OpenFox

The deciding trait is that it comes to you. First run detects the backend you already have serving models and configures itself around it, then hands you a browser page on a local port. Nobody enjoys writing provider configuration before finding out whether a tool is any good, and skipping that step changes how likely you are to try it twice.

What you should expect is rough edges in exchange. This is a small project with big ambitions about planning. Pick it if local models are the point. Pick Aider if they are not.

An immutable contract plus an unbounded retry is a hang waiting for the right bug. If a criterion is unsatisfiable, because the plan was wrong or a dependency is missing, the described control flow has no exit: it keeps building. No ceiling is documented, no failure state is named, and the shell commands it runs on the way have no sandbox around them.

What it does right is refusing to move the goalposts. Criteria are fixed once agreed, so success is not quietly redefined mid-run.

1. Using the language server is the strongest design decision on this row. A compiler-adjacent tool reports whether a symbol exists; a model reports whether it believes one exists, and the first of those is evidence. Feeding it back immediately narrows the class of errors that survive to the diff. 2. Plans are elicited interactively before execution, so intent is recorded rather than inferred afterwards.

2. No measurement accompanies the approach. The architecture is arguable on its merits and the capability remains asserted.

This is a personal project with product-shaped ambitions, and the two do not usually stay together. Nothing about the design creates a revenue surface: it runs on the user's own hardware, by intention, which is admirable and is also the reason no business could be attached without breaking the premise.

Moat: none, and the audience for local-first tooling is small and self-sufficient. Likely path: steady solo maintenance, or a quiet stop with the source still standing. Position: fine to run, wrong to depend on, and cheap enough to replace when it stops.

The cost model is inverted from everything else I review. There is no per-seat charge, so the spend moves to servers with accelerators and the people who keep them running, which is capital expenditure and a rota rather than a subscription. That can be cheaper at sixty engineers, and it is never simpler.

The runtime floor is a fleet task: Node 24 across every workstation, ahead of a tool nobody has approved yet. It does not run unattended, so it never appears in a pipeline. Not yet: revisit once the inference platform exists for other reasons.

Four local backends recognised by name, and the plugin surface lets me add an auth method or a transport without patching anything I would have to re-apply next release. That is the design of someone who runs their own inference and got tired of forking tools to do it.

Then the licence, or the absence of one. No file in the repository means the terms are unstated, which is a legal hole rather than a permissive grant. I would use this on my own machine and I would not build anything on it until a file lands.