The AI app builder for non developers that refuses to ship AI slop

Those tools give you a visual editor and an AI hint box. You still choose the theme, the font, the card layout, and the hero pattern. mk0r ships the opposite: no visual editor, and an opinionated agent that already refuses the aesthetic choices most non developers would accidentally pick. The opinions live in a plain text file at src/core/vm-claude-md.ts. The file is 2,354 lines long, it is written into the sandbox as /root/.claude/CLAUDE.md and /app/CLAUDE.md before your first prompt arrives, and Claude reads both on its very first tool call. When you ask for a dashboard, the agent has already been told not to produce purple indigo gradients on a white background, not to reach for Inter or Roboto, not to put decorative icons on feature cards, and not to center every hero section with a gradient button. That list is quoted below.

Non developers do not choose templates in isolation. They choose a template, accept its defaults, then have the AI fill in the blanks. Every default becomes an anchor the AI will pattern match on. If the template already has purple to indigo gradients and Inter, the AI will extend them. mk0r skips the template step entirely. The agent is told, in English, what good looks like and what it is not allowed to produce. That moves the aesthetic decision from a dropdown menu the user did not understand onto a written rule the model applies consistently.

src/core/vm-claude-md.ts lines 168 to 176, inside the Design Constraints section of the globalClaudeMd export. The heading is '### Anti-Patterns (Never Do These)'. The six bullets are: 'Purple/indigo gradients on white backgrounds', 'Uniform rounded cards in a grid with icons', 'Generic hero section with centered text and a gradient button', 'Every section looking the same with slight color variations', 'Cookie-cutter layouts that could be any app', and 'Placeholder images or Lorem Ipsum left in place'. Those lines are shipped to the agent verbatim. You can open the repo and verify the line numbers.

A headful Chromium runs inside the same sandbox on an Xvfb display at :99, 1600 by 900, with its remote debugging port on 9222. A Playwright MCP server is attached to Chromium over CDP at 127.0.0.1:9222. In vm-claude-md.ts lines 278 to 283, under 'Browser Testing', the agent is told after every UI change to 1) navigate to http://localhost:5173 via Playwright MCP, 2) take a snapshot to verify the DOM rendered correctly, 3) check browser_console_messages for runtime errors, 4) confirm the component is imported in App.tsx if the page is blank, and 5) not report completion until the browser shows the expected result. That is a closed loop. The agent writes code, looks at the page in a real Chromium tab, and only hands control back if the render matched its intent.

The rule is scoped by a single sentence above it: 'These rules apply unless the user explicitly overrides them.' If you type 'I want a purple and indigo landing page,' the agent obeys the override. The anti pattern list is a default, not a lock. The point is that a non developer who types 'build me a habit tracker' will not get a purple and indigo habit tracker by accident, because the default is the taste file, not the model's underlying aesthetic tendencies.

No. /app is already a Vite plus React plus TypeScript plus Tailwind v4 project. The dev server is already running on port 5173 with hot module reload. When the agent edits files in /app/src/ the page reloads automatically, and the agent's browser check sees the new render in seconds. You never pick a stack, you never install a framework, and you never open a terminal. The agent does all of that inside the sandbox.

Yes. Every prompt commits to a local git repo initialized at /app by ensureSessionRepo in src/core/e2b.ts. You can revert a commit, redo it, and jump between shas using the built in UI. The agent can also push your app to a private GitHub repo that was auto created under m13v/ at session start; the repo url is in /app/.env from turn one. If you want the source, it is real TypeScript on your own private repo, not a Bubble export.

Prompting Claude in a chat window gives you a model with no agentic tools, no browser, and no opinions about your app. mk0r wraps Claude in three things: the taste file above, a running Vite dev server with Chromium attached, and the Playwright MCP that lets the agent open the app and read the DOM. The agent does not have to imagine the page, it can look at it. That is the difference between 'here is the code you asked for' and 'here is the code, I opened it, it renders, I fixed the bug the console complained about, your turn.'

No. The landing page writes a session key to localStorage as mk0r_session_key on first visit. POST /api/vm/prewarm fires on page mount to warm a sandbox in the background via a Firestore pool. By the time you have picked a prompt, there is already a booted VM with Chromium, Vite, Playwright MCP, git, and a pre provisioned /app/.env waiting for the claim transaction. No email, no OAuth, no pricing modal before the product.