Grok × Codex Bridge
A local, read-only MCP bridge that lets OpenAI Codex ask your authenticated Grok Build CLI for a second opinion without copying API keys into Codex.
README
<p align="center"> <img src="./public/og.png" width="100%" alt="Grok Codex Bridge — two agents, one sharper answer" /> </p>
<h1 align="center">Grok × Codex Bridge</h1>
<p align="center"> <strong>One local bridge. Two independent agent perspectives.</strong><br /> Use your authenticated Grok Build CLI as a safe, read-only second-opinion tool inside OpenAI Codex. </p>
<p align="center"> <a href="./README.md"><b>English</b></a> · <a href="./README.zh-CN.md"><b>简体中文</b></a> · <a href="https://grok-codex-bridge.xyropy.chatgpt.site"><b>Interactive Demo</b></a> · <a href="./docs/ARCHITECTURE.md"><b>Architecture</b></a> · <a href="./docs/SECURITY.md"><b>Security</b></a> </p>
<p align="center"> <a href="https://github.com/wangsiyi7/grok-codex-bridge/stargazers"><img src="https://img.shields.io/github/stars/wangsiyi7/grok-codex-bridge?style=flat-square&color=A2FF59&labelColor=10131D" alt="GitHub stars" /></a> <a href="./LICENSE"><img src="https://img.shields.io/badge/license-MIT-A2FF59?style=flat-square&labelColor=10131D" alt="MIT license" /></a> <img src="https://img.shields.io/badge/Windows-PowerShell-52E8FF?style=flat-square&labelColor=10131D" alt="Windows" /> <img src="https://img.shields.io/badge/macOS-bash-8A72FF?style=flat-square&labelColor=10131D" alt="macOS" /> <img src="https://img.shields.io/badge/protocol-MCP-F3F5F7?style=flat-square&labelColor=10131D" alt="MCP" /> </p>
[!NOTE] Unofficial community project. It is not affiliated with, endorsed by, or sponsored by xAI or OpenAI.
What this project does
Codex remains the primary coding agent. The bridge gives it four explicit MCP tools that invoke your locally authenticated Grok Build CLI in headless, advisory-only mode. Grok returns text; Codex verifies the result and owns the final answer.
Your task in Codex
│
▼
Grok second-opinion skill
│ explicit, bounded context
▼
Local MCP bridge ─────► Official Grok Build login
│ │
└──── advisory text ◄───┘
│
▼
Codex verifies, reconciles, and delivers
<table> <tr> <td width="50%" valign="top"> <h3>Independent perspective</h3> <p>Ask a second model to challenge an architecture, plan, review, or draft answer before Codex commits to a recommendation.</p> </td> <td width="50%" valign="top"> <h3>Local trust boundary</h3> <p>No hosted relay and no copied API key. Grok Build owns its official browser login and refresh lifecycle.</p> </td> </tr> <tr> <td width="50%" valign="top"> <h3>Read-only by design</h3> <p>No automatic approvals, repository mutation, background file access, Grok memory, or hidden subagents.</p> </td> <td width="50%" valign="top"> <h3>Cross-platform setup</h3> <p>One-command installers for Windows PowerShell and macOS bash, plus a self-contained Codex plugin marketplace.</p> </td> </tr> </table>
Start in one command
<details open> <summary><b>Windows PowerShell</b></summary>
irm https://raw.githubusercontent.com/wangsiyi7/grok-codex-bridge/main/install.ps1 | iex
</details>
<details> <summary><b>macOS</b></summary>
curl -fsSL https://raw.githubusercontent.com/wangsiyi7/grok-codex-bridge/main/install.sh | bash
</details>
The installer checks Codex, installs Grok Build through xAI's official installer when needed, starts the official browser login, registers the plugin marketplace, and runs a local doctor check.
Interactive demo
The companion experience is a full product walkthrough, not a static documentation page. It includes:
- bilingual English / 简体中文 switching;
- an editorial, scroll-driven explanation of the Codex → local MCP → Grok Build flow;
- an interactive browser-style demo that simulates a complete advisory handoff;
- floating depth, a focused security boundary, and responsive typography;
- platform-aware quickstart commands and copy feedback;
- complete documentation, roadmap, FAQ, and contributor entry points.
MCP tools
| Tool | Best for | Input boundary | Output |
|---|---|---|---|
ask_grok |
Focused second opinions | Question + optional context | Concise advisory response |
plan_with_grok |
Independent implementation planning | Objective + constraints | Assumptions, risks, sequence, validation |
review_with_grok |
Diff, snippet, or document review | Supplied material only | Findings prioritized by severity |
compare_with_grok |
Cross-model answer critique | Question + Codex draft | Agreements, disagreements, synthesis |
Try these prompts in a new Codex task:
Ask Grok for a second opinion on this architecture.
Have Grok review this diff for hidden security regressions.
Compare your answer with Grok and synthesize the strongest result.
Architecture
flowchart LR
U["User task"] --> C["Codex"]
C --> S["grok-second-opinion skill"]
S --> M["Local MCP server"]
M --> G["Authenticated Grok Build CLI"]
G --> A["Advisory text"]
A --> V["Codex verification + synthesis"]
V --> U
classDef primary fill:#10131d,stroke:#52e8ff,color:#f3f5f7
classDef bridge fill:#10131d,stroke:#a2ff59,color:#f3f5f7
classDef second fill:#10131d,stroke:#8a72ff,color:#f3f5f7
class C,V primary
class S,M bridge
class G,A second
The adapter uses Grok Build's documented headless JSON mode rather than pretending ACP and MCP are the same protocol. See the full architecture notes.
Safety contract
| The bridge does | The bridge deliberately does not |
|---|---|
| Pass explicit text selected by Codex | Read the contents of ~/.grok/auth.json |
| Separate Grok stdout from diagnostics | Export passwords, browser cookies, or tokens |
| Cap input, output, and runtime | Pass --always-approve to Grok |
| Return advisory text to Codex | Edit files or run repository commands through Grok |
| Tell the user when Grok is consulted | Treat Grok output as unquestioned authority |
Read the complete security and privacy model.
Project map
grok-codex-bridge/
├── plugins/grok-codex-bridge/
│ ├── .codex-plugin/plugin.json # Codex plugin manifest
│ ├── .mcp.json # Local MCP registration
│ ├── scripts/mcp-server.mjs # Dependency-free bridge
│ └── skills/grok-second-opinion/ # Codex routing workflow
├── app/ # Interactive bilingual demo
├── docs/ # Architecture, security, install guides
├── install.ps1 # Windows one-command installer
├── install.sh # macOS one-command installer
└── tests/ # MCP and rendered-page regression tests
Verification
npm ci
npm run lint
npm run test:bridge
npm test
Validated locally on Windows with:
- a real Premium+ Grok Build browser login;
- a direct Grok headless response (
GROK_BRIDGE_OK); - a complete MCP request/response (
MCP_GROK_OK); - a clean public-repository installer replay;
- plugin manifest, protocol, lint, production build, and server-render tests.
Documentation by goal
| Goal | Document |
|---|---|
| Install, verify, troubleshoot, uninstall | Installation guide |
| Understand the process and protocol boundary | Architecture |
| Review data flow, credentials, and guardrails | Security |
| Contribute safely | Contributing guide |
| Read in Chinese | 简体中文说明 |
Roadmap
- [x] Official Grok Build login reuse
- [x] Four read-only MCP advisory tools
- [x] Windows and macOS one-command installers
- [x] Personal and repository Codex marketplace support
- [x] Bilingual interactive product demo
- [ ] Signed release artifacts
- [ ] Public compatibility matrix across Codex releases
- [ ] Opt-in structured citations for advisory responses
- [ ] Community-maintained prompt recipes
Contributing
Security-preserving improvements are welcome. Start with CONTRIBUTING.md, open an issue for trust-boundary changes, and keep prompt or credential logging out of pull requests and bug reports.
Acknowledgements
This project is inspired by the extensibility models documented by xAI Grok Build and OpenAI Codex. Product and company names belong to their respective owners.
License
MIT © 2026 wangsiyi7
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。
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。
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。
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。