Snow CLI

The deciding trait is that it does not make you choose a home. There are editor extensions for both major families shipping alongside the command-line tool, so the agent you configured in the terminal is the one answering inside your editor, with the same settings and the same behaviour. Most projects pick one surface and leave the other to somebody else.

You are the wrong buyer if you live entirely in one place, because you would be carrying the complexity of the other. Pick it for both. Pick a dedicated extension for one.

Count the surfaces. An interactive interface, a headless mode, a service mode with its own permission model, sub-agents, a team mode, hooks, a skill system, protocol integration, language server tooling and two editor extensions, all maintained from one repository by one person. Each of those has its own upstream that breaks on its own schedule. The arithmetic is not encouraging.

What it does right is gate the dangerous path. Background tasks require approval for sensitive commands rather than running them silently.

1. The context-gathering design is the interesting part. Where most agents in this class locate a symbol by searching text and hoping, this one asks a language server, which returns the definition the compiler would agree with. That converts a probabilistic step into a deterministic one.

2. The consequence is fewer tokens spent on candidate files that were never relevant, and fewer edits made against the wrong definition of an overloaded name. 3. No evaluation quantifies the improvement, and the mechanism is sound enough that the absence is disappointing rather than disqualifying.

1,063 stars for a project of this scope is a good number and an uncomfortable one, because the ratio of work to reward here is severe. Nothing is charged, no entity is named beyond a personal handle, and the roadmap is one person's remaining enthusiasm.

Moat: none. Likely acquirer: none, since there is nothing to acquire. Likely path: it stays remarkable until it stops, with no transition period, which is how single-maintainer projects of this size usually end. Position: use it, and do not put it anywhere a sudden stop would hurt.

One line in the documentation decides this. Third-party relays are listed as a way to reach provider endpoints, which means an engineer following the quick start could route our source code through an intermediary that has never been through a data-handling review. That is not a hypothetical for sixty people, it is a Tuesday.

Zero licence cost, uncapped provider spend, no SSO, no audit log and no retention policy. Approved with conditions: direct provider endpoints only, enforced in configuration we manage, and no relays.

Skills use another agent's format, which means the library I already wrote works here without translation, and that compatibility choice is worth more than any feature on the list. Hooks and protocol services extend it further, so the capability surface is genuinely mine to shape.

On the licence: the repository ships the permissive text, and the forge classifies it as something else, which is the kind of ambiguity that matters to nobody until it matters to a lawyer. Read the file, not the badge.