CodeBeaver

You will like the shape of the offer. It notices code without tests, writes some, and hands them back as a change you review rather than pushing them into your branch, and when something fails it explains which lines were responsible instead of pasting a stack trace. The deciding daily trait is that it arrives with work already done, so the cost of ignoring it is zero.

Pick it if coverage is a reporting problem you keep deferring. Pick Qodo when you would rather generate tests while writing the code they cover.

This is the failure mode of every generated regression suite and nothing here addresses it. The tool reads what the code does and asserts that it does that, which is correct as a description and wrong as a specification, and the difference only becomes visible when someone fixes the underlying defect and the suite fails them for it. Volume makes it worse: a hundred generated assertions are a hundred small commitments nobody read.

What it does right: it runs the existing suite first, so a real failure is diagnosed against real lines rather than guessed at.

1. The self-run module documents two languages, which is a specific and checkable claim. 2. The hosted service is described as framework agnostic, which is the same claim with the boundary removed and no evidence supplied, and agnosticism about frameworks is exactly where generated tests break. 3. Browser-level tests are specified as natural-language steps in a configuration file, so the test itself becomes a model output evaluated by another model run.

No benchmark and no accuracy figure are published. The observation: the narrower half of this product is the better documented half.

The open module was meant to be the top of the funnel and it never became one. Thirty-four stars is not distribution, and archiving the repository in March 2025 removed the only free thing that might have grown into some. Above it sits a hosted service whose price is not published anywhere, so there is no visible conversion path and no way for an outsider to size the business.

Likely path: a quiet wind-down or an acquihire. Position: do not build a pipeline dependency on this, and if you already have, keep the generated tests and expect to maintain them yourself.

One sentence on the demo: it opened a pull request full of tests and some of them were good. The blocker is that there is no rate card for the hosted product, so I cannot forecast a year, and the self-run alternative requires a provider credential configured per repository, which is a secrets-management job across dozens of repositories rather than one contract. It does attach to pull requests properly, which is the one thing it gets right for us.

Onboarding is an afternoon. Not yet, pending a published price.

The licence is the good part and it is genuinely good: permissive, so whatever I change stays mine, and installation is a single package command with no account and no daemon. That is the correct amount of ceremony for something that writes tests.

Then it asks for exactly one credential and offers no way to substitute the model behind it, which means my own hardware is not an option and neither is any provider I would prefer. There is no tool-protocol client either, so nothing I already run plugs in. A permissive licence around a hardcoded vendor is freedom to do the porting myself.