2 minute read

On March 31, 2026, OpenAI published openai/codex-plugin-cc — an official plugin that lets Claude Code users run Codex reviews and delegate tasks without leaving their session. Within hours it had 3,700+ GitHub stars. This is OpenAI shipping its agent inside a competitor’s tool.

What It Does

Six slash commands, all runnable in the background:

Command Purpose
/codex:review Read-only code review on uncommitted changes (supports --base <ref>)
/codex:adversarial-review Steerable challenge review — questions design decisions and pressure-tests assumptions
/codex:rescue Delegates investigation/fixes to Codex via a subagent
/codex:status Shows running and completed Codex jobs
/codex:result Retrieves output; includes session IDs for codex resume
/codex:cancel Terminates active background operations

All commands support --background and --wait flags. The plugin wraps the local Codex CLI binary — no separate runtime, same credentials, same usage limits.

Installation

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

Requirements: Node.js 18.18+, ChatGPT subscription or OpenAI API key.

Why This Matters Strategically

  1. Ecosystem penetration over platform lock-in. Claude Code accounts for ~4% of public GitHub commits (~135K/day) and ~$2.5B annualised revenue. OpenAI cannot ignore that install base — so they’re embedding Codex inside it rather than competing for switching.

  2. Cross-model adversarial review becomes first-class. The /codex:adversarial-review command formalises what power users already do manually: have one model critique another’s work. Claude Code authors code; Codex acts as sceptical reviewer. This is the strongest use case — one developer noted: “Claude more often finds big-picture or taste issues with Codex. And Codex more often finds correctness and code quality issues with Claude.”

  3. The “review gate” pattern. An optional feature blocks Claude Code from finalising changes until Codex has reviewed them. This is essentially automated cross-model CI — with cost implications (feedback loops between agents can spike usage).

  4. Signal for the market. If Anthropic or Google reciprocate with reverse plugins, competition shifts from raw model performance to integrated ecosystem maturity. The era of “pick one tool” may give way to “pick a primary agent, supplement with plugins.”

Relevance to Daniel’s Work

  • Agentic pod architecture: This plugin is a concrete implementation of the cross-model review pattern Daniel writes about. The adversarial review is exactly the “second opinion” layer in his pod design.
  • Dual-tool workflow validation: Daniel already uses Codex for throughput/precision and Claude Code for architectural reasoning. This plugin makes that workflow seamless.
  • Book chapter material: The strategic shift from platform lock-in to ecosystem penetration deserves a chapter section.

Friction Points

  • Dual authentication overhead for teams (Anthropic + OpenAI accounts)
  • Split usage limits across platforms
  • Background latency — explicit model selection recommended to avoid expensive auto-selection
  • Review gate feedback loops can cause cost spikes

Sources

Published: 2026-03-31

Updated:

You May Also Enjoy