Review a pull request with AI

What you get

• Findings in the format a session's review step records: a title, a severity (critical, high, medium, low or info), a category such as security or reliability, the file, what is wrong and a suggested fix. At most 12.
• A verdict from the release gate's findings check: would be held when any finding is high or critical, passes the findings check otherwise.
• A link to the result, for the pull request's thread or your team's chat. Result pages are kept out of search engines.

What it doesn't do

It reads the pull request's title, description and diff as GitHub shows them, and nothing else. It doesn't clone the repository, run the code or its tests, or see CI results, so it can miss problems that only show up in context, and it can flag something that isn't a problem. Treat the findings as a second reader's notes, not as a decision. Files GitHub shows no text diff for (binary files, or files too large to diff) are listed as not reviewed.

Limits

• Public repositories on github.com only. A private repository looks the same as a missing one, and is refused.
• Up to 50 changed files, 2,000 changed lines and a 60,000-character diff. A larger pull request is refused rather than reviewed in part.
• 5 requests an hour from each network, a daily number of reviews for all visitors together, and a pause when the platform's monthly AI budget is used up. When you meet a limit, the page says which.
• One review per pull request commit: asking again for the same commit shows the same result at no cost.

Prompt injection

A pull request is written by whoever opened it, so the tool treats all of it as untrusted data. The model gets no tools: it can only answer. The pull request is fenced in the prompt with a marker made fresh for each review, the model is told never to follow instructions inside it and to report any that try to steer a reviewer, and the answer is shown here as plain text.

In AI CodeOps

In a workspace, AI agents write, test and review each change in your own repository, then hold it at a release gate until people approve; an open finding of high severity or above keeps it held. See AI code review and the AI code review guide.

Questions

Do I need an account?

No. Paste the pull request's URL and the review runs without signing in. Anyone with the result's link can open it.

Which pull requests can it review?

Pull requests on github.com in public repositories, of up to 50 changed files, 2,000 changed lines and a 60,000-character diff. Private repositories, other hosts and larger pull requests are refused with a message saying why.

What does the AI read?

The pull request's title, its description and its diff, as GitHub shows them. It doesn't read the rest of the repository, run the code or its tests, or see CI results. Files GitHub shows no text diff for (binary files, or files too large to diff) are listed as not reviewed.

What does the verdict mean?

It is the release gate's findings check applied to this review: a pull request with a finding of high or critical severity would be held, one without would pass that check. It isn't an approval: the gate in AI CodeOps also checks test results and waits for people to approve.

Is the result kept?

Yes. A review is kept at its link and reused for the same pull request at the same commit, so asking again costs nothing and gives the same answer. Pushing a new commit to the pull request gives it a fresh review. Result pages are kept out of search engines.

Why was my request refused?

To keep the tool free, each network can ask for 5 reviews an hour, all visitors together share a daily number of reviews, and the tool pauses when the platform's monthly AI budget is used up. The message says which limit you met.