Codex Peer
Lets one Codex instance send a task to another Codex instance on a different computer, enabling symmetric host-to-host collaboration including projectless tasks and local environment dependent work.
README
Codex Peer
Codex Peer lets one Codex instance send a task to another Codex instance running on a different computer, then follow the resulting thread until it finishes.
It is aimed at people who actively use Codex on two or more machines and need symmetric host-to-host collaboration, including projectless tasks and work that depends on the peer computer's local environment.
[!WARNING] Codex Peer is a Developer Preview. It uses the experimental Codex app-server WebSocket API, which may change without backward compatibility.
When to use it
Use Codex's native remote features first when they cover the workflow. Codex Peer is useful when you need a separately running Codex on another computer to:
- start or continue a projectless task;
- inspect files, applications, or state that exist only on that computer;
- exchange natural-language task reports with the calling Codex;
- work in either direction, such as Mac to Windows and Windows to Mac.
Codex Peer does not provide remote desktop control. The peer Codex can only do what its own tools, permissions, and local environment allow.
Install
Node.js 22 or later and the Codex CLI are required.
codex plugin marketplace add shotaro311/codex-peer
codex plugin add codex-peer@codex-peer
The repository contains a bundled MCP server, so plugin users do not need to run npm install.
Continue with Setup. Keep the app-server on loopback unless you have a TLS-protected and authenticated network route.
Tools
peer_health: verifies the connection and app-server initialization.peer_message: starts or continues a peer task.peer_wait: checks one existing peer turn.peer_wait_until_complete: quietly follows a turn until it ends or the local wait window expires.peer_threads: lists threads on the peer.peer_read: reads recent turns from a peer thread.
A running turn returns ok: true and turnCompleted: false. A failed or interrupted turn returns ok: false, its turnStatus, and a sanitized turnError; an empty response is never treated as proof of success.
Privacy defaults
- Remote
wss://peers require a token. - Plain
ws://is accepted only for loopback addresses. - Credentials, query parameters, and fragments are rejected in peer URLs.
- Transcript recording is disabled by default.
- When explicitly enabled, transcripts contain event metadata only, not prompts or responses.
Read Security before exposing any route beyond localhost.
Development
npm ci
npm run check
npm audit --omit=dev
npm run check rebuilds dist/codex-peer-mcp.mjs, runs the test suite, and validates the plugin layout.
Status and scope
The initial release is intentionally small: remote app-server messaging, thread discovery, readback, and long-turn tracking. It does not include scheduled jobs, remote desktop, local LLM summarization, or a relay service.
See Troubleshooting, Contributing, the MIT License, and third-party notices.
Official references
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。