Moa X — Development skill for Claude Code

A small, CLI-native take on the 2024 [Mixture-of-Agents paper](https://arxiv.org/abs/2406.04692), pointed at a different job: producing **repo-grounded implementation plans** for coding agents instead of chat answers. The default roster puts proposers from three model labs to work — Google `agy-gemini-pro` (`gemini-3.1-pro-high`), xAI `grok` (`opencode-go/grok-4.5`), and OpenAI `codex-luna` (`gpt-5.6-luna`) — reading the repo in parallel, doing their own web research, and each writing an independent plan. Three refiners—Alibaba Qwen `qwen3.8-max-preview`, DeepSeek V4 Pro, and Anthropic `claude-opus-5`—then refine in broadcast mode (every refiner sees every plan). The shipped defaults and all recommended Web UI modes run Layer 3 through the recorded Codex path with `gpt-5.6-sol` at `xhigh` reasoning. Claude Opus 5 is also available as an aggregator, while Fable 5 is available only as a warning-gated, quota-heavy aggregator option.

Built to run **inside Claude Code** as a skill, or from the local Web UI. Standalone Python works too. The harness ships curated routes across four execution harnesses (`codex`, `claude`, `opencode`, `agy`) and th