Kota

The deciding trait is how little arrives with it. Few dependencies, a fast start, and a deliberate refusal to guess what you want, which means the useful version of this tool is the one you build on top of it over a few evenings. That is a real pleasure for some people and a waste of a week for everyone else.

Be honest about which you are. Pick it if configuring things is the part you enjoy. Pick a finished agent if the code was the point.

Recovery is the missing half. Multi-file editing is supported, branch and commit handling is not, and there is no container between the agent and the working tree, so the only safety net is whatever state the developer happened to save beforehand. Tools that write across a repository without owning a checkpoint mechanism put that burden on a human at exactly the wrong moment.

What it does right is starting fast and carrying few dependencies, which keeps the amount of code that can surprise you small.

1. Minimalism and plain configuration are engineering values, not architectural claims about how an agent should reason, and this project is unusually clear that its principles are borrowed from an editor rather than from any theory of the task. 2. What follows is an absence: nothing states how files are selected for a change, and nothing states how a change is checked afterwards.

2. In a category where those two loops determine everything, silence on both is the finding. No benchmark is offered, and none would be interpretable without them.

Two mentions in twelve months is the number that matters, not the stars. It means nobody is talking about this outside the repository, which in a category with hundreds of entrants is the difference between a project and a product nobody found. There is no company, no revenue and nothing to price.

Moat: none. Likely acquirer: none; an embeddable agent crate is a weekend for anyone who wants one. Likely path: a personal tool that stays personal. Position: pass, and watch the mention count rather than the stars.

Being usable as a library inside something we build is the only route by which this reaches production, because then it inherits the controls of the service around it and my team owns the whole surface. Handed to engineers directly, it governs nothing.

There is no identity layer, no directory sync, no activity record and no unattended mode, so it never becomes a stage in delivery I can measure or require. Across sixty engineers there is nothing to license and only model spend to reconcile. Not yet.

Configuration in a real scripting language is the feature I keep asking for. Commands I define myself, hooks at the points I choose, and behaviour expressed as code rather than as keys in a map somebody else designed. Paired with permissive terms and a compiled binary, that is genuine ownership rather than the appearance of it.

What is absent is the outside world: no port for attaching the servers I run, and my own weights are not a destination, so every request leaves the machine. Superb inner loop, closed outer one.