Proval

The deciding trait is that findings arrive sorted by severity as inline comments. That sounds procedural and it is the whole difference between a review bot people use and one they mute: when the important remarks are separated from the pedantic ones, you can read the top of the list and ignore the rest without guilt.

It reviews, it does not fix, so nothing lands in your branch without you. Pick it to shorten the wait for a first opinion. Pick a pair-programming agent if you wanted the change written.

Review agents are judged on precision and this one is built to maximise recall. Several explorers hunting for cross-file problems will surface more candidates than a diff-local reader, and every additional candidate is either a caught defect or noise a human must dismiss. The row records no precision figure, no confidence scoring and no way to suppress a class of finding.

What it does right is choosing when to speak. It waits for a meaningful push or for a draft to be marked ready, rather than commenting on every intermediate state.

1. Segmentation before analysis is the design decision that matters most in automated review, because a model shown one file at a time cannot see the defect that spans two. Grouping first makes the boundary explicit rather than incidental to how the diff was ordered. 2. The grouping rule itself is not stated, and a wrong grouping produces confident nonsense.

2. No evaluation against labelled defects is published. For a category where precision is the only meaningful score, that absence is the finding.

Self-hosted with no hosted option is the honest configuration and the one with no revenue in it. Every competitor in this category charges per repository or per seat for a service they operate, which funds the ongoing accuracy work that makes a review bot tolerable. This has no such engine.

Moat: none, though the copyleft terms do stop a company quietly reselling it. Likely acquirer: none; the incumbents already ship this. Likely path: maintained while the author needs it. Position: run it, own the deployment, expect no roadmap.

A single image deployed onto infrastructure we already run, talking to the code host we already have, is a deployment my team completes in an afternoon rather than a procurement cycle. It works unattended by design, so it becomes a measurable stage in review rather than a desktop habit, and nothing is charged per engineer.

What is missing is administration. No single sign-on, no directory sync, no export of what it commented on and why, so evidence for an audit means scraping the code host. Approved with conditions: our deployment, our credentials, and a documented mute path.

This is the right place to insist on local inference and one of the very few projects that actually supports it. A review agent sees every line of every change, which makes it the worst possible component to hand to somebody else's endpoint, and here the weights can sit on a box in the same rack as the code host.

Copyleft terms mean a company cannot take this closed, and the whole thing arrives as one image I can inspect. What is missing is a protocol port, which for a tool that only reads diffs I can live without.