Skip to content

Prerequisites

You need five things before this Skill is useful.

Codex

Codex is the controller.

It plans the work, calls MCP tools, launches workers, reviews logs, checks diffs, and gives the final answer.

Claude Code

Claude Code is the worker process.

The orchestrator launches claude with a selected role, prompt, permission mode, and provider environment.

Check that the command is available:

bash
claude --version

You can also set a custom binary:

bash
export CLAUDE_CODE_BIN="/path/to/claude"

On Windows:

powershell
$env:CLAUDE_CODE_BIN = "C:\Path\To\claude.exe"

CCSwitch

CCSwitch is the local profile registry and model router.

The orchestrator reads CCSwitch in read-only mode. It does not rewrite global CCSwitch state. It injects the selected provider variables only into the launched Claude Code process.

Discovery checks:

  • $env:CCSWITCH_HOME
  • $env:USERPROFILE\.cc-switch
  • the current user home

Multiple models in CCSwitch

Configure more than one Claude-compatible model inside CCSwitch.

The Skill becomes stronger when each model has a job:

Model classGood for
Strong reasoning modelarchitecture, review, security
Strong code modelimplementation and refactoring
Fast low-cost modelquick checks and small planning tasks
Stable review modelrisk, testing, compatibility
Fallback modelresilience when a route is missing

After adding models, rerun:

bash
python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" score-models
python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" write-auto-policy
python "$CC_ORCHESTRATOR_HOME/cc_orchestrator.py" write-reports

Python 3.10+

The bundled CLI and MCP server are Python programs.

Check:

bash
python --version

If your system uses python3, use that command in the examples.

MIT licensed. Not affiliated with OpenAI, Anthropic, Claude, Claude Code, or CCSwitch.