Grok Plugin Codex
An MCP server that exposes the local Grok CLI to Codex for bounded repo work, reviews, rescue analysis, adversarial checks, session listing/export, and background job management while keeping Codex hidden context out of scope.
README
Grok Plugin Codex
grok-plugin-codex is a Codex plugin that exposes the local Grok CLI through a bundled Node/TypeScript stdio MCP server.
It lets Codex ask Grok for bounded repo work, reviews, rescue analysis, adversarial failure-mode checks, session listing/export, and background job management while keeping Codex hidden runtime context out of scope.
Repository: https://github.com/handong66/grok-plugin-codex
Tools
grok_check: discover Grok CLI, rungrok --version, and check login/model availability withgrok models.ok: truemeans both CLI discovery andgrok modelssucceeded.grok_models: return raw and parsedgrok modelsoutput.grok_run: rungrok --cwd <cwd> -p <prompt> --output-format jsonin foreground orstreaming-jsonin background.grok_continue: continue with--resume <sessionId>or with--continueonly whencontinueLatest: true.grok_rescue: read-only independent diagnosis prompt wrapper.grok_review: bounded findings-first review prompt wrapper.grok_adversarial_review: bounded failure-mode review prompt wrapper that asks Grok for at most 5 findings. The cap is prompt-level guidance, not output truncation.grok_sessions: wrapgrok sessions list/search.grok_export: wrapgrok export <sessionId>and return Markdown from stdout by default.grok_status,grok_result,grok_cancel: manage background Grok jobs.
There is no grok_transfer tool in v1. Grok has import, but this repository does not claim a verified Codex rollout JSONL to Grok import schema.
Requirements
- Node.js
>=22 - npm
- Codex local plugin marketplace support
- Grok CLI installed and logged in
Check Grok directly:
grok --version
grok models
Install
From the repository root:
npm install
npm run check
codex plugin marketplace add .
codex plugin add grok-plugin-codex --marketplace grok-plugin-codex
Then start a new Codex thread so the MCP tools and grok skill are loaded.
npm run check builds plugins/grok-plugin-codex/dist/server.js; the installed MCP server runs that bundle from the plugin directory.
Usage Notes
Use the Grok tools as a second-agent surface. Codex remains responsible for scope, workspace state, verification, git, and final judgment.
Pass cwd whenever Grok should inspect a specific workspace. If cwd is omitted, the plugin falls back to the MCP server process directory, which is normally the installed plugin directory rather than the user's active repo.
For long tasks, pass background: true to use --output-format streaming-json, then poll:
grok_status -> grok_result -> grok_cancel if needed
Use the same cwd when polling or cancelling a background job that was used when the job was started; job records live under <cwd>/.grok-plugin-codex/jobs.
grok_result treats a background job as complete only when the process succeeded, at least one text event was observed, and an end event was observed. Partial logs are process evidence, not a finished review.
Use very low maxTurns values only for sentinel checks or prompts that do not need file/tool work. Repo reviews and rescue analysis usually need enough turns for Grok to inspect the requested files and produce a final answer.
Shared Arguments
Run, continue, rescue, review, and adversarial-review tools accept:
cwdgrokBinmodeltimeoutMsbackgrounddisableWebSearchnoSubagentsmaxTurnsalwaysApprovereasoningEffortallowCodexPrivatePaths
grok_check accepts cwd, grokBin, timeoutMs, and optional includeModels. grok_models accepts cwd, grokBin, and timeoutMs.
grok_sessions accepts cwd, grokBin, timeoutMs, query, and limit. Search queries are separated from Grok CLI flags before execution. grok_export accepts cwd, grokBin, timeoutMs, sessionId, and optional outputFile; outputFile must resolve inside cwd. Discovery-style subcommands pass cwd as Grok's global --cwd option. Job tools accept cwd plus a jobId in the generated job_<timestamp>_<8-hex> format; grok_result also accepts maxChars.
grok_check also accepts includeModels: false to skip grok models and verify only CLI discovery/version. With the default behavior, grok_check runs grok models and ok: true means both discovery and model probing succeeded.
alwaysApprove defaults to false and is only passed when explicitly true.
reasoningEffort is passed through only when the plugin does not know it is unsupported. The local default model grok-composer-2.5-fast does not support --reasoning-effort; the plugin warns and does not pass that flag for that model. If no model is specified, the plugin also warns and does not pass --reasoning-effort, because the local default may be grok-composer-2.5-fast.
Privacy Boundary
This plugin does not copy Codex hidden context, system/developer messages, tool outputs, hidden reasoning, secrets, or Grok auth tokens into prompts.
Grok CLI child processes receive only the plugin-declared environment allowlist: GROK_BIN, HOME, and PATH.
Prompts that ask Grok to read Codex private runtime paths such as ~/.codex are rejected by default. Set allowCodexPrivatePaths: true only when the user explicitly asks for that risk and understands it.
This boundary does not redact arbitrary user-provided text. If a caller pastes secrets, private tool output, or sensitive file contents into prompt, problem, or target, that text is passed to Grok.
Development
npm install
npm run check
git diff --check
Optional authenticated live smoke:
npm run smoke:live-grok
Keep these files aligned whenever tools change:
README.mdplugins/grok-plugin-codex/README.mdplugins/grok-plugin-codex/skills/grok/SKILL.mdplugins/grok-plugin-codex/src/tools.tsplugins/grok-plugin-codex/src/server.tsscripts/smoke-mcp.mjs
See docs/development.md and docs/verification.md.
Project Policies
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。