MCO

The deciding trait is that nobody blends the answers. You give one task to several agents you already have, and what comes back is their raw output, next to each other, with the judging left to you. That is the honest version of a multi-model tool, and it is also the version that costs you attention every single time.

You will hate it if you wanted a decision. Pick it when the stakes are high enough that reading three opinions beats trusting one, which is most architecture calls and almost no small edits. Pick a single agent you trust for the daily work.

Parallel is the risk. It dispatches implementation work to several coding agents on the same machine, and nothing in the row describes isolation between them: no sandbox, no per-agent worktree, no lock. Two agents told to fix the same bug in the same checkout is a merge conflict you did not ask for, at best.

Cost compounds the same way. Every run bills every provider you selected, for the same task, whether or not you use more than one answer. What it does right is refusing to pick a winner for you, which at least makes the waste visible.

1. The review path is read-only and separately named, which matters more than it sounds: comparing several agents on a task that cannot write is a controlled condition, and comparing them on a task that can is not. Holding the side effects at zero is the difference between an observation and an anecdote. 2. The comparison is between whole agent harnesses, not models, so scaffold and model vary together and no single-variable conclusion is available from a run.

2. Nothing defines how disagreement resolves. Three answers arrive with no rubric, no scoring and no published methodology for reading them, which leaves the hardest step entirely to the reader.

This is a wrapper, and the interesting question is where the value lands. Every run monetises somebody else's subscription; MCO itself never touches an invoice. Five hundred and nine stars on a repository with no company behind it is a popular utility, not a business, and there is nothing in it that a first-party vendor could not ship as a flag.

Moat: none. Distribution belongs to the CLIs it orchestrates, and each of them would rather you used only theirs. Likely path: absorbed as a feature or abandoned when one vendor ships parallel review natively. Position: use it, do not expect it to become a company.

The arithmetic is the problem. Sixty engineers running this means sixty engineers each holding several vendor subscriptions, because it orchestrates the CLIs already on the machine rather than replacing them. My cost is not this tool, it is the four tools it needs, multiplied.

It does run in CI and it emits JSON, so a review step could gate a pull request, which is the first thing that would justify the spend. Against that: macOS and Linux only, so my Windows contingent is excluded from day one, and there is nothing resembling a central console. Approved with conditions: one shared CI runner, not sixty desktops.

MIT, which is the licence I stop worrying about. The part I did not expect is the bundled Skill: it ships an interface for another coding agent to drive it, so the orchestrator is itself scriptable by the thing it orchestrates. That is a small, slightly recursive idea and it is the right one.

What is missing is protocol. No MCP client, so nothing I run as a server is reachable, and no local models either, since it delegates to installed vendor CLIs rather than to an endpoint. I own the wrapper completely and none of the agents inside it. Grudging respect for the wrapper.