wtf is claude-model-relay?

Claude Model Relay is a small set of skills and subagents for Claude Code that splits a coding task across three different AI models instead of using one model for everything. The idea is that the most expensive, most capable model is best spent on planning and on reviewing the finished work, while a mid-tier model can handle the actual writing of code once it has a clear plan to follow. The process runs in three stages. First, a read-only planner using Fable 5 looks at the real codebase, writes a step-by-step plan naming exact files and pieces of code, and classifies the task as either high complexity, meaning architectural changes, concurrency, security work, or an unclear root cause across many files, or standard complexity, meaning a contained and well-specified change. Second, an implementer stage carries out the plan, using Opus 4.8 for high complexity tasks and Sonnet 5 for standard ones, then runs the plan's own verification step and reports honestly if the actual work turned out to differ from the plan. Third, a separate read-only reviewer, again using Fable 5, independently re-checks the verification and the real code changes, then returns either an approval or a request to revise, capped at two revision rounds before the disagreement gets handed back to the person who asked for the change. Because each model has its own usage limits, the project also defines fallback chains, so if one model runs out of usage the task moves to a specified backup model instead of quietly failing, and the tool always reports out loud whenever this kind of fallback happens. Installation is a single command that copies one skill file and three subagent files into Claude Code's configuration folders, with alternative install paths through a Claude Code plugin or a manual git clone. The project is released under the MIT license, which allows free use, including commercial use.