$0 forever
Use your existing Claude Pro/Max coding-agent subscription—or an API key. Latch charges only when you choose the hosted gate.
Install from GitHub ↗Open source · independent · human merges
Your agent wrote the PR. Latch tries to break it.
A reviewer attacks the diff. A separate fixer patches the real defects, fights the bad findings, and re-runs the review. Then Latch stops for you.
No new subscription
Use the coding-agent plan you already pay for.Claude Pro/Max today. Latch charges only if you choose hosted.
Runs automatically
No manual review request.Open a PR—or mark it ready—and the review → fix → re-review loop starts itself.
One-time setup per repo · runs in your GitHub Actions · human merges
#412add math-drill generator
✓ CI PASSED
const choices = [answer - 1, answer, answer + 1];
-
Reviewer
Green CI missed this. The correct answer is always the middle tile.
-
Fixer
Fixed in a3f9c1e. Randomized the answer position per problem.
-
Reviewer
Seed the RNG so the test is reproducible.
-
Fixer
Declined. A fixed seed restores the pattern. Leaving this open for a human.
✓
RE-REVIEW CLEAN1 fixed · 1 open for you
YOUR CALL
The automatic loop
Configure once. Then it runs without you.
Every new or ready-for-review PR starts Latch automatically. The only manual action is the one Latch refuses to take: merge.
-
PR opens. Reviewer wakes.
GitHub starts the independent review automatically—no command, dashboard, or review request.
-
The loop turns itself.
The review wakes the fixer. A real fix re-dispatches the reviewer. A bad finding gets a reasoned refusal.
-
Clean—or capped—means stop.
A clean review parks the PR for you. Three failed rounds escalate to a human instead of spinning forever.
- automatic on PR open
- different reviewer identity
- max three fix cycles
- never merges
Why it is different
Different agent. Different instructions. Different blind spots.
Point the reviewer at a model unlike the one that wrote the PR. Then teach it the failures your codebase has already survived.
Latch keeps that institutional scar tissue in a versioned doctrine—not in a magical prompt you cannot inspect. The community doctrine library is the part that gets sharper over time.
skeptical-senior-engineer.md
# Green CI proves only what CI knows. 01 Falsify the PR's claims. 02 Distrust regenerated goldens. 03 Find every producer and consumer. 04 State the concrete failure scenario.
Use it
The whole loop is free. Hosting is the convenience.
$49 / active repo / month
No workflow file, token, or runner to manage. Includes 25 gated PRs. Join the design-partner list.
Join the waitlistThree honest answers
The trust boundary is the product.
Isn't this still AI reviewing AI?
Yes. It only helps if the reviewer is independent of the author: separate context, adversarial doctrine, and ideally a different model family. Latch makes that mismatch a setting instead of pretending two prompts are two independent minds.
What happens when the reviewer is wrong?
The fixer can refuse a finding with reasons and leave it open for a human. The verdict is non-blocking by default, and the loop stops after three fix cycles instead of thrashing.
Can Latch merge or rewrite its own rules?
No. A human always merges. The fixer also skips any PR that changes Latch's workflows or policy, so the loop cannot edit the rules governing itself.