Questions about this page

Code, one file at a time, against up to 34 questions from the chapters of Robert C. Martin's Clean Code: names, functions, comments, formatting, objects and data structures, error handling, unit tests, classes and the smells chapter. Of those answers, up to 31 are probabilities and 3 are scores on a five-level scale, so the meters compare between files. Markdown and other prose files sit beside the review and are never judged.

Two. Jev, TypeSafe's evaluation model, answers the whole question set for one window of one file in a single call, through the Vercel AI Gateway; a file longer than one window takes one call per window, and another of each with the comments removed. It returns probabilities and scores, no prose. Luna (openai/gpt-5.6-luna-fast) writes the words. Each file's section comes from Jev's findings and that file's code, and the decision at the top from every file's findings and the pull request's title and description.

No. Each browser tab is one eve session, and it holds the files only for the turn being judged. The page clears that history before every turn, and the session ends with the tab. There is no account and no database. An identical turn can come back from a one-hour cache: that is what the "from cache" note means.

A pull request arrives as one unified diff and is split per file. The question set adjusts: a diff is also asked whether it leaves the code worse than it found it, and only a test file is asked whether its tests are clear. One turn judges at most 24 code files, whichever changed most, and shows up to 10 prose files beside them. Images, lockfiles, generated files and data, config and markup files (JSON, YAML, TOML, XML, HTML, CSS, SQL, templates, Dockerfiles) are skipped. Public repositories only.

Yes, over MCP, at https://clean-code-review.vercel.app/api/mcp. It has two tools: one reviews a public GitHub pull request, the other reviews pasted code or a diff. Both return every answer, the written review and what the call cost. Each address can make 10 calls every 10 minutes, and all callers share $0.25 an hour and $1.00 a UTC day of model time. A server card at https://clean-code-review.vercel.app/api/mcp/server-card describes it for clients that look one up.

Yes, and there is nothing to sign in to. Each browser tab carries its own cap of 50 cents of model time, so one session cannot run up a bill. When a tab reaches the cap the meters keep their last answers, and a reload starts a fresh session. The whole site shares a model budget of $0.40 an hour and $1.00 a UTC day. Once that is spent, new reviews pause until it resets. The source is MIT-licensed, at github.com/frostney/clean-code-review.