Coop is a single-user operating layer that runs the recurring, deadline-bound decisions I would otherwise miss, and keeps a dated record of whether that helped. It is a Node codebase on one Mac: 165 files, 20,632 tracked lines, and no runtime dependencies of any kind. State lives in markdown and JSONL files a human can open, so the system can be audited by reading it.
The split is not across managed services. It runs on one axis: can this piece of work write to disk. Tier one is the Mac under launchd, where five scheduled jobs produce every durable artifact. Tier two is scheduled cloud tasks for work that needs the open web, running with no path to the filesystem. Tier three is a private git remote holding the code and the measurement files that cannot be regenerated.
Model routing turns on one question, and it is not difficulty. It is whether the output will be checked by something other than a model. Work that passes through an invariant, a parser, or a threshold runs on a cheaper tier, because a wrong answer gets caught mechanically. Judgment that nothing downstream validates gets the expensive tier. Every numeric result is computed by the engine, never by a model.
Known limit
The context caps are documented at roughly 20k tokens for monitoring and 50k for execution, and nothing enforces them. Searching the codebase for a token budget or a spend limit returns nothing. It is a written policy with no mechanism, in a system whose whole thesis is that mechanisms beat intentions.