Example page made by Ghostlaunch. This product doesn't exist yet.
Back to the exampleRisk-ranked review for agent PRs
Diffrank is a GitHub App that scores every hunk in your agents' PRs by blast radius, so tech leads review auth, money and migrations first and skim the rest.
Analyzed in memory. Never stored or used for training.
Agents make creation cheap and verification expensive. Every line gets the same attention, so the risky ones hide in the noise.
“My agents open 15 PRs a day and I'm the only one who can review them.”
“Tests were green, style was fine, and it still broke billing.”
“I read 900 lines to find the 20 that mattered.”
“Half the time I skim and hit approve.”
“The review bot left 30 nitpicks and missed the one that mattered.”
Diffrank sits on your pull requests. It points your attention where it's needed. You still make the call.
Pick the repos your agents push to. Setup takes a few minutes and needs no changes to your CI.
Claude Code, Cursor, Codex or a teammate: Diffrank reads each new diff along with the parts of the repo it touches.
A comment sorts each hunk into Read first, Skim or Safe to skip, with a one-line reason for each.
Start with the 40 lines that carry the risk. Skim the rest with confidence, and approve or request changes as you normally would.
Diffrank doesn't add another layer of review to read. It tells you which parts of a PR need a human and why.
Each hunk is scored by blast radius: auth, money, data migrations, public APIs and cross-module contracts. The riskiest lines sit at the top of the PR summary.
Diffrank compares each change with patterns used elsewhere in your repo. It flags an agent calling the user service directly when every other module goes through the shared client.
A queue across your open PRs, sorted by the highest-risk hunk in each. The PR that touches billing comes before the twelve that rename props.
Diffs and repo context are processed in memory to build the ranking. Diffrank doesn't keep your code and never uses it for training.
You're on a 3–20 person team, run coding agents daily, and most PRs end up on your desk.
You trust agents with the boilerplate but need to know exactly where they touched money, auth or data.
You get plenty of comments but no order. Diffrank adds the priority list on top.
Probably not for: Teams not on GitHub. Diffrank is a GitHub App for now.; Anyone hoping for auto-approval. You still review and decide..
Review bots add comments and line-by-line reading doesn't scale. Diffrank gives you a ranked reading list instead.
| Criteria | Diffrank | AI review bots | Reading every line |
|---|---|---|---|
| Tells you what to read first | Yes | No | No |
| Flags code that doesn't fit the rest of the repo | Yes | Partly | If you catch it |
| Keeps up with 15+ agent PRs a day | Yes | Yes | No |
| Output you can take in within two minutes | Yes | Often 20–30 comments | No |
| Code never stored or used for training | Yes | Check vendor terms | Yes |
Run as many agents as you like. You only pay for the people who review their PRs. Every plan starts with a 14-day free trial.
Per reviewer seat. Cancel anytime.
$25/ month
Per reviewer seat, billed yearly. Two months free.
$240/ year
Seats are for reviewers only. Agents and PR authors don't need one.
$25 per reviewer seat per month, or $240 per seat per year. Every plan starts with a 14-day free trial. Agents and PR authors don't need seats, only the people who review.
Diffs and the repo context needed to score them are analyzed in memory. Diffrank doesn't store your code and never uses it to train models.
Review bots comment on everything with roughly equal weight, which gives you more text to read. Diffrank ranks hunks by blast radius and tells you which ones to read, skim or skip. You can run it alongside a review bot.
Treat it as a reading order, not an approval. Each label comes with a one-line reason, so you can check the logic quickly. Anything touching auth, money, migrations or public APIs is never placed in Skip.
Install the GitHub App, choose your repos, and the next PR gets a ranked summary. Nothing to add to CI.
Diffrank reads pull requests, so it works with any agent that opens them: Claude Code, Cursor, Codex, or a human. It's GitHub-only for now.
Get a ranked list of the hunks that need a human on every PR, and get your review hours back.