Parallel Codex workers inside Claude Code — multi-agent orchestration.
How to install magic-cc-codex-worker?
wenqingyu/magic-cc-codex-worker
$git clone https://github.com/wenqingyu/magic-cc-codex-workerInstalls into the current project.
›Prefer a prompt? Paste this to your agent
Install magic-cc-codex-worker by running `git clone https://github.com/wenqingyu/magic-cc-codex-worker`, then use it for the current task and follow its documentation at https://github.com/wenqingyu/magic-cc-codex-worker.
README.md
| 1 | # magic-cc-codex-worker |
| 2 | |
| 3 | ### Parallel OpenAI Codex workers inside Claude Code. |
| 4 | |
| 5 | [](https://github.com/wenqingyu/magic-cc-codex-worker/actions/workflows/ci.yml) |
| 6 | [](LICENSE) |
| 7 | [](.nvmrc) |
| 8 | [](tsconfig.json) |
| 9 | |
| 10 | **Languages:** English · [简体中文](README.cn.md) |
| 11 | |
| 12 | > **Turn Claude Code into a multi-agent coding system powered by OpenAI Codex.** A zero-install Claude Code plugin for multi-agent orchestration — delegate implementation, code reviews, and planning to a pool of parallel Codex workers, each running in its own isolated git worktree, via the MCP (Model Context Protocol) standard. |
| 13 | |
| 14 | The bridge between Claude Code and the OpenAI Codex ecosystem: Claude stays in orchestrator mode (planning, synthesis, interactive work), Codex workers absorb the grunt work. Get dual-model PR review, resumable sessions across conversations, and role-based agent specialization — while preserving your Claude quota for what Claude does best. |
| 15 | |
| 16 | ## Why use this? |
| 17 | |
| 18 | - ⚡ **Parallel execution.** Fan out N Codex workers on independent subtasks in N isolated worktrees. Finish work that would serialize in a single Claude session. |
| 19 | - 🛡️ **Isolated experimentation.** Every implementer runs in its own `git worktree` on its own branch. Try three approaches in parallel; keep the best; discard the rest. Zero risk to your main tree. |
| 20 | - 🔀 **Two model families beat one.** Launch a Codex (GPT) reviewer alongside your Claude review — different models catch different classes of bugs. The plugin materializes PRs in detached worktrees so the reviewer reads real files, not a diff blob. |
| 21 | - 💰 **Quota arbitrage.** Claude budget running low? Dial delegation up to `max` and route everything Codex can handle over there — Claude stays in orchestrator mode. One knob (`minimal` / `balance` / `max`) controls the split. |
| 22 | - 🎯 **Role-tuned, observable delegation.** Not a thin "forward the prompt" wrapper — a full orchestration layer with role-based specialization (implementer / reviewer / planner / generic), resumable sessions, per-role sandbox + timeout, and first-class session tracking in a persisted registry. |
| 23 | - 🧰 **Production engineering.** 62 unit tests, strict TypeScript, CI on Node 20/22. Designed from an actual spike of Codex's MCP protocol — no stdout parsing, no brittle scraping. Git worktrees for parallelism, MCP protocol for transport, TOML for configuration, sandboxed execution for safety. |
| 24 | |
| 25 | ## How it compares to the official Codex plugin |
| 26 | |
| 27 | | | Official OpenAI Codex plugin | **magic-cc-codex-worker** | |
| 28 | |---------------------------------------|:---------------------:|:-------------------------:| |
| 29 | | Single Codex session in Claude Code | ✅ | ✅ | |
| 30 | | Multi-agent orchestration | ❌ | ✅ | |
| 31 | | Parallel worker execution | ❌ | ✅ | |
| 32 | | Git worktree isolation per worker | ❌ | ✅ | |
| 33 | | Role-based specialization | ❌ | ✅ | |
| 34 | | Resumable session continuity | ❌ | ✅ | |
| 35 | | Dual-model PR review | ❌ | ✅ | |
| 36 | | Epic / batch fan-out | ❌ | ✅ | |
| 37 | |
| 38 | OpenAI's official Codex plugin lets you **use** Codex. This plugin lets you **scale** Codex into a multi-agent coding system inside Claude Code. |
| 39 | |
| 40 | --- |
| 41 | |
| 42 | ## Quick start |
| 43 | |
| 44 | ### Install (two slash commands) |
| 45 | |
| 46 | Claude Code plugin distribution is a two-step pattern — same as `brew tap` + `brew install` or `apt-add-repository` + `apt install`. First you register a **marketplace** (a catalog that lists plugins), then you **install** one of its plugins. Our marketplace happens to contain only this one plugin, so the names look duplicated — that's normal. |
| 47 | |
| 48 | #### Step 1 — Register the marketplace catalog |
| 49 | |
| 50 | Tells Claude Code "this GitHub repo publishes a plugin catalog." It clones the repo's `.claude-plugin/marketplace.json` and lists the plugins available from it. No plugin is installed yet. |
| 51 | |
| 52 | ```text |
| 53 | /plugin marketplace add wenqingyu/magic-cc-codex-worker |
| 54 | ``` |
| 55 | |
| 56 | #### Step 2 — Install the plugin from that marketplace |
| 57 | |
| 58 | Picks one plugin out of the catalog and attaches it to your Claude Code session. The `<plugin-name>@<marketplace-name>` format disambiguates when a plugin name exists in multiple catalogs. |
| 59 | |
| 60 | ```text |
| 61 | /plugin install magic-codex@magic-codex |
| 62 | ``` |
| 63 | |
| 64 | #### Step 3 — Reload plugins |
| 65 | |
| 66 | Activates the newly-installed plugin in the current session without a full restart. |
| 67 | |
| 68 | ```text |
| 69 | /reload-plugins |
| 70 | ``` |
| 71 | |
| 72 | After |
Related
6 picksType
- langchain-mcp-adaptersMake Anthropic Model Context Protocol (MCP) tools compatible with LangChain and LangGraph agents.Jul 2026
- mcporterTypeScript runtime and CLI for connecting to configured Model Context Protocol servers.Jul 2026
- mcp-proxyA MCP server which proxies requests to a remote MCP server over streamable HTTP or SSE.Jul 2026
- mcp-atlassianThe Model Context Protocol (MCP) Atlassian integration is an open-source implementation that bridges Atlassian products (Jira and Confluence) with AI language models following Anthropic's MCP specification.Jul 2026
- open-webuiOpen WebUIJul 2026
- rulesyncUnified AI rules management CLI tool that generates configuration files for various AI development toolsJul 2026
