MothX

The deciding trait is that it does not make you choose an interface. The agent protocol support hosts the same binary inside VS Code, Zed and the JetBrains editors, so the tool you configure once follows you between the terminal and whichever editor a given project pushes you into. Very little else in this category is willing to be a guest.

The cost is a project that is doing a great many things at once and is young. Pick it if the editor-and-terminal split is a real annoyance for you. Pick a dedicated extension otherwise.

Two things compound. A permissive mode that hands over the whole system is a documented feature, and the shell protection is a denylist of forbidden commands rather than an allowlist of permitted ones. Denylists fail in one direction only: they miss what nobody thought to add, and a model generating shell is a very effective generator of things nobody thought to add.

What it does right is name the modes. Three explicit settings mean a user knows which contract they are operating under, which is better than a single mode with hidden gradations.

1. Branching sessions in durable storage treats a conversation as a tree rather than a line, so an alternative approach is explored without discarding the path that produced it. Compaction alongside it makes the retention policy explicit rather than emergent. 2. This is a stronger persistence design than most terminal agents attempt.

2. The economic claim is unsupported. An extremely high cache hit rate is asserted with no figure, no workload and no comparison, so it describes an intention. The architecture is evidence for the intention, not for the number.

The marketplace is the tell. A free tool with a marketplace attached is a project that has thought about a second act, because a marketplace is the one structure that can carry a take rate without changing the licence of the client. Whether it fills is another question; a thousand weekly installs is a start and not a market.

Moat: the distribution that comes with a prominent regional hosting namespace, which is narrow geographically and real within it. Likely path: regional adoption, marketplace monetisation. Position: watch the marketplace, not the star count.

That HTTP surface changes the conversation. A compatible endpoint means it can sit behind the gateway we already run, where requests are logged, rate limited and attributed, instead of sixty binaries each talking directly to a vendor. That is the difference between a tool I can measure and sixty I cannot.

Everything else is missing: no directory integration, no provisioning, no retention statement, and nothing that runs without a person present, so it is not a delivery step. Approved with conditions: served centrally, keys held by us, and the permissive mode disabled by policy.

This is the most configurable row I have read in a while. A permissive licence, a binary with nothing underneath it to install, local model support so my own hardware is a first-class target, and MCP so the servers I already run arrive as tools. Choosing a low-cost provider as the default rather than the expensive one tells me who wrote it.

Twenty adapters means the model layer is genuinely abstracted rather than one integration with aspirations. I can fork this, read it, and run it with nothing leaving my network.