Rather review than babysit.

The story

One Claude Code session is remarkable. It is also a part-time job. It asks before edits and commands, it stalls, it hits rate limits, and it declares victory early, so you sit beside it pressing yes and checking its work. Run six of them and the part-time job becomes a full-time one.

Loom started from a simple refusal: babysitting is work for software, not for you. So we built a Conductor, a reasoning model you bring your own key for, that plans a goal into tasks, dispatches them across six real Claude Code CLI sessions, watches every terminal, answers permission prompts, recovers stalls and rate limits, rotates pooled accounts when a login hits its limit, and verifies the work before it counts. Your job shrinks to the two parts that were always yours: deciding what to build, and reviewing what got built.

The principle

Delegation only works when you can check it. That is why every terminal in Loom is the real Claude Code TUI you can watch or type into, every change lands in git where the source control panel can show you the full graph, and nothing counts as done until verification re-checks it. We do not ask for trust. We make the work inspectable and let it earn trust on its own.

Where to go next

If you are curious how it feels, getting started takes a few minutes and your first goal shows a full mission end to end. The broader MingLLM documentation lives at docs.mingllm.com.