What is codex-claude-orchestrator?

Codex Claude Orchestrator is a plugin that lets two separate AI coding tools, OpenAI's Codex and Anthropic's Claude Code, work together on the same local project instead of forcing a developer to pick one. Codex stays in overall charge: it decides what the user wants, whether handing work off to Claude is worth it, and it independently checks the result afterward rather than trusting it automatically. When delegation does happen, a single long-running Claude Code session acts as a parent that keeps context across related tasks and routes smaller pieces of work to different Claude models, Haiku for searching and quick triage, Sonnet for implementation and debugging, and Opus for review, with a rarely used Fable model reserved for unusually long tasks. The project describes itself as a transport and policy layer rather than a hosted service: there is no central coordinator, no credential relay, and no daemon running somewhere else. It simply launches the user's already-logged-in Claude Code command line tool locally and stores private state about which model is handling which task. If Claude becomes unavailable partway through, the system can fall back to Codex's own native roles, mapped to discovery, implementation, review, and security tasks, but only after Codex confirms Claude's work has stopped and control has properly returned. One of the stated motivations is combining two separate paid subscriptions, a ChatGPT Plus plan that includes Codex and a Claude Pro plan that includes Claude Code, into one workflow, rather than needing multiple accounts with a single provider. The project is explicit that it does not copy login credentials, share accounts, or turn a subscription into an unofficial API, each tool authenticates through its own official client, and the user remains responsible for following each provider's terms. A typical task flows through the system in stages: the user gives Codex a goal, Codex decides whether to handle it directly or delegate, and if delegating, it creates a bounded contract describing the outcome, boundaries, and what evidence would count as proof the task is actually done. The full README is longer than what was shown.