dangerousyams/muxer

Muxer is a plugin for Claude Code that routes different parts of a coding session to different AI models based on what the work actually requires. The idea is that an expensive model like Fable handles planning and judgment while cheaper models handle the execution: file exploration on Haiku, documentation and boilerplate on Sonnet, implementation on Opus. The orchestrating model stays in a lean main loop, only reading summaries and plans rather than raw files, which keeps its token cost low even at premium rates. The routing happens in three places. Each agent definition file includes a hard model assignment in its header, so a scout agent always runs on Haiku and a builder always runs on Opus regardless of what the main session uses. A startup hook injects a short policy into every session telling the orchestrator to delegate rather than do everything itself. A third guard hook catches Claude Code's built-in subagents (like file exploration) before they inherit the session's expensive model, and pins them to Opus instead. The plugin includes seven named agents: a scout for read-only exploration, a writer for low-risk edits and copy, a builder for implementation, a reviewer for adversarial verification, an oracle that escalates to Fable for hard decisions, and two external delegates that dispatch work to OpenAI Codex and Google Gemini through their command-line tools, consuming zero Anthropic tokens. The quality rules are deliberately conservative: never hand a whole big task to a cheap model at once, always build one example at a verified tier before fanning out, and never use a cheaper model to review work that a more expensive model built. A task that fails review twice at one tier gets redone at the tier above. After each session, a cost report compares actual spending against what the same work would have cost if run entirely on the main model. The comparison uses API list prices with cache-adjustment. The license is not mentioned in the README.