you ask it to hold. There's no copy on your laptop to keep in step with the live one.
The agent is told never to say "fixed" or "works" because a server answered OK. It runs the check that fits the change: it reads the page for the new text, or takes a phone-size screenshot in a real browser and looks, clicking a button or filling in a form when that's what changed. The screenshot check also reports console errors on load. When you ask it to publish, or tell it something is broken, it publishes the fix itself and then checks the live page.
One case works differently. For an Expo phone app, the live address keeps serving the last published version until the platform publishes the new edits after the reply. So while the agent is still working, the check renders a web preview of the current code, and if the preview can't be built, the check reports that nothing was verified.
What none of this covers: a web preview is not the phone app itself, and sound, the microphone, the camera, haptics, sign-in and payments stay untested until you try them. The agent is told to say so rather than claim they work. To see the app as a visitor would, open its address in a private window. Expo apps open on your phone in Expo Go.
If you're still choosing how to build, our guide can Claude build an app? compares the routes, from Claude's own artifacts to Claude Code to app platforms.
FAQ
Do I need to know how to code to test my app? No. Steps 1 and 2 are requests you type to Claude, and Step 3 is using the app the way your users will.
Does Claude Code test the app by itself? It runs the checks you ask for. Ask it to write and run tests after every change, and to say which checks it ran.
Is Claude in Chrome free? It needs a paid direct Anthropic plan: Pro, Max, Ultra, Team or Enterprise. Playwright MCP is free and open source.
What's the one test I should never skip? Reload. Add something, reload, and check it's still there, both on your laptop and on the live address.
Should I test on the live site or on my computer? Both. Your computer is where you find most bugs fast; the live site is the only version your users see.