Kodus

The deciding trait is that you write the review standard in plain sentences and scope it to a directory. The architectural rule your senior engineer explains twice a month becomes something the bot says instead, at the exact path where it matters, which is a much smaller and more useful promise than a bot that has opinions of its own.

That means it is only as good as the rules you bother to write, and writing them is real work. Pick it when your conventions are known and ignored. Pick Greptile when they are neither.

The failure mode is a rule that does nothing. Instructions are written as prose and applied by a model, and no dry-run, no fixture set and no regression check appears in the documentation. A team therefore cannot distinguish a rule that never fires because the code is clean from one that never fires because the wording is wrong, and both look identical in the pull request.

What it does right is triage: findings arrive ranked by severity rather than as an undifferentiated list, which is the difference between a reviewer and a linter.

1. The scope is deliberately narrow. This reads a change and proposes; it does not apply multi-file edits, so the verification step remains a person reading a diff, which is the loop the team already has. 2. That avoids the hardest problem in the category rather than solving it, and the choice is defensible.

2. Precision is the only number that matters for a review agent, and none is published: no false positive rate, no comparison set, no methodology. The documentation describes behaviour and demonstrates nothing.

The commercial design is legible: give away the reviewer, reserve a marked subset of the source under a paid licence, and sell the hosted version to teams who do not want to operate it. It works only if the reserved subset stays valuable enough to buy, and open-core boundaries have a habit of moving toward the customer over time.

Charging nothing on top of inference removes the easiest revenue line, which is principled and narrows the model. Moat: the rule library a customer accumulates. Position: adopt the free edition, and read the licence boundary before you build a workflow across it.

Four code hosts are supported, and we run three of them, so this is one deployment rather than a per-host negotiation. Running it ourselves means no seat licence at all and inference billed to an account finance already reconciles, which makes the sixty-developer figure a compute line rather than a contract.

A command-line entry point means it runs as a pipeline step, not only as a webhook, so coverage is enforceable. Identity integration and retention terms are not documented for the hosted edition. Approved with conditions: self-hosted only, behind our own gateway.

The licence boundary is stated precisely, which I appreciate more than a vague community edition: everything outside the marked enterprise paths is copyleft, so a fork stays open and stays possible. The model layer takes any endpoint that speaks the common API, which means a server on my own hardware is a configuration value rather than a feature request.

No MCP client, so it will not reach the tools I already run. For a reviewer that matters less than it would elsewhere. This one I would actually maintain.