Kery – comments on your PR with a video of the feature working

I ship all of my code via different agents, mainly Claude Code and Codex. And with 5 different sessions running on different worktrees, the only way for me to know if something broke is once it gets to the deployment stage or production.

I even use Claude's /pr-review skill and it does catch most of the errors, but most of them are at code level, and it does not really catch what happens at the UI / UX / functionality side. If there's a section entirely rendered orange, your CC skill might just be useless.

So I built Kery; a platform that checks your PRs, reads the diff and descriptions to understand the change intent, opens your preview deployment in a real browser, logs in, and tries to perform the exact thing the PR is for, and provides you evidence of that thing working or not.

The comment is the product. I wanted the artifact to be the evidence rather than an opinion; it reduces the cognitive load from your brain needing to process every PR for what it is, instead you get an image or a video of it working.

The engine underneath is Apache-2.0 and I've open-sourced it, and going to support it for the foreseeable future. Here are a few interesting features of Kery.

* It gets past your login. Clerk, Supabase, Auth0, Firebase, magic links, 2FA codes, and Vercel preview protection. Everything worth checking is behind a login screen, and that's where most browser agents quietly stop. * Video and screenshots of the moment it broke, with a box drawn on the exact element. You watch what happened instead of reading a claim about it. * Nothing to write. No test scripts, no acceptance criteria, no ticket to paste. It works out what to check from the diff. * It drives off the accessibility tree instead of CSS selectors. Stagehand handles the case where the DOM moved and the element is semantically still there. Agent-written UI churns too fast for selectors to survive. * It gets quieter every run. A triage pass dedupes findings and remembers the ones you called false positives, with confidence decay so old memory doesn't compound. * MCP server, so Claude Code or Cursor can run a check inline without you leaving the session.

Where it stands - roughly 60% of the planted bugs are caught on my benchmark, and it finds a few real bugs as well alongside a sprinkle of false positives. It's not exactly where I want it to be, and I'd rather say that here than have you find out on your own repo. The number that annoys me most is about two invented findings per run on a page with nothing wrong, down from three and a half, still too high.

That's also the thing I most need help with. If Kery tells you something is broken and it isn't, replying here with which one is genuinely more useful to me than a compliment; false positives are the reason tools like this get switched off in two weeks, and I only know about the ones I've been able to imagine.

Nobody except me has run this against a repo I don't own, so I'd rather hear that it fell over than that it's neat.

Engine, Apache-2.0 — npx keryai sets up docker-compose and a dashboard on localhost:11111, your own keys, runs against any web app: https://github.com/Kery-HQ/Kery

Kery Cloud — GitHub App, connect a repo and your next PR gets checked. 1,000 credits free on signup: https://www.kery.dev/

Happy to answer anything about the agent loop, the auth handling, or the model setup.