kimchi

The deciding trait is that you can start simple and grow into the complicated version. Run everything through one model on day one, and when a task is big enough to want a division of labour, switch modes rather than tools. Most agents make you choose that shape at install time and live with it.

You are the wrong buyer if you want one predictable thing that behaves the same way every time, because the interesting mode is by definition several things. Pick it for variety. Pick a single-model terminal agent for consistency.

Delegation across models makes failure hard to locate. An orchestrator hands a task to a role, the role runs on a different model, and a wrong answer could have come from the delegation, the role's model, or the handoff between them. Nothing in the documentation describes attribution, so debugging means running the same work again in the simple mode to see whether it survives.

What it gets right is committing. Work reaches the repository with real history, so at least the output of a confusing run is inspectable.

1. The repository ships a benchmark adapter rather than a benchmark result. That is an unusual and commendable order of operations: the harness that would let someone else measure the tool is published, and the number that would flatter it is not. 2. The adapter aggregates usage across session files, which means cost per task is measurable alongside success.

2. What is absent is any published run. A reader has the apparatus and no findings, so every capability statement here remains asserted rather than demonstrated.

The structure is legible and the structure is the business. Give away the client, run the models yourself, charge for the tokens. 2,220 stars means the marketing works, and a free tier does not exist, which tells me the unit economics were considered before the launch rather than after it.

Moat: the inference relationship, as long as the built-in models are worth choosing. Likely acquirer: a model provider buying distribution, or an infrastructure company buying the funnel. Position: a real business with a real risk, and the risk is that the escape hatch works too well.

I cannot budget this. Billing is by usage against a key, and the documentation states the key requirement without stating a price, so the number I would take to finance does not exist. Sixty engineers times an unknown rate is not a forecast.

It does run unattended, which is a point in its favour, and remote sessions are documented, which is a second procurement question about where those sessions live. No SSO, no audit log, no retention policy. Not yet, and the blocker is a rate card.

Apache-2.0 on the client, which means I can read every path the code takes before it talks to anybody. The part that decides it for me is that an outside provider can be configured in place of the built-in models, so the vendor's infrastructure is the default rather than the only road.

MCP servers attach, so my own tools are available without a shim. Installation covers a shell script and PowerShell as well as the usual, which is more platform care than most projects this size bother with.