Soreqai reads the AI agent code in your GitHub repo and tells you what's wrong, with the fix.
Three kinds of AI live in your repo. Usually only one gets reviewed.
One PR for each kind of AI in your repo: the agents you ship, the AI that writes your code, and the AI in your pipeline. Code review was built for the first. Tap the one you think got an approval.
The PRs are examples. The number below is measured.
Before any logic runs, something has already decided how your agents behave. Which model, which prompt, which tools they may call, which MCP servers they trust, and the CLAUDE.md telling a coding agent what it may do here. All of it is read first.
Most scanners never publish how often they are right. We check ours by hand against a written standard, and we commit which findings we will check before checking them, so the score can't be improved by picking the easy ones.
A finding blocks only if it is new, as serious as the level you chose, and sure enough by a number you also chose. No model decides that. Today the verdict is on every report; blocking a pull request straight from the GitHub App is next.
A page per repository. Every finding carries the file and the line, how far it reaches, and what to change. This one is from a real scan.
What it isA call to the model with no timeout, no retry and no cap on how many tokens it can spend. One slow response holds the request open until something else gives up.
How much50 places across 23 files. Seven of them are not on the path your customer hits, and the report says which.
What to doSend these through one configured client that sets all three, instead of calling the SDK directly in fifty places.
One tap installs the GitHub App on one repository. It reads and never writes. Starter credits come with the install, and you get a page for the repository.