grok
Allows Grok to review GitHub pull requests and branches, post pending review comments, and monitor open PRs.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@grokReview the current branch against main with focus on security issues."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
grok-plugin-codex
Chinese | English
A plugin that uses Grok directly in Codex.
You still talk to Codex. Reviews, hard bug fixes, proposals, design docs, PR babysitting, docs / images / videos — hand those to your local Grok Build.
Chinese search terms | Codex 调用 Grok · 在 Codex 上用 Grok · Codex Grok 插件 · OpenAI Codex 用 xAI Grok · Codex MCP Grok |
English search terms | Codex Grok plugin · use Grok in Codex · Grok MCP for Codex · Grok Build Codex · delegate Codex tasks to Grok |
Not an official xAI / OpenAI product. Current version 0.6.0. Apache-2.0. Derived from stdevMac/grok-in-codex.
What this is
Codex stays the foreman: you talk to it, it breaks down tasks and tracks progress.
This repo is just a very thin MCP plugin — it doesn't replace Codex, and it doesn't embed the Grok model into Codex. It calls the grok CLI you already have installed locally.
Good for: adding a second capable agent inside Codex; getting a structured review before merging; following a "proposal → design → implementation" flow; or using Grok's image / video / document capabilities (which Codex doesn't have).
Not needed for: one-line fixes, renames, or chit-chat that never touches the repo.
Related MCP server: Grok-in-Codex
What it can do
You want | Say to Codex | Tool |
Fix / troubleshoot | "Have Grok look into why the tests fail" |
|
Proposal only, no code changes | "Use Grok to plan an auth rewrite" |
|
Design doc + PR breakdown | "Have Grok write a design doc" |
|
Implement per design | "Execute the latest Grok plan, dry-run first" |
|
Review a branch / PR | "Have Grok review this branch against main" |
|
Challenge assumptions | "Do an adversarial review of the billing design" |
|
Babysit CI / review comments | "Have Grok watch the open PRs" |
|
Named workflows | "List / run a Grok workflow" |
|
PDF / Word / PPT | "Make a one-page release PDF" |
|
Images / short videos | "Use Grok to make a 16:9 release banner" |
|
Task board | "Check Grok task status" |
|
Don't cram bigger product work into one giant rescue:
Artifacts land in the project directory (gitignore recommended): .grok-plans/ .grok-designs/ .grok-reviews/ .grok-docs/ .grok-media/ .grok-workflows/.
Environment
Node.js 18.18+
Grok Build CLI (
grok) onPATH, commonly at~/.grok/bin/grokgrok loginGrok Build ≥ 1.0.5 recommended (0.2.118 still starts; specific flags are gated by capability detection)
ghonly needed if you want to post comments on PRs
This plugin doesn't ship a model and won't log you into Grok for you.
Installation
codex plugin marketplace add /path/to/grok-plugin-codex/.agents/plugins
codex plugin add grok@grok-plugin-codexOpen a new Codex session for the MCP tools to load. Then:
node plugins/grok/scripts/grok-companion.mjs setupOr have Codex call grok_setup.
Codex may launch the plugin from its install cache. Pass the real project path as cwd on every tool call so tasks and artifacts land in that repo:
grok_review cwd="/path/to/project" base=main
grok_status cwd="/path/to/project" json=trueGetting started
Just say it in Codex:
让 Grok 对照 main 评审这个分支。
用 Grok 规划一次鉴权重写。
让 Grok 出一份设计文档,然后 dry-run 执行最新方案。
给重试重构开一个后台 Grok rescue。
用 Grok 生成一张 16:9 发布 banner。
看看 Grok 任务状态。MCP example:
grok_plan prompt="plan the auth rewrite" background=true
grok_design prompt="design multi-tenant billing" background=true
grok_execute_plan latest=true dryRun=true
grok_workflow action=list
grok_review base=main focus="auth, data loss, and race conditions"
grok_rescue prompt="investigate why npm test is failing" background=true
grok_babysit action=list
grok_document type=pdf prompt="one-pager for the launch"
grok_image aspect="16:9" prompt="Dark developer-tool launch banner"
grok_video image="./.grok-media/image/hero.png" duration="6" prompt="gentle camera push-in"
grok_status
grok_result jobId="plan-abc123"Tool overview
Tool | Purpose |
| CLI + login + version + doctor; optional stop-review gate |
| Troubleshoot / implement (writable by default) |
| Proposal only → |
| Read-only review (tree / branch / PR; optional |
| Challenge designs and assumptions |
| List / run Grok Rhai workflows |
| Design doc + PR plan → |
| Execute the PR DAG from a design doc |
| Babysit PRs / fix CI and review comments ( |
| docx / pdf / pptx → |
| Generate or edit images → |
| Short videos → |
| List / search / export Grok sessions |
| Context handoff notes for Grok |
| Tasks + live progress / log tail + usage |
| Final output ( |
| Cancel background tasks |
Control parameters for long tasks: sandbox (workspace, read-only, strict, devbox, off; the old alias workspace-write normalizes to workspace), planMode / permissionMode, agent, noSubagents, memory / noMemory, allow / deny, disableWebSearch, forkSession, maxTurns.
Aligned with Grok CLI 1.0.x: if the CLI has no --check / --best-of-n, they aren't force-passed. check=true is written into the prompt. Write tasks that only read for many turns without writing files are stopped on execution drift.
Usage notes
Rescue — writable by default. readOnly=true for troubleshooting only. worktree=true for safer changes.
Plan / design / execute — artifacts go to .grok-plans/ and .grok-designs/. latest=true picks the newest design. dryRun=true is read-only.
Review — never applies patches. With postPending=true on a PR, it only posts PENDING GitHub comments when it finds something.
Media — files are copied into .grok-media/ to stay within the project path contract.
Jobs — background=true returns a job id. Multiple Grok tasks can run at once. Track them with grok_status / grok_result / grok_cancel.
CLI posture — uses a denylist (--disallowed-tools) rather than an allowlist. Media, dry-run / validate-only, and babysit list are read-only (no yolo). Reviews default to --sandbox read-only and don't strip shell tools (Grok 1.0.x needs them to produce task results).
Environment variables
Variable | Purpose |
| Overrides the |
| Root directory for this plugin's task state |
| Host plugin data directory; only trusted when the basename is |
Default state directory: ~/.grok/codex-plugin/state/. Not shared with the Claude plugin state.
Task control
No global single-task lock. Prefer
background=truefor long tasks.Status follows the text stream and thought stream; still shows
runningon whitespace-only tokens. Status also records the most recent tool and whether a write tool has run.Plan results prefer harvesting
plan.md. Completed tasks storeconfig,usage,artifacts(v3).Reaper: pid dead and
result.jsoncomplete → completed; pid dead but result truncated → failed (never stuck on running forever).Background
result.jsonis written atomically (tmp + rename).PR post-pending also runs when background jobs finish; empty / oversized diffs fail closed, findings go to
.grok-reviews/.
Development
npm test
node plugins/grok/scripts/grok-companion.mjs setup --json
node plugins/grok/mcp/server.mjs # stdio NDJSON MCP serverVersion numbers in package.json, plugins/grok/.codex-plugin/plugin.json, and .agents/plugins/marketplace.json must match.
License
Apache-2.0. See LICENSE and NOTICE.
This project is derived from stdevMac's stdevMac/grok-in-codex.
This server cannot be deployed
Maintenance
Related MCP Connectors
Source-checked CLI guides and model-aware planning for Claude Code, Codex, and Grok Build.
A paid remote MCP for OpenAI Codex agent coordination MCP, built to return verdicts, receipts, usage
No-data MCP handoff for local Claude Code to Codex harness moves. $49 lifetime.
- QuallaaOAuthcom.quallaa
Talk to your public-facing AI from any MCP client — Claude, ChatGPT, Cursor, Cline, Windsurf.
Related MCP Servers
- FlicenseAqualityBmaintenanceEnables using the xAI Grok CLI as an MCP sub-agent for code review, asking questions, and continuing conversations within MCP hosts like Claude Code.4-
- AlicenseNot gradedqualityDmaintenanceEnables Codex to delegate code reviews, coding tasks, image/video generation, and background investigations to Grok via MCP tools.25Apache 2.0
- AlicenseNot gradedqualityAmaintenanceEnables Codex to use Grok Build CLI as a controlled subagent via MCP tools for independent investigation, review, and isolated implementation tasks.6MIT
- AlicenseNot gradedqualityBmaintenanceConnects OpenAI Codex and Claude Code to the local Grok Build CLI for second-opinion code reviews, bounded coding tasks, web research, media generation, and TTS.2Apache 2.0