@wayen_ai: 离谱的是——昨天 xAI 开源了 Grok Build,大家都在刷屏。但五天前,同一个 xai-org 账号悄悄发布了另一件东西:一个官方的 Claude Code 插件。 零公告。零推文。连作者都没提过。 7 个星标。1 个 fork。…
摘要
xAI 悄然发布了一个官方 Claude Code 插件,允许在 Claude Code 中直接调用 Grok Build CLI 进行代码审查、批评设计、委派任务和导入会话。该插件零公告发布,仅在 GitHub 上低调上线。
查看缓存全文
缓存时间: 2026/07/16 18:22
离谱的是——昨天 xAI 开源了 Grok Build,大家都在刷屏。但五天前,同一个 xai-org 账号悄悄发布了另一件东西:一个官方的 Claude Code 插件。
零公告。零推文。连作者都没提过。
7 个星标。1 个 fork。
这个插件做了这几个命令,让 Claude Code 可以直接调用 Grok Build CLI:
- 审查代码(/grok-build:review)
2.批评设计决策(/grok-build:critique)
3.委派调查任务(/grok-build:delegate)
4.导入 Claude 会话到 Grok(/grok-build:import)
它本质上是在竞争对手的终端上,为自己建了一座桥。
这不是“多模型策略”的宣传稿,是 xAI 的工程师直接写出来的代码。
建议直接收藏,在 xAI 注意到之前装好试试。
GitHub:
xai-org/grok-build-plugin-cc
Source: https://github.com/xai-org/grok-build-plugin-cc
Grok Build ↔ Claude Code Bridge
Bridge Grok Build into Claude Code for review, critique, delegation, and session import.
This repository is a Claude Code marketplace plugin that shells out to the real grok CLI. Run status, results, and stop are owned by the plugin (PID + log files). There is no app-server broker.
Requirements
- Node.js
>= 18.18 - Grok Build CLI (
grok) onPATH, or setGROK_BINARY - A logged-in Grok CLI session (
grok modelssucceeds)
Local install
From this repository root (path must be absolute):
# Resolve an absolute path to this clone, then add it as a local marketplace
claude plugin marketplace add "$(pwd)"
# example: claude plugin marketplace add /absolute/path/to/grok-build-plugin-cc
# Install the plugin from that marketplace
claude plugin install grok-build@xai-grok-build
Or, when Claude Code is already open, use /plugin and add the local marketplace path, then install grok-build@xai-grok-build.
Check readiness
/grok-build:check
Ready means: Node is available, grok is available, and soft auth via grok models succeeds.
Commands
/grok-build:check
Probe Node + Grok CLI availability and authentication.
/grok-build:review
Read-only review of local git state:
/grok-build:review --wait
/grok-build:review --background --scope working-tree
/grok-build:review --base main
/grok-build:review --wait --model grok-build --effort high
Runs:
grok -p <prompt> --agent explore --permission-mode plan --sandbox read-only --cwd <ws> --output-format plain
Optional: pass --model / --effort (low|medium|high). If omitted, Grok chooses defaults.
/grok-build:critique
Same target selection as review, with a design/risk critique prompt and structured JSON output when possible:
/grok-build:critique --wait
/grok-build:critique --base main challenge whether this was the right caching and retry design
/grok-build:critique --wait --model grok-build --effort high focus on failure modes
Optional: same --model / --effort flags as review.
/grok-build:delegate
Delegate investigation or implementation to Grok via the grok-build:grok-delegate subagent:
/grok-build:delegate investigate the flaky test in auth
/grok-build:delegate --resume apply the top fix
/grok-build:delegate --model grok-build --effort high fix the race
Write policy layering:
| Layer | Default |
|---|---|
Bridge run CLI | Read-only (--permission-mode plan + --sandbox read-only) unless you pass --write |
| Delegate agent / skill | Adds --write by policy (write-capable delegate) unless the user asks for read-only |
- Direct
node …/grok-bridge.mjs run "…"is therefore read-only unless--writeis passed. --resume/--resume-lastcontinues the last stored Grok session id viagrok -r <id>.- Prefer bridge
--backgroundfor long work so runs record bothbridgePid(Node worker) andagentPid(grok child). /grok-build:stopterminates both process trees when present (agent then bridge/worker).- If you do not pass
--modelor--effort, Grok chooses its own defaults.
/grok-build:import
Import the current Claude transcript into Grok:
/grok-build:import
/grok-build:import --source ~/.claude/projects/.../session.jsonl
Uses grok import and prints a resume hint: grok -r <id>.
/grok-build:runs
List active and recent plugin-owned runs:
/grok-build:runs
/grok-build:runs <run-id> --wait
/grok-build:show
Show stored output for a finished run:
/grok-build:show
/grok-build:show <run-id>
/grok-build:stop
Stop an active run by terminating tracked process trees:
/grok-build:stop
/grok-build:stop <run-id>
Kills every distinct pid among agentPid (detached grok child) and bridgePid / legacy companionPid / legacy pid (bridge or run-worker). Terminal status is claimed under a locked CAS so a finishing worker cannot overwrite cancelled with completed.
Environment
| Variable | Purpose |
|---|---|
GROK_BINARY | Optional override for the grok executable |
GROK_CC_SESSION_ID | Claude session id (set by SessionStart hook) |
GROK_CC_TRANSCRIPT_PATH | Claude transcript path (set by SessionStart hook) |
CLAUDE_PLUGIN_ROOT | Plugin install root (host) |
CLAUDE_PLUGIN_DATA | Plugin data root; state lives under .../state |
CLAUDE_ENV_FILE | Host env file for session hooks |
CLAUDE_PROJECT_DIR | Project directory from the host |
State fallback when CLAUDE_PLUGIN_DATA is unset: $TMPDIR/grok-cc-runs.
Development
npm test
Tests use Node’s built-in test runner and a fake grok binary on PATH. Runtime code uses Node stdlib only.
Version: 0.2.0.
License
Apache-2.0. See LICENSE and NOTICE.
相似文章
xai-org/grok-build, 现已开源
xAI 在其 Grok CLI 工具被发现将整个目录上传至其服务器后,遭到了强烈反对。作为回应,xAI 删除了保留的数据,禁用了默认数据保留功能,并将整个 Grok Build 代码库以 Apache 2.0 许可证开源,以重新赢得用户信任。
@vista8: 让 AI 分析拆解了 Grok Build 的源码,完整文档在这里 https://xiangyangqiaomu.feishu.cn/docx/OAoYdiZ1eoc2HTxmmZ3clWIPnLf…
用户利用AI(Codex)分析拆解了刚开源的Grok Build源码,并分享了完整分析文档。
@ai_super_niko: 1/5 用 Claude Code 写代码,切到 Codex 做 review,再切回来改。 每次切换都是一次思路谋杀。 你不是在写代码,你在疲于奔命。 直到 OpenAI 给竞争对手 Claude Code,出了一个官方插件:`code…
OpenAI 为竞争对手 Claude Code 推出了官方插件 codex-plugin-cc,用于在 Claude Code 中集成 Codex 进行代码审查,引发了关于工作流切换效率的讨论。
@yibie: 推荐 Simon Willison 对 Grok Build 开源代码的深度分析。不是因为代码本身——而是因为它让我们能窥见一个终端编程 agent 的内部运作:真实世界的系统 prompt、从 Codex 和 OpenCode "移植"…
Simon Willison 深度分析了 Grok Build 的开源代码,揭示了终端编程 agent 的内部运作,包括系统 prompt、工具实现和隐私问题。
@XAMTO_AI: OpenAI 创始成员将一份 CLAUDE.md 丢上 GitHub,一周内揽下 44k 星,总星数冲到 100k——连 Stack Overflow 都没这待遇。 他观察到 AI 编程最大的问题不是不会干活,而是“太爱发挥”:让它修个 …
OpenAI 创始成员发布 CLAUDE.md 文件,通过四条原则约束 AI 编程助手避免过度发挥和复杂化,一周内获得 44k GitHub stars。