Skip to content

AI code review for GitHub

Only what it can prove.

An AI reviewer for engineering teams on GitHub. It comments only when it can name a concrete failure scenario, and a second pass has checked that scenario against your code. No scenario, no comment.

Free for open source · No card · Flat price per PR author

THE PR

  • candidate

    nit: trailing whitespace on line 12

  • candidate

    consider adding a comment here

  • candidate

    prefer const over let

+ 3 more dropped before posting

superdiffsPOSTED

retry creates a new idempotency key → customer charged twice

✓ scenario checked against the code

Seven candidate comments. Six had no failure scenario that held up against the code — dropped. One did.

✗ dropped before posting✓ checked — posted

superdiffs / checkoutPR #482

Fix: retry the charge when Stripe times out

OPEN
ConversationFiles +18 −4
MK
mira.kopened 12m ago

Stripe timed out twice yesterday. This retries the charge — review before I merge?

src/checkout/charge.ts+18 −4
138async function chargeWithRetry(order) {
139 for (let i = 0; i < 3; i++) {
140- const key = order.idempotencyKey;
140+ const key = randomUUID();
141 const res = await stripe.charge(order, { key });
7 candidates in · press run

illustrative PR · the real pipeline is replayable from a stored trace

Your code

Never used for training. Retention published.

Each review reads one repository in a disposable sandbox, with a one-hour, read-only token. The code snippets kept with findings are deleted when you uninstall; every retention window is in the table.

Read the retention table

Where it works

GitHub, and only GitHub.

It reviews pull requests on github.com through a GitHub App. No GitLab, Bitbucket or Azure DevOps, and no IDE plugin — the review lives in the pull request.

How setup works

Open source

Free for public repositories

Public repositories under an OSI-approved license get the full reviewer — every depth, no card, no trial clock.

What qualifies

Built on GitHub · Anthropic · E2B · Railway · Paddle — what each one receives

Who it is for

For teams that turned the last review bot off.

A reviewer that comments on everything gets the same treatment as a noisy alert: twenty comments on a small change, the same nit again after every push, and the one real bug buried where nobody reads it. People stop reading, and then someone uninstalls it.

superdiffs is built the other way round. It says less, and backs up what it says.

A good fit

  • Roughly 10 to 200 engineers opening pull requests on GitHub.
  • You tried an AI reviewer, and your team learned to scroll past it.
  • You want fewer comments you can trust, not more coverage you have to triage.
  • You would rather pay per pull request author than per review run.

Not a fit yet

  • Your code lives on GitLab, Bitbucket or Azure DevOps.
  • You need SSO/SAML, an audit log export or self-hosting today.
  • You want the reviewer inside your editor rather than in the pull request.
  • Procurement requires a SOC 2 report. There is none yet.

Not sure? The comparisons say where other reviewers are ahead, with sources.

What is different

Built to be read, not muted.

01

Quiet by default

When nothing survives the checks below, the pull request gets a short summary and no comments. The default depth posts minor issues and up, never more than 20 comments.

nothing held up → summary only
02

A failure scenario on every comment

A finding has to say how the code breaks: the inputs or state, then the wrong result. If it cannot, it is not posted. You can check the scenario yourself in seconds.

given X, Y breaks — or nothing
03

Checked against the code first

The code a finding quotes must be in the diff, and a second pass re-reads the claim against that evidence — it can open other files in the checkout when it needs to. Rejected findings are dropped, not softened. The second pass is the same model family: a strict filter, not independent proof.

find → check → post
04

Never tells you twice

Findings are fingerprinted by content, not by line number, so a finding you resolved is not posted again on the next push — through rebases and force-pushes too.

resolved → remembered → suppressed
05

Priced per author, nothing metered

Team is $19 per pull request author per month, billed annually ($24 monthly). No per-run charges, no per-file overage, no Actions minutes. Past 30 reviews in an hour, a review waits; it is never billed.

bots and outside contributors never billed · how others bill
06

And the table stakes

It resolves its own thread when a push fixes the code, learns from remember: … replies and @superdiffs resolve <reason>, and reads the guidance your repository already keeps (CLAUDE.md, .cursorrules, Copilot instructions). Other reviewers do these too; they are included, not the pitch.

teach it in the thread · how

GitHub only — no GitLab, Bitbucket or Azure DevOps, and no IDE plugin.

Team · per pull request author

$19 / seat

per month billed annually ($24 monthly). Bots and outside contributors are never billed, and nothing is metered per run.

To start

14 days

of Team on every new installation, with no card. When it ends you are on Free unless you subscribe.

GitHub permissions

Read-only code

contents: read · writes only to pull request conversations and checks — it cannot push to your branches.

drawn, not decorated — every number above comes from the code: the price catalog, the trial setting, the token permissions

How it works

Three steps, then it stays out of your way.

No new review workflow to learn: the reviewer shows up in the pull request you were already looking at. Setup takes about five minutes, most of it GitHub’s install screen. It reads your code and writes only to pull request conversations and checks — it cannot push to your branches.

  1. 1

    Install the GitHub App

    Pick an organization or a few repositories. Installing starts a 14-day Team trial, no card, and queues reviews of up to three of your most recently updated open pull requests, so you see a real review straight away.

    contents: read · pull requests, issues, checks: write

  2. 2

    It reads the diff, and what the diff depends on

    The changed code is checked out in a disposable sandbox, read alongside the guidance your repository already writes for reviewers (CLAUDE.md, .cursorrules, Copilot instructions). On Team, the trial and open source, deterministic checks look for committed secrets, conflict markers and leftover debugger statements.

    configured in .superdiffs.yaml, committed with your code

  3. 3

    It posts only what held up

    Inline comments anchored to the right line, each with its failure scenario and, when the fix is unambiguous, a suggestion you can commit. Plus one summary comment and a check run. Everything else is dropped without a word.

    a clean pull request: a summary, and nothing else

Steer it in the thread

9 commands, in the pull request.

@superdiffs review@superdiffs full-review@superdiffs summary@superdiffs resolve@superdiffs pause@superdiffs resume@superdiffs ignore@superdiffs config@superdiffs help

commands run only for people with write access to the repository — text in a diff cannot issue one, and a pull request description can carry only ignore. Reply remember: … in a thread to teach it a rule.

Design partners

Team free for 6 months, for the first 10 teams.

There is no logo wall on this site because we have not earned one yet. What we have is a reviewer we are willing to measure in public, and room for a few teams who want to shape it while it is still cheap to change.

What you get

  • The Team plan, free for 6 months, for every pull request author on your team.
  • A direct line to the person who writes the reviewer, not a support queue.
  • A monthly look at your numbers together: what it posted, what your team acted on, and what it got wrong.
  • Anything we publish that draws on your repositories, you see and approve first.

What we ask

  • A real repository with real pull request traffic.
  • Honest, specific feedback when a comment wastes your time.
  • About 30 minutes a month to talk through what it got wrong.

Put it on one repository.

Install it where a noisy reviewer would annoy you most. If it talks too much, uninstall it — that deletes the code snippets and learnings we hold — and tell us exactly which comment lost you.