Baz

The distinguishing habit is that it reviews plans, not only diffs. Catching a wrong approach while it is still a paragraph is worth more than catching it as four hundred changed lines, and it is the one thing here that changes how a team works rather than just adding a commenter to the pull request.

Pick it if your engineers already write plans and would benefit from a second reader on them. Pick CodeRabbit if they do not, because a plan reviewer with no plans to read is an expensive way to get ordinary review comments.

The gap is disclosure. No backbone model is named and there is no way to bring your own, which means the component doing the judging is opaque and can be swapped without notice. A review bot whose behaviour shifts silently is one your engineers stop trusting, and you will have no changelog to point at when the comment quality moves.

Ask for the model policy in writing before a rollout. What it does right: it connects to GitHub, GitLab and Azure DevOps rather than assuming one host, which is unusual in this category and matters to anyone with a mixed estate.

Diff-local review misses the class of defect that matters most: a change correct in isolation and wrong given a caller three modules away. Traversing the repository addresses exactly that, at a token cost proportional to the traversal, so the design choice is principled and expensive in the same breath.

What is missing is measurement. No false-positive rate, no recall figure and no methodology accompany the claim, and for a review product those numbers are the product. A reader is asked to accept that broader context yields better comments, which is plausible, undemonstrated, and the only question that matters.

Two signals point the same way. Charging from the first user with no free plan removes the funnel most competitors depend on, which implies a sales-led motion and a belief that the buyer is an engineering leader rather than a curious developer. Pairing the seat with consumption credits at a hundredth of a dollar gives them a lever to raise revenue per account without renegotiating anything.

Moat: the review history accumulating per repository. Likely acquirer: a code host or a developer platform that wants review coverage. Position: interesting, and priced like they know it.

The seat price is $30 per active developer, which is $1,800 a month for sixty and predictable only once the word active has a definition in the agreement. Contractors, interns and anyone who pushed twice in a quarter all need to fall on a known side of that line, otherwise the invoice becomes a monthly argument.

The private mode and self-hosted deployment sit in the enterprise tier, which is where our security review would land regardless. It runs unattended against pull requests, so the workflow fit is genuine. Approved with conditions: define active, get retention in writing, pilot on two repositories first.

There is nothing to read and nothing to change. No source, no self-hosted community build, no way to point it at a model I chose, no protocol support in either direction. If it starts behaving badly I file a ticket and wait, which is the arrangement I spend my life avoiding.

One grudging note. It exposes a CLI and can be driven from Claude Code, Codex and Cursor, so at least it reaches into the environment I already work in rather than demanding I live in a dashboard. That is the difference between a closed tool I dislike and one I refuse outright.