Albatross

The deciding trait is the number on the status line. Each turn and each session shows what it cost, so the relationship between a lazy prompt and a real charge stops being theoretical. Once you have watched that figure move you write shorter prompts, which no pricing page ever achieved.

The caveat is that the figure only appears when the price is known, so the moment you point it somewhere unusual the meter goes quiet. Pick it if cost awareness is why you are here and you live in a terminal. Pick Aider if you want a longer track record behind the same idea.

The problem is attribution. A single command swaps the model mid-session, and the log is the same log, so a transcript can contain answers from several providers with no boundary marked between them. When a change turns out to be wrong, there is no way to tell which model produced it.

That matters more here than elsewhere, because switching is the headline feature rather than an escape hatch, so the mixed transcript is the normal case. What it does right is refusing to change the tools and commands across the switch: the interface stays constant, which is the half of the problem worth solving.

1. The routing is described as transparent and as scoring models per task; the word transparent is asked to do the work of a specification. 2. A per-task score implies features, weights and a threshold, none of which appear in the documentation, so a design presented as legible is legible only in its outcome.

2. The Fusion mode for deliberative work is a stronger claim still, since deliberation between models is precisely the sort of design that needs an evaluation to distinguish it from an expensive average. None is published and none is claimed. The architecture is interesting; the record supporting it is a README.

The signal to read is the size of the audience. Stars in the low hundreds describe a project that has not yet found the users who would keep it alive, and there is no company, no paid tier and no funding to buy time while it looks for them. Moat: none.

Likely path: a single-author terminal agent either finds a niche and compounds, or the author's attention moves and the last release becomes the last release. There is nobody to acquire and nothing to acquire them for. Position: enjoy it, contribute if you like it, and keep your workflow portable to something with more shoulders behind it.

The install is the cost. Cargo means every desk needs a Rust toolchain, a compile step and a story for keeping both current, which is a platform project rather than a rollout. Licence cost is zero and the model spend sits on keys the engineers already hold, so finance is not where this stalls.

It stalls on operations. No single sign-on, no directory sync, no audit record, and it does not run unattended, so it contributes nothing to a pipeline and reports nothing upward. Onboarding a mid-level engineer is an afternoon. Not yet: I will revisit it when there is a packaged build.

Four local runtimes named explicitly, rather than implied through an OpenAI-compatible escape hatch, is the part I want to see. My hardware is a first-class provider here and not a tolerated one, and MCP servers attach, so the tools I already run are available inside the session.

MIT means a fork is both legally and practically possible: one binary, no runtime to drag along, and a codebase small enough to read in a weekend. The bring-your-own-key story is complete rather than partial. Grudging respect for a project this young getting the ownership questions right first.