CodeAlta

The deciding trait is what happens when you are faster than the agent. Send a second prompt while the first is still running and it is queued or used to steer the work in flight, instead of being dropped or garbling the turn. That sounds small until you have lost a thought to a busy spinner three times in an afternoon.

What you accept in exchange is software the authors themselves call unfinished. Pick it if the runtime it installs into is one you already have. Pick a JavaScript CLI if adding a second toolchain is the real cost.

The combination is what matters. Tool calls are carried out by host processes against the real filesystem, and the project states plainly that it has not reached a stable release. Either fact alone would be ordinary. Together they mean the least-tested part of the system is the part with the most reach, and there is no isolation layer to absorb a mistake.

What it does right is disclose the maturity in its own documentation instead of letting a user discover it. That is rarer on this board than it should be.

1. Owning session state locally rather than deferring to a provider's thread abstraction is the decision that lets the same conversation survive a provider swap, and it is the reason this design should outlast individual API shapes. 2. Serialising prompts through a single actor gives the ordering guarantee that most implementations approximate with a lock and get wrong under cancellation.

2. Compaction is described without a stated policy for what survives it, which is the detail that determines whether a long session degrades gracefully. No measurement accompanies any of this.

Projects built by a respected individual in a specific language community follow a predictable curve. Quality is high because the author has taste and standards, adoption comes from an existing following rather than marketing, and continuity depends on one person's interest holding. There is no company, no tier and nothing that produces revenue.

Moat: the author's standing inside one ecosystem, which is real and not transferable. Likely path: continued personal maintenance, or a platform vendor hiring the author and the project slowing. Position: use it, pin the version, and treat yourself as the support contract.

The distribution story is unusually easy for this category. One command through a toolchain my desktop team already manages, and the provider options include subscriptions that are on the invoice rather than a new vendor relationship. That removes two of the three things that normally block a rollout.

The third is still there. No administrative console, no directory login, no audit trail and no unattended mode, so sixty installations are sixty independent configurations. Approved with conditions: the platform group packages it, provider selection is set centrally, and nobody configures a custom endpoint on their own.

Compatibility with an existing skill folder format is the detail that decides this for me. The instructions I already wrote for another tool work here without translation, which means trying it costs an evening rather than a migration. Plugins load from my own machine, so extending it does not go through anybody's marketplace.

Tool servers attach through a documented configuration file, and the licence is about as permissive as they come, so a fork stays legal and cheap. The one gap is inference: every provider on the list ends up somewhere that is not my hardware.