{"slug": "claude-code-antigravity-cli-agents-skill", "title": "Claude Code Antigravity CLI Agents Skill", "summary": "Mark Fulton released a Claude Code skill that turns the AI assistant into an orchestrator, delegating heavy coding tasks to Google Antigravity CLI sub-agents while Claude continues working. The skill saves tokens by offloading full-repo audits and refactors to background jobs, enables parallel work, and lets users run tasks on different models like Gemini for independent verification.", "body_md": "*Turn Claude Code into an orchestrator — delegate heavy jobs to Google Antigravity sub-agents while it keeps building.*\n\n**Created by Mark Fulton** · Founder of the\n\n**community (300k+ members)**\n\n[Vibe Coding is Life](https://facebook.com/groups/vibecodinglife)Co-authored by Claude Fable 5\n\nA single Claude Code **skill** — one markdown instruction file (`SKILL.md`\n\n) named `antigravity-agents`\n\n. Skills are capabilities [Claude Code](https://claude.com/claude-code) loads on demand. This one teaches it to hand off work to the [Google Antigravity CLI](https://antigravity.google/docs/cli/getting-started) (`agy`\n\n), an autonomous terminal coding agent powered by Gemini and other models.\n\nWith the skill loaded, Claude Code stops being a single worker and becomes an **orchestrator**: it writes a self-contained brief, launches an `agy`\n\njob in the background, keeps working on your task, then collects and **verifies** the result before trusting a line of it.\n\n**Save your tokens.** Full-repo audits, big refactors, and research sweeps are expensive to run inside a Claude session. Offload them to Antigravity instead of burning your Claude context and subscription tokens on the grind work.**Broaden your model choice.** Run a job on Gemini (or whatever Antigravity offers) for a genuinely independent second opinion — or just to put the right model on the right task.**Work in parallel.** Fan out several sub-agent jobs at once while Claude Code stays responsive to you. More gets done per unit of wall-clock time.\n\n| Solo Claude Code session | With `antigravity-agents` |\n|\n|---|---|---|\n| Long audits / refactors | Run inline, spending your context | Handed off to a background sub-agent |\n| Second opinion | Same model, same session | An independent model reviews the work |\n| Throughput | One task at a time | Multiple jobs run while Claude stays responsive |\n| Your Claude tokens | Spent on the heavy lifting | Spent on orchestration and verification |\n\nClaude writes the brief, launches the job non-interactively (`agy -p \"<task>\"`\n\n), and immediately returns to your work. The sub-agent runs on its own. When it finishes, Claude reads the output and **verifies it** — `git diff`\n\n, typecheck, build — before any of it is trusted or merged.\n\n``` php\nflowchart LR\n    A[Claude Code<br/>orchestrator] -->|writes self-contained brief| B[\"agy -p '&lt;task&gt;'\"]\n    B -->|launched in background| C{Job type}\n    C -->|read-only| D[Sandbox mode<br/>runs concurrently]\n    C -->|writes files| E[Isolated worktree<br/>or separate repo]\n    A -. keeps building your task .-> A\n    D --> F[Result]\n    E --> F[Result]\n    F -->|git diff · typecheck · build| G[Claude verifies]\n    G -->|trusted| H[Folded into your work]\n```\n\nTwo one-time steps.\n\n**1. Install the Antigravity CLI.**\n\nOn macOS / Linux:\n\n```\ncurl -fsSL https://antigravity.google/cli/install.sh | bash\n```\n\nOn Windows (PowerShell):\n\n```\nirm https://antigravity.google/cli/install.ps1 | iex\n```\n\nFull details are in the official guide: [https://antigravity.google/docs/cli/getting-started](https://antigravity.google/docs/cli/getting-started)\n\n**2. Sign in to Google.**\n\nRun `agy`\n\nonce in a terminal and complete the one-time Google OAuth sign-in in your browser:\n\n```\nagy\n```\n\nAfter that, Claude Code can launch `agy`\n\njobs non-interactively — no further prompts.\n\nClone this repo and copy the `antigravity-agents`\n\nfolder into your Claude Code skills directory. Claude Code auto-discovers it on the next run.\n\n**macOS / Linux**\n\n```\ngit clone https://github.com/markfulton/claude-antigravity-agents.git\ncp -r claude-antigravity-agents/antigravity-agents ~/.claude/skills/\n```\n\n**Windows (PowerShell)**\n\n```\ngit clone https://github.com/markfulton/claude-antigravity-agents.git\nCopy-Item -Recurse claude-antigravity-agents\\antigravity-agents \"$env:USERPROFILE\\.claude\\skills\\\"\n```\n\nThat's it. Start a Claude Code session and ask it to delegate something — the skill loads on demand.\n\nJust talk to Claude Code the way you already do. A few prompts that trigger the skill:\n\n\"Have Antigravity review my current branch diff for bugs while you keep building the checkout flow.\"\n\n\"Spin up a sub-agent to audit this repo's dependencies for known vulnerabilities and report back.\"\n\n\"Get a second opinion from Gemini on this caching design, and keep working on the API in the meantime.\"\n\nClaude decides the job type, picks the right isolation mode, launches it in the background, and reports back once it has collected and verified the result.\n\nDelegation only helps if you can trust what comes back. One rule sits above everything else:\n\nAn Antigravity job never writes to the same files Claude Code is editing at that moment.\n\nThat rule drives how every job runs:\n\n| Job type | Writes files? | Isolation | Concurrent with Claude? |\n|---|---|---|---|\n| Code review | No | Sandbox (read-only) | Yes |\n| Architecture / security analysis | No | Sandbox (read-only) | Yes |\n| Research sweep | No | Sandbox (read-only) | Yes |\n| Dependency / config audit | No | Sandbox (read-only) | Yes |\n| Second opinion (different model) | No | Sandbox (read-only) | Yes |\n| Coding task | Yes | Isolated git worktree | Yes — on separate files |\n| Large refactor | Yes | Worktree / separate repo | Yes — on separate files |\n\n**Everything is verified before it's trusted** — Claude inspects the `git diff`\n\nand runs the project's typecheck / build on any returned work before folding it in. No sub-agent change is taken on faith. Keep a task in Claude when it's small, tightly coupled to what you're editing right now, or needs live back-and-forth; delegation shines on the heavy, separable work.\n\n**Created by**[Mark Fulton](https://markfulton.com)- Founder of the\ncommunity (300k+ members)[Vibe Coding is Life](https://facebook.com/groups/vibecodinglife) - Co-authored by\n**Claude Fable 5**\n\nReleased under the [MIT License](/markfulton/claude-antigravity-agents/blob/main/LICENSE).", "url": "https://wpnews.pro/news/claude-code-antigravity-cli-agents-skill", "canonical_source": "https://github.com/markfulton/claude-antigravity-agents", "published_at": "2026-07-07 13:07:48+00:00", "updated_at": "2026-07-07 13:30:24.374230+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "large-language-models", "generative-ai"], "entities": ["Mark Fulton", "Claude Code", "Google Antigravity CLI", "Gemini", "Claude Fable 5", "Vibe Coding is Life", "GitHub"], "alternates": {"html": "https://wpnews.pro/news/claude-code-antigravity-cli-agents-skill", "markdown": "https://wpnews.pro/news/claude-code-antigravity-cli-agents-skill.md", "text": "https://wpnews.pro/news/claude-code-antigravity-cli-agents-skill.txt", "jsonld": "https://wpnews.pro/news/claude-code-antigravity-cli-agents-skill.jsonld"}}