Add MCP servers to Antigravity CLI (2026 setup)

Antigravity CLI is Google's terminal-first agent, and it quietly replaces Gemini CLI as the way to drive Gemini models from the command line. It's built in Go, so it starts and runs noticeably faster than its predecessor — and it shares the same agent "harness" as Antigravity 2.0 and the IDE, which means one MCP configuration covers the whole suite. Here's how to add a server and the one bug to watch for.

One config, three surfaces

The headline change for MCP users is a single source of truth. Because Antigravity CLI, the IDE, and Antigravity 2.0 use a common harness, the servers and skills you configure once show up everywhere. No more re-adding the same Postgres or GitHub server in each tool — you edit one file and every surface picks it up.

Where the config lives

Antigravity CLI reads MCP servers from two places, global and per-workspace:

# global, applies everywhere
~/.gemini/config/mcp_config.json

# workspace-local, checked into the repo
.agents/mcp_config.json

The workspace file is the one to commit so a whole team shares the same tools — the same idea as sharing an MCP config across a team.

Add a server

The schema is the familiar mcpServers map. A local stdio server is a command plus args; a remote server is a url, since mcp_config.json supports URL-based servers directly:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/absolute/path"]
    },
    "github": {
      "url": "https://api.githubcopilot.com/mcp/"
    }
  }
}

Save, restart the CLI, and the tools appear in the agent. Good first picks are filesystem and GitHub for coding, or Exa and Firecrawl for research — all curated, categorised and vetted in the directory.

The env-var gotcha

There's a real rough edge worth flagging: a bug in the global MCP configuration currently mishandles environment variables, which forces you to hardcode API keys in mcp_config.json for now. That's fine for a throwaway local test, but do not commit a file with a live key in it. Keep secret-bearing servers in the un-tracked global file until the env handling is fixed, and treat the workspace file as key-free. Our MCP security best practices and credential rotation guide cover safer patterns.

Google Cloud over MCP

Antigravity leans hard into Google's own MCP surface: the core Cloud services expose connections over MCP, and the CLI ships a full MCP client for Google Workspace and Gemini. To reach managed Vertex AI and project resources from the same agent, pair it with the Gemini Enterprise remote MCP server. If you're migrating from Gemini CLI, the Gemini CLI MCP setup maps the old config to the new one, and the DevOps loadout is a ready-made stack to start from.