CodeJ

The deciding trait is which desk it fits on. Packaged builds exist for Windows and for Linux, and the project says plainly that a Mac build is not published. Almost every tool in this category was written by somebody on a laptop with a fruit on the lid, so a terminal agent that treats Windows as a first destination is a genuinely different offer.

Everything else is competent and familiar rather than exciting: it reads the repository, plans, edits, runs things and asks first. Pick it for the platform. Pick anything else if you have a choice.

Bundling is the cost of the architecture. A single download that carries its own execution environments removes an installation problem and creates a supply problem: the versions inside are frozen at build time, and nothing in the row records how quickly a rebuild follows an upstream fix, or whether a user can substitute a patched runtime themselves.

What it does right is refusing to be a thin shim. The loop, the tool pipeline and the extension surface are the project's own code rather than borrowed behaviour, so a defect here is fixable here.

1. Most agents in this category end a task when the model stops producing tool calls, which is not a criterion. Requiring evidence before a plan step is considered finished converts completion into something checkable, and pairing it with declared termination states means the loop has named exits instead of one implicit one. 2. Task dependencies and sub-agent authorisation give delegation a structure to inspect.

2. No measurement accompanies any of it. The stopping rule is a design claim, and only a run against real tasks would show whether the evidence bar holds.

Below a hundred stars there is no market signal to read, only an author's own conviction, and the measured adoption here is nil. There is no entity, no paid tier and no distribution channel that anyone would pay to reach, so nothing about this can be bought, funded or wound down in an orderly way.

Moat: none. Likely acquirer: none exists; the realistic outcome is the author moving on or being hired for unrelated work. Likely path: a strong first year of commits, then silence. Position: pass, and revisit if a second maintainer appears.

The install line is where this stops. Fetching a script from a domain and executing it is not a thing my security team permits on managed hardware, and repackaging it for the fleet is work nobody has budgeted. The licence costs nothing, so the only recurring number is model spend on keys we would issue.

Beyond that there is no identity integration, no directory sync and no central record of what an agent touched. It cannot run in a pipeline, so it never becomes a stage I can gate a release on. Not yet.

Pointing it at an arbitrary compatible endpoint is the flag that matters, because it means the provider list is a default rather than a wall, and the permissive licence keeps the fork open if the author disappears. That is a decent ownership floor for something this young.

Then it narrows. Extending it means writing against a runtime I would not have chosen, my own weights are not a supported target, and there is no port for the servers I already run, so the tools I have built stay outside. Readable, extensible, and only on its own terms.