OpenAI Codex plugin adds Claude Code delegation through the app server

• OpenAI shipped a Codex plugin for Claude Code that lets developers delegate tasks and run Codex reviews without leaving Claude Code, according to the launch post and the README screenshot.
• The integration is built on the Codex app server rather than an MCP server, so it uses the same "open-source agent harness" as Codex itself and can run "multiple Codex tasks in parallel in the background," per the thread and a practitioner note.
• The first release centers on review and delegation commands: the README lists /codex:review, /codex:adversarial-review, plus /codex:rescue, /codex:status, /codex:result, and /codex:cancel for background job control.
• OpenAI positioned the plugin as subscription-friendly and cross-workflow: the launch post says it can use a ChatGPT subscription, while a follow-up reply says the goal is to make Codex usable "everywhere," including Claude Code alongside OpenCode, Xcode, and JetBrains tools.

The new plugin brings Codex directly into Claude Code as a delegated worker and reviewer. In the announcement, Daniel Kundel says users can "trigger Codex from Claude Code" to "delegate tasks to Codex" or "have Codex review your changes," and the linked GitHub repo positions it for teams that want Codex inside an existing Claude Code workflow.

The concrete surface area is broader than a single review command. As shown in the README screenshot, the plugin exposes a normal read-only /codex:review, a steerable /codex:adversarial-review, and a small job-control set — /codex:rescue, /codex:status, /codex:result, and /codex:cancel — for dispatching and managing background work. The repo summary in the linked overview also notes setup requirements: Node.js 18.18+ plus either a ChatGPT subscription or an OpenAI API key.

The key implementation detail is that this plugin rides on the Codex app server, not a separate MCP integration path. In the architecture thread, Kundel says it is "built on top of the Codex app server" and therefore uses "the same open-source agent harness as the rest of Codex," which is what enables parallel background tasks, shared models, and the same Codex review flow inside Claude Code.

That app-server choice stood out to practitioners because it keeps Claude Code as the front end while Codex handles orchestration behind the scenes. Kevin Kern writes in his post that the plugin "doesn't use the codex mcp server" and instead wraps the app server; the OCR text attached to the thread image describes a "forwarder-not-orchestrator pattern," "background/foreground modes," and scope controls like auto|working-tree|branch. Early usage ideas are already appearing around those hooks: one practitioner suggests wiring stop hooks to automatically ask for a Codex review, and Kundel's reply frames Claude Code as one of several IDE surfaces where the same Codex backend should work.