Portable agent skill: call Grok Build / Super Grok 4.5 from any coding agent that can run a shell.
Works with Claude Code, Codex, Cursor, Copilot, OpenClaw, Hermes Agent, and custom multi-agent stacks. One SKILL.md — install, login, health-check, then shell out for code review, adversarial critique, research, second opinions, images, video, and optional implementation.
OpenClaw · Hermes · Claude Code · Codex · Cursor · …
│ shell (portable skill)
▼
grok -p "…" ← Grok Build CLI (local Super Grok)
│
▼
stdout → return as-is (reviews) or asset paths (media)
| Install time | ~1 minute if Grok is already on PATH |
| Platforms | Windows · macOS · Linux |
| Surface | Single portable SKILL.md — no Node broker required |
| Default safety | Review / research = read-only |
| Convention | .agents/skills/ multi-harness skill path |
Keywords people search: invoke grok · grok skill · use grok from claude code · grok second opinion · multi-agent grok · openclaw skill · hermes skill · agent skills · grok build headless · cross-agent review
- What is Invoke Grok?
- Why it exists
- Works with your agent stack
- Quick start
- Install the skill
- Health check
- What you can do
- Choose a model
- Multi-agent patterns
- How it compares
- FAQ
- Documentation
- Contributing
- License
Invoke Grok is a portable agent skill (SKILL.md) that teaches coding agents and multi-agent harnesses how to call the local Grok Build CLI (grok) — including Super Grok / Grok 4.5 workflows for review, research, and media.
It is not a full marketplace plugin with background job queues. It is the smallest useful product in the agent skills ecosystem:
- Install + login instructions for every OS
- A smoke test so the host knows Grok is alive
- Flag recipes for review, critique, research, ask, image, video, implement
- Host rules: read-only by default, return review text verbatim
- Drop-in under
.agents/skills/(shared skill layer across harnesses)
Canonical skill file: SKILL.md
2026 agent reality: people run OpenClaw, Hermes Agent, Claude Code, Codex, and Cursor side by side. Plans and skills get shared; reviews still collapse into one model’s bias.
Coding agents get boxed: limited tools, one model’s blind spots, usage caps, no web, weak media.
You already pay for (or can install) Grok Build. This skill is the escape hatch:
Stay in your main agent — when stuck, invoke Grok.
Same idea as multi-agent bridges (Codex-from-Claude, Hermes↔Claude Code delegation), but deliberately one markdown skill so any host with shell access can use it — including OpenClaw and Hermes skill directories.
| Host / harness | How Invoke Grok fits |
|---|---|
| Claude Code | Copy skill to ~/.agents/skills/ or ~/.claude/skills/; ask for Grok review |
| OpenAI Codex | Load skill + shell grok -p; second pass after Codex implements |
| Cursor | Project or user skills path; @ / agent rules can point at AGENTS.md |
| OpenClaw | Drop into workspace/user skills; Grok as external specialist (review / research) |
| Hermes Agent | Portable skill alongside Hermes’ skill library; shell out when you want Super Grok |
| Copilot / custom | Any agent that runs shell + reads SKILL.md |
| Multi-agent orchestrators | One worker plans, one builds, Grok reviews — classic plan/build/review loop |
Gap this fills: OpenClaw + Hermes dominate “run agents / share skills” talk. Invoke Grok is the missing cross-harness Super Grok peer — not another full agent runtime.
macOS / Linux / Git Bash:
curl -fsSL https://x.ai/cli/install.sh | bash
grok --versionWindows (PowerShell):
irm https://x.ai/cli/install.ps1 | iex
grok --versiongrok login
# headless / SSH:
grok login --device-authOr set XAI_API_KEY for CI.
User-global (recommended):
# Unix
mkdir -p ~/.agents/skills/invoke-grok
curl -fsSL https://raw.githubusercontent.com/aalhadxx/invoke-grok/main/SKILL.md \
-o ~/.agents/skills/invoke-grok/SKILL.md# Windows PowerShell
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.agents\skills\invoke-grok" | Out-Null
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/aalhadxx/invoke-grok/main/SKILL.md" `
-OutFile "$env:USERPROFILE\.agents\skills\invoke-grok\SKILL.md"From a clone:
git clone https://github.com/aalhadxx/invoke-grok.git
cp invoke-grok/SKILL.md ~/.agents/skills/invoke-grok/SKILL.md
# Windows: copy to %USERPROFILE%\.agents\skills\invoke-grok\SKILL.mdProject-local (team share):
mkdir -p .agents/skills/invoke-grok
cp path/to/invoke-grok/SKILL.md .agents/skills/invoke-grok/Optional Claude Code path: also copy to ~/.claude/skills/invoke-grok/SKILL.md.
Use the invoke-grok skill. Health-check Grok, then review my uncommitted changes.
Or natural language: “Ask Grok for a second opinion on this auth change.”
| Location | Who loads it |
|---|---|
~/.agents/skills/invoke-grok/SKILL.md |
Multi-agent convention; Grok scans this too |
~/.claude/skills/invoke-grok/SKILL.md |
Claude Code |
~/.cursor/skills/invoke-grok/SKILL.md |
Cursor (when skills enabled) |
.agents/skills/invoke-grok/SKILL.md |
This repo / any project |
See docs/install.md for OS notes and GROK_BINARY.
Agents should run this before the first real handoff:
grok --version
grok models
grok -p "hi. reply with exactly: pong"Success: stdout contains pong.
Failure: fix install/login — do not invent answers.
| Mode | Purpose | Default safety |
|---|---|---|
| review | Independent code review | Read-only (plan + sandbox read-only + explore agent) |
| critique | Adversarial design / risk | Read-only |
| research | Web + repo synthesis | Read-only; web enabled |
| ask | Second opinion / unbox stuck host | Read-only |
| image | Generate or edit images | Save path reported by Grok |
| video | Short animate / multi-ref clips | Save path reported by Grok |
| implement | Fix / code (only if user asks) | --yolo / write-capable |
Example — read-only review:
grok -p "Review uncommitted changes for bugs and risks. Be specific with file paths." \
--agent explore \
--permission-mode plan \
--sandbox read-only \
--cwd .Full recipes: SKILL.md · docs/modes.md
grok models # list what your account supports
grok -p "…" -m <model-id> # pin a model
grok -p "…" -m <model-id> --effort highPolicy for host agents: omit -m unless the user names a model; omit --effort unless they ask for thoroughness.
Trending stacks (OpenClaw, Hermes, Claude Code, Codex) already do plan / build / monitor. Add Grok as the independent review and research lane:
| Pattern | Host does | Grok does (invoke-grok) |
|---|---|---|
| Ship gate | Implements feature | Read-only review / critique before merge |
| Unbox | Stuck in a loop | ask second opinion with fresh tools/model |
| Research | Coding only | research with web + repo context |
| Media | Text-only agent | image / video generation |
| Cost split | Cheap model builds | Grok only on judgment passes |
Classic loop:
1. Main agent (Claude / Hermes / Codex / OpenClaw) implements
2. invoke-grok → review (read-only)
3. Main agent fixes top findings
4. Optional: invoke-grok → critique before release
| invoke-grok | Full agent runtimes (OpenClaw, Hermes, …) | Claude ↔ Codex plugins | Grok↔Claude bridges | |
|---|---|---|---|---|
| Role | Portable skill to call Grok | Full agent product / harness | Deep Claude integration | Deep Claude ↔ Grok |
| Install | Copy one SKILL.md |
Runtime install | Marketplace + Node | Marketplace / local plugin |
| Hosts | Any shell-capable agent | Own ecosystem | Usually Claude Code | Usually Claude Code |
| Jobs / cancel UI | No (keep simple) | Yes | Yes | Often yes |
| Best for | Cross-harness Super Grok peer | Daily driver agent | Codex from Claude | Grok from Claude only |
Use a full runtime when you want a primary agent. Use invoke-grok when you already have an agent and need Grok as a specialist — portable, boring, one-minute install.
A single portable agent skill that teaches coding agents (Claude Code, Codex, Cursor, OpenClaw, Hermes Agent, and others) to shell out to the local Grok Build CLI for review, research, media, and optional implementation.
No. Any host that can run shell commands and load skills (or follow the recipes manually) works: Claude Code, Codex, Cursor, Copilot, OpenClaw, Hermes, custom multi-agent orchestrators.
Yes. Place SKILL.md on the host’s skill path (often .agents/skills/invoke-grok/ or the harness’s user skills directory). The host shells out to grok -p using this skill’s recipes. Invoke Grok is not a replacement for OpenClaw/Hermes — it is a cross-harness skill so those stacks can call Super Grok.
Yes. Install via PowerShell (irm https://x.ai/cli/install.ps1 | iex), put the skill under %USERPROFILE%\.agents\skills\invoke-grok\, and use the same grok -p commands.
Yes by default. Review and critique use plan mode and read-only sandbox flags. Write mode (implement + --yolo) only when the user clearly asks Grok to change code.
Run grok models, then grok -p "…" -m <id>. If you do not care, omit -m and use the default.
You can always run Grok yourself. This skill makes other agents do it correctly: install, auth, health check, safe flags, and “return the review as-is” so they do not rewrite Grok’s judgment. That matters in multi-agent setups where Claude/Codex/OpenClaw/Hermes should not “soft filter” a second opinion.
No. Community skill that shells out to the official Grok Build CLI. Grok Build: x.ai/cli · github.com/xai-org/grok-build.
OpenClaw and Hermes are full agent harnesses (skills ecosystems, memory, long-running agents). Invoke Grok is a thin portable skill that only teaches hosts to call Grok Build. Install both: run your preferred agent daily; invoke Grok for independent review, research, and media.
Yes — each grok -p run uses your Grok account / API limits. Prefer read-only review and avoid auto-loops.
Yes. The skill includes image and video modes: the host asks Grok (via grok -p) to generate or animate and report saved paths. Quality and availability depend on your Grok Build install and account.
More answers: docs/faq.md
| Doc | Contents |
|---|---|
SKILL.md |
Full agent contract (source of truth) |
| docs/install.md | OS install, auth, skill paths |
| docs/modes.md | Mode recipes |
| docs/faq.md | Expanded FAQ for answer engines |
AGENTS.md |
One-liner for agent bootstraps |
PRs welcome: clearer install paths, host-specific notes, better review prompts, translations of the README FAQ.
Keep the product one skill + docs. Prefer not to grow a broker unless there is a strong reason.
Apache License 2.0 — see LICENSE.
Grok Build and related marks belong to their owners. This project is not affiliated with or endorsed by xAI unless stated otherwise.