Your AI writes code. malveon reads it back.

• Your plan. A checklist in any format you already use: Markdown, JSON or plain text. Each line is one thing you want built.
• Your prompts. Optional. What you type to Claude Code, Codex, Cursor or Antigravity. Most requests never make it into a plan, so this is where a lot of “done” claims hide.
• Your code and its history. Read straight from your project folder and git.

1. 01Mark the start. malveon saves a snapshot of your code before the AI begins. That’s how it can tell what the AI changed. Anything you had already written but not saved to git is part of the snapshot, so it isn’t blamed on the AI.
2. 02Read the code. It builds a map of your project: the pages, the requests they make, the server addresses that answer them, and the database calls behind them. It does this by reading files. Nothing is started.
3. 03Check every line. For each item on your plan, malveon looks up what the line names (a URL, a file, a component) and follows the chain from the button to where the data is saved.
4. 04Give a verdict. If something is proven wrong, the result is “blocked” and the command exits with an error. A git hook can use that to stop the commit.

The same code always gets the same answer, because nothing in these steps guesses.

Every plan line gets one of these. Each answer points to the file and line it is based on.

The report opens with one verdict. FAILED means something is proven wrong in the code. NEEDS YOU means nothing is proven wrong, but something is yours to settle, such as a request like “make it faster” that no check can cover. VERIFIED BY CODE means everything malveon could check is built and connected.

And it keeps three things apart, on the first line of every report:

• Read from the code. Done on every run.
• Your build and tests. Passed, failed, or not run. Only when you ask.
• Your real app. Never verified. malveon doesn’t run it.

• In the terminal. One command before your AI starts, one after it says it’s done. Best if you want it in a git hook.
• Your AI runs it. Connect malveon to Claude Code, Cursor, Codex or Antigravity, and the AI checks its own work before it tells you it’s done.
• In your editor. The VS Code and Cursor extension checks when you save and puts problems next to your code.

They all run the same checks. Use more than one if you like.

• It never runs your app, opens a browser, or checks a live site. CONNECTED means built and connected in the code, not clicked through.
• It doesn’t prove your business logic is right.
• It runs your project’s own build and test commands only after you say yes, once per project.
• It sends a few anonymous usage signals (your operating system, which command ran, whether a check came back clean). Never your code, file names, plans or prompts. You can turn them off. Details in the privacy policy.