Multi Agent Orchestrator — AI skill for Claude Code

MAO: Multi-Agent Orchestrator

A personal AI runtime that runs on your own PC. It treats Claude Code, Codex (GPT), DeepSeek, Doubao, Qwen and other models as interchangeable workers: it routes each task by difficulty and cost, accepts work only when its tests pass, and hands failed work to another model with a handoff note. Each call gets only the minimum context it needs, which keeps token use and cost down.

Python 3.11+, standard library only. Built and used on Windows. A full guide in Chinese is in [docs/README.zh-CN.md](docs/README.zh-CN.md).

What you can do with it today

• Give a coding task to your own models from the command line, or from a chat app such as Claude Desktop (MCP). It counts as done only when your tests pass.
• Use the accounts you already have: Claude Code and Codex subscriptions, Gemini CLI, and API keys for DeepSeek, GPT, Claude, Qwen, Gemini and Doubao (orch connect shows what works).
• Start each task on the cheapest suitable model and move it to a stronger one when it fails, crashes, hits a rate limit or gets stuck; the next model continues from a handoff note.
• Turn a larger goal into tasks you approve (orch plan), or let two vendors propose and critique a design and a judge write the decision record (orch decide).
• Run several tasks at once, each in its own git worktree, and merge them automatically or by hand.
• See what every model cost and how often it passed first time (orch report), and exactly what each call was sent (orch context).
• See every agent on one panel: what it is working on, what it used today and this week, its subscription windows, recent success rate and last error; pause or resume an agent, or give a task to another one (orch agents).
• See the Claude Code and Codex sessions you open yourself on this computer, next to orch's own: project, model, last use, whether one is in use now and its tokens (orch sessions). Only times, folders and token counts are read from the CLIs' se