Cloi

The trait that decides it is that setup does the thinking for you. It measures your memory and cores, chooses a primary model that fits, picks a fallback, and then checks where each actually landed, which removes the single most tedious part of running agents on your own hardware.

What you accept is the capability of models that fit on a desk, which on hard problems is a real step down. Pick it when independence and privacy matter most. Pick Aider when you want the best answer and are willing to pay per token for it.

The fallback is where the design gives something up. Setup deliberately sizes the larger model to main memory rather than to the graphics card, which is what makes it available at all and what makes it slow when invoked. A stuck turn therefore hands off to a model that answers in minutes, and that is the turn the user is already frustrated by.

What it does right is consent granularity. Writes and shell commands each ask, with once, always-for-this-tool or refuse as the answers, which is a better vocabulary than a yes-or-no box.

1. Explicit stuck-turn detection is a rarity. Most harnesses either escalate on every turn or never, whereas naming six conditions makes the policy inspectable and, in principle, tunable. 2. Validating a response against the actual workspace before display is verification placed correctly, ahead of the human rather than after.

2. Nothing quantifies either mechanism: no false-trigger rate, no measurement of how often validation catches an error, and no comparison against always using the larger model.

A tool defined by needing no provider account has removed every place a business could attach. There is no consumption to mark up, no hosting to sell and no enterprise tier that would not contradict the premise, which makes this permanently a labour of enthusiasm.

Moat: none, and none possible. Likely path: it stays a small well-regarded project, or it stops when its author's interest moves, and the permissive licence means either way users keep what exists. Position: install it, enjoy it, and never let a team practice depend on it continuing.

Data residency answers itself, which is the shortest version of my longest conversation: no provider, no egress, no processing agreement. Licensing costs nothing. That is a genuinely strong starting position for a regulated environment.

The cost moves to procurement of a different kind, because the usable version of this needs capable machines across the team, and the fleet I have was specified for spreadsheets. There is also no console, no policy and no way to see what anyone ran. Approved with conditions: the subset of the team with suitable hardware, and a standard model configuration.

Sanitising the environment before spawning a subprocess is the detail that tells me someone thought about this properly: my exported tokens do not silently become available to whatever the model decided to run. Paths are confined to the workspace, which is the other half of the same instinct.

No account, no key, no telemetry destination, and a permissive licence over a small tree. This is the purest ownership story in the batch, and the only thing I would add is a protocol client so my servers could join in.