How to Add and Use Antigravity Skills
Skills run from workspace folders under .agents/skills while MCP servers connect through a separate configuration file.
To add skills to Google Antigravity, create a dedicated subfolder for each skill under .agents/skills inside the project workspace directory. The agent loads the matching skill automatically whenever a user prompt or assigned task aligns with that skill's capabilities.
Developers migrating from the Gemini command-line interface (CLI) can carry over existing skills, hooks, and subagents directly into Google Antigravity.
Google launched Antigravity 2.0 at Google I/O on May 19, 2026, and the Antigravity CLI replaced Gemini CLI for individual users. As of October 2026, skills work in both the CLI and the 2.0 desktop app, which share one agent harness. The Antigravity software development kit (SDK), a Python research preview, includes skills as well.
How to Add a Skill to Antigravity
Adding a skill to Google Antigravity requires creating an individual folder under .agents/skills in the active project workspace.
The agent loads the matching skill when given a matching task. Google's official documentation defines the exact file format and structure requirements at the Google Antigravity overview documentation, so developers should consult those specifications before writing custom logic.
- 1. Open the project root directory in your terminal or integrated development environment (IDE).
- 2. Create the
.agents/skillsdirectory if it does not already exist in the workspace. - 3. Add a dedicated subfolder for the specific skill inside
.agents/skills, using a descriptive name for the task. - 4. Place the skill definition files inside that subfolder according to the structure specified in the Google Antigravity overview documentation.
- 5. Prompt the agent with a task matching the skill domain to verify execution.
Migration of Gemini CLI Skills to Antigravity
Gemini CLI skills transfer into Antigravity by moving their definition folders from .gemini/skills into .agents/skills.
On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions stopped serving requests for Google AI Pro, Google AI Ultra, and free individual users, as announced on the Google Developers Blog. Google's Gemini CLI migration guide lists what moves where.
Existing skills, hooks, and subagents carry over directly into Antigravity workspaces. Project-level instruction files named GEMINI.md and AGENTS.md continue to apply without renaming. Gemini CLI extensions convert into Antigravity plugins with agy plugin import gemini.
Google noted at launch that Antigravity did not offer complete feature parity with Gemini CLI. Teams migrating custom automation must test each imported skill individually inside the new runtime to ensure consistent results. More details on the command line interface appear in the Google Antigravity CLI guide.
- Move existing skill directories out of
.gemini/skillsand place them directly into.agents/skills. - Retain existing
GEMINI.mdandAGENTS.mdfiles, which still apply in Antigravity. - Convert legacy Gemini CLI extensions into plugins using the terminal command
agy plugin import gemini. - Test each migrated skill with targeted prompts to confirm behavior under the updated Antigravity execution engine.
Roles of Skills, Rules, Hooks, Subagents, and Plugins
Google Antigravity gives the agent five separate mechanisms: skills, rules, hooks, subagents and plugins.
The agent loads a skill when an assigned task matches that skill. Rules appear in Google's overview documentation alongside skills, MCP and plugins, so read that page before writing one.
Hooks execute shell scripts configured via JavaScript Object Notation (JSON) before or after tool and model calls. Workspace-level hooks override global hooks, so a single project can carry its own checks. Subagents run in parallel and inherit the main agent's permissions, so a subagent can do anything the main agent is allowed to do. Plugins are distributed through a marketplace, and Gemini CLI extensions convert into plugins with agy plugin import gemini.
| Mechanism | What It Does for You |
|---|---|
| Skills | Each skill sits in its own folder under .agents/skills, and the agent loads it when a task matches |
| Hooks | JSON-configured shell scripts run before or after tool and model calls, and workspace hooks override global ones |
| Subagents | Run in parallel with the main agent's permissions, so review what a subagent may touch |
| Plugins | Come from the plugin marketplace, or from Gemini CLI extensions via agy plugin import gemini |
Configuration of Antigravity MCP Servers
Model Context Protocol servers in Google Antigravity are configured in a file named mcp_config.json.
Developers coming from Gemini CLI move their MCP servers into mcp_config.json as a separate step from moving skills out of .gemini/skills into .agents/skills. The Antigravity SDK also includes MCP.
Antigravity is included in eligible Gemini Enterprise Standard and Plus plans, where admins can restrict MCP, according to the Google Antigravity enterprise announcement. Admins there also get sandbox controls, audit logging and budget caps, so on a company account, ask your admin which MCP servers are allowed before adding one to mcp_config.json.
Skills in the Antigravity CLI and SDK
Skills work in both the Antigravity CLI and the standalone 2.0 desktop app, and the Antigravity SDK includes skills too.
The Antigravity CLI runs as agy and is written in Go. The CLI uses the same agent harness and settings as the 2.0 desktop app, which TechCrunch reported Google launched at I/O on May 19, 2026. Terminal conversations initiated with agy can open directly in the 2.0 application for inspection. The Google Antigravity 2.0 guide covers the desktop app in more detail.
The Antigravity SDK is a Python research preview released under the Apache 2.0 license, installed with pip install google-antigravity. According to the Google Antigravity SDK announcement, the SDK includes built-in support for skills, hooks, a policy engine, subagents, structured output and MCP. Google has announced future SDK support for TypeScript and Go.
The original Antigravity IDE, launched on November 18, 2025, is not deprecated, and Google recommends that developers retain it alongside the 2.0 desktop app. Google has announced that the Agent Manager will be removed from the IDE in an upcoming release, and the 2.0 desktop application succeeds the IDE's Agent Manager.
- CLI users install the tool via
curl -fsSL https://antigravity.google/cli/install.sh | bashon macOS, Linux, and Googlebook, orirm https://antigravity.google/cli/install.ps1 | iexon Windows PowerShell. - CLI sessions authenticate through a browser Google account login or headlessly via
GEMINI_API_KEYwith"modelProvider":"gemini". - The Python SDK provides programmatic access to skills and policy engines through
pip install google-antigravity.
Security Risks and Defenses for Third-Party Skills
Loading third-party skills into Google Antigravity introduces execution risks that require terminal sandboxing and strict manual code review.
In April 2026, CyberScoop reported on security research showing that prompt injection could lead to code execution in Antigravity even in strict mode. Google resolved that specific vulnerability on February 28, 2026, but the finding illustrates the risk of executing unvetted procedural instructions inside agent runtimes.
Operational hazards have also occurred during unconstrained local execution. In December 2025, The Register reported that an Antigravity Turbo-mode command wiped a user's drive. In February 2026, VentureBeat reported that Google banned accounts of users who connected the third-party agent OpenClaw, so connecting outside agents to an Antigravity account carries a ban risk.
To mitigate these risks, Google Antigravity provides built-in isolation mechanisms. The runtime includes a terminal sandbox, project permission boundaries, and browser controls featuring domain allow and deny lists operating inside an isolated Chrome profile. Developers should audit every file in a third-party skill directory before placing it into .agents/skills, and keep sandbox protections active.
Teams that need a fixed, published skill file specification before committing should start with Google's overview documentation, which covers skills.
To get started, create one folder for your first skill under .agents/skills in your workspace. Then give agy or the 2.0 app a task that matches the skill and check that the agent loads it.
Frequently Asked Questions
- Add skills by creating a dedicated subfolder for each skill inside
.agents/skillsin your workspace. The agent loads the skill when the task matches. - The Antigravity skills folder is located at
.agents/skillswithin the active project workspace. Each specific skill resides in its own named subfolder under that directory path. - Gemini CLI skills carry over to Antigravity once you move their folders from
.gemini/skillsinto.agents/skills. Project instructions inGEMINI.mdandAGENTS.mdcontinue to function, and extensions convert viaagy plugin import gemini. Google says the Antigravity CLI did not have full feature parity with Gemini CLI at launch, so test each skill after the move. - Add an MCP server by configuring it in
mcp_config.json, the file Antigravity uses for MCP servers. On a Gemini Enterprise account, admins can restrict MCP, so ask them first. - Yes. On eligible Gemini Enterprise Standard and Plus plans, which include Antigravity, admins can restrict MCP and also set budget caps, capped overages and audit logging.
- Yes. The Antigravity CLI shares the underlying agent execution harness and settings with the 2.0 desktop application, allowing custom workspace skills in
.agents/skillsto execute during CLI sessions.
Planning Your Antigravity Workflow Implementation?
Layer3Labs designs structured agent integrations, enterprise guardrails, and automated developer tooling across cloud and local environments. Book an architecture audit to align agent capabilities with security requirements.
Book a Consultation