cc-orchestrator

Let Claude Code orchestrate Kimi (write code) and Codex (review code) as sub-agents via MCP.

Why

Claude Code is great at architecture and decision-making, but burns your Claude quota on code generation too. This package wires up Kimi and Codex as MCP tools so CC delegates the heavy lifting — saving 50–70% of Claude token usage on code-heavy tasks.

Architecture

You
 └── Claude Code (CC)          ← architect, task router, result synthesizer
       ├── mcp__kimi__kimi_task    ← Kimi (kimi-for-coding)  writes code
       └── mcp__codex__codex       ← Codex (gpt-5.3-codex)   reviews code

Flow:

1. You give CC a task
2. CC breaks it down
3. CC dispatches coding subtasks → Kimi writes files
4. CC dispatches review          → Codex finds bugs
5. CC synthesizes, decides next step
6. Repeat until done

Token cost per task:

Claude (CC)  = architecture + routing + synthesis  [small]
Kimi         = code generation                     [Moonshot quota]
Codex        = code review                         [OpenAI quota]

Requirements

macOS and Linux supported. Windows: use WSL.

Install

git clone https://github.com/Craybreeding/cc-orchestrator
cd cc-orchestrator
bash setup.sh

setup.sh will:

• Auto-detect or install Codex / Kimi / mcp
• Copy kimi_mcp_server.py to ~/.cc-orchestrator/ (permanent location)
• Write MCP config into Claude desktop app
• Warn if Codex or Kimi are not logged in

Custom install path:

bash setup.sh --install-dir ~/my-tools

After Install

# If prompted, log in:
codex login
kimi login

# Then restart Claude desktop app

Verify

In Claude Code, say:

Use kimi_task to write a Python function that adds two numbers

If it returns code — you're connected.

Files

Uninstall

Remove the codex and kimi entries from:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Then restart Claude desktop app.