AI App Maker with Browser Testing: Your App Gets QA Before You See It

The gap in every AI app builder

AI app builders are fast. You describe what you want, and code appears in seconds. But "code that compiles" and "code that works" are different things. A button might render but do nothing when clicked. A form might look right but throw a JavaScript error on submit. An export feature might fail silently because of a CORS restriction.

Traditional AI builders hand you the output and leave you to discover these problems yourself. You become the QA team. mk0r takes a different approach: the AI agent that builds your app also tests it with a real browser, in the same session, before you see the result.

What happens inside the VM

When you choose VM mode, mk0r boots a Freestyle cloud sandbox with seven coordinated services. The AI agent writes your React code, Vite hot-reloads it instantly, and Playwright automates Chromium to test every interaction. Here is the actual message flow between components:

Seven services, one VM, zero configuration

The testing infrastructure is not bolted on as an afterthought. It is part of the base VM image, pre-warmed and snapshot-cached so the entire stack boots in 2 to 3 seconds. Every service is configured and wired together before your first keystroke.

What the AI actually tests

The agent does not just render your app and screenshot it. It interacts with it the way a real user would. It navigates to pages, clicks buttons, fills forms, checks that state updates correctly, reads the browser console for errors, and verifies that exported files are valid. Here is a real testing session for an icon generator app:

Why testing during generation changes everything

When testing happens after delivery, every bug becomes a round trip. You find the problem, describe it, wait for a fix, and test again. Each cycle costs minutes of your attention.

When testing happens during generation, the AI catches its own mistakes. The CORS error in the terminal output above never reaches you. The agent detects it, rewrites the component, hot-reloads through Vite, retests with Playwright, and confirms the fix. By the time you see the app, the export button works.

This is possible because the entire stack lives in one VM. The agent has write access to the codebase (Vite at port 5173), read access to the browser (Chromium CDP at port 9222), and automation access through Playwright MCP (port 3001). There is no context switch between "build tool" and "test tool." They are the same process.

Under the hood: the snapshot hash

Speed comes from snapshot caching. The file src/core/freestyle.ts defines the complete VM specification: the base image setup (Chromium, Playwright MCP version, Node packages), the startup script (Xvfb, CDP, VNC, Vite), the proxy routing, and the ACP bridge. A SHA-256 hash of these four strings becomes the snapshot key. If nothing changed, the VM boots from cache in 2 to 3 seconds. If you upgrade Playwright from 0.0.70 to a newer version, the hash changes and CI automatically rebuilds the snapshot (5 to 10 minutes) so production never hits a cold start.

This design means the testing infrastructure is versioned and deterministic. Every user session gets the same browser, the same automation layer, and the same dev server configuration. There is no "works on my machine" variance because there is no local machine involved.

Watch the AI test your app live

The VM does not run tests in a black box. A screencast relay streams live JPEG frames from Chromium over WebSocket to your browser. You see the AI navigating your app, clicking elements, and watching results update in real time. Frames are captured at 1600 by 900 resolution, JPEG quality 60, every other frame, giving you a smooth ~15 FPS preview of the entire testing process.

If you want full control, VNC access is available on port 5901 through a WebSocket bridge. You can watch, and (in future updates) interact alongside the AI to guide testing toward specific areas of your app.

Who benefits most

Non-developers who cannot debug JavaScript errors themselves. The AI handles the entire feedback loop: build, break, fix, verify. You describe what you want; you get something that works.

Developers prototyping fast who do not want to set up a test harness for a throwaway prototype. The Playwright integration runs automatically. No config files, no test scripts, no CI pipeline.

Teams evaluating AI builders who care about output quality, not just output speed. Browser testing during generation is a quality gate that other builders skip entirely.