LaReview

The deciding trait is that it brings no model of its own. You point it at the coding agent you already run, and the thing you get for free is structure: a plan built from a pull request or a pasted diff, organised so you start where the danger is rather than at the top of the file list. That is the part reviewing large changes actually lacks.

What it will not do is review anything while you are asleep. Pick it if you want a better hour of reviewing. Pick a hosted bot if you wanted the hour back entirely.

The consequence is that nothing here is reproducible between two users. Point it at a different agent and the risk ordering, the issue list and the architectural reading all change, while the interface presenting them looks identical and carries no indication that it might. A checklist that looks authoritative and is not comparable is a specific kind of hazard in review, because people trust checklists.

What it does right is scope the reading. Per-task isolated diff hunks mean a reviewer sees the change relevant to one concern rather than the whole file.

1. Learning from dismissed findings is the correct signal to collect, because false positives are what destroy trust in review tooling and dismissal is the cheapest available label for them. Most products in this class collect nothing. 2. Grouping by flow rather than by file imposes a semantic partition on a diff, which is a stronger organising principle than proximity.

2. No precision or recall figure is published for any of it, so the calibration is a described mechanism rather than a demonstrated improvement. The difference matters here more than usual.

Code review is the hardest segment to enter from the outside, because the buyers already have a bot commenting on their pull requests and switching means changing a habit sixty people share. A free local workbench with no hosted component cannot buy that distribution and has nothing to sell that would fund trying.

Moat: the local-first stance, which appeals to a real but small set of buyers who refuse to send code out. Likely path: the risk-ordering idea is copied by an incumbent within a year. Position: use it personally, do not wait for a company to appear behind it.

Two things here are worth procurement's time. Finished feedback syncs into the pull request where our review already happens, so this adds a step without adding a destination. And team rules are configurable, which means a standard we argue about once becomes something enforced consistently rather than depending on who reviewed.

Against that: no identity integration, no audit trail, no unattended run, and every developer installs a downloaded binary themselves. Approved with conditions: managed distribution, one shared rules file under version control, and a named owner for it.

The local-first claim is narrower than it sounds and still worth having. Repository search runs against my linked checkout, so the tool is not shipping my tree somewhere to index it; the model call still leaves, because it goes wherever the agent I selected sends it. That is an honest division and the row says so.

Permissive licence, Rust source, no daemon phoning anywhere. No MCP client, which for a review workbench I mind less than usual, and no local model path of its own. I would run this on an air-gapped box if the agent underneath it could follow.