What you get

omp works like Claude Code, but each kind of work goes to the model you picked for it. omp calls these kinds roles: default for normal turns, smol for subagents, slow for deep reasoning, plan for plan mode and commit for commit messages. You set a model for each role once, in one file.

• One file, three tiers. Frontier on default and slow, a cheap model on smol and commit, and a free local model where it copes. Change any of it mid-session with /model.
• Rate limits don’t stop you. When a model gets rate-limited, its work moves down a fallback chain until the cooldown ends.
• Rules by folder. You can restrict models by directory, say local-only under ~/clients.

Price a week of agent work

The simulator prices a week of work three ways: all frontier, cheap models on subagents and commits, and the same with subagents on a local model. It assumes rough public price ratios: a frontier model costs about 25 times a small one per token, and local costs nothing. The slider sets the number of subagent calls.

Subagent calls and boilerplate make up most turns. So most of the 25x price gap pays for work that never needed the expensive model.