CLIO

The deciding trait is that it goes where you go. No dependency tree to resolve, nothing to fetch from a package index, so the agent runs inside an SSH session on a box you do not administer, and a container that was never meant to host tooling. Anyone who has tried to get a modern agent onto an old server knows what that is worth.

You are the wrong buyer if all your work happens on a laptop you control, because then the constraint bought you nothing. Pick it for the awkward machines. Pick Aider for the comfortable ones.

The constraint has a bill. Restricting to core modules means the transport, the parsing and every one of the sixteen provider configurations are written and maintained in this repository rather than delegated to libraries with their own maintainers. Providers change their APIs on their own schedule. Each change is a fix that only one project can make.

What it does right is concurrency hygiene. Parallel sub-agents take file locks and git locks and run under rate limiting, which is more care than most tools take before letting two workers touch one tree.

1. The control flow is stated plainly: investigate, propose a plan, wait for approval, then edit, test and commit. Placing the gate at the plan rather than at each edit is a deliberate trade of granularity for tempo, and the documentation says which it chose. 2. Any turn can be undone, so the loop is reversible as well as gated.

2. Memory is separated into short-term session state and long-term cross-project state, which is a distinction most tools collapse. No benchmark is published and none is asserted.

230 stars, no company, no price, and a technology choice that filters the contributor pool down to people who made that choice deliberately. That is a small, committed group, which is a good predictor of quality and a poor predictor of succession.

Moat: none commercially, and the interesting asset is a niche nobody else wants. Likely acquirer: none, and no acquisition path exists for a project with no entity behind it. Likely path: maintained as long as its author has the itch. Position: use it where nothing else fits, and do not standardise on it.

Nothing per seat, so the cost across sixty engineers is provider spend and my time. The item that stops the review is reach: it executes tasks on remote systems over a developer's own credentials, and there is no SSO, no audit log and no retention policy recording which host was touched by whom.

There is no unattended mode, so it never becomes a pipeline step I can gate. A container image exists, which helps the deployment story and not the identity one. Not yet, and the missing piece is an audit trail.

GPL-3.0, which is the licence that keeps a fork honest, and I will take that trade. Skills load from git repositories, so my prompt library is a repo I version rather than a directory I copy, and MCP servers attach on top of that.

The part that matters most: llama.cpp and LM Studio are configured endpoints, so nothing has to leave the machine and the bill can be electricity. GitHub Copilot OAuth is also in the list, which is a strange neighbour for llama.cpp and I am not complaining.