Setting up the Google Cloud MCP Server (gcloud-mcp)

This guide walks through setting up Google's official Google Cloud MCP server (gcloud-mcp) so an AI assistant can control Google Cloud resources in natural language through the gcloud CLI.

Prerequisites

Node.js version 20 or later (node --version)
An installed and authenticated gcloud CLI (gcloud auth login)
An MCP-capable AI client (Gemini CLI, Claude Desktop, Cline, Cursor, VS Code, or similar)

Option 1: Gemini CLI / Gemini Code Assist (recommended for Gemini users)

Run from your home directory:

npx @google-cloud/gcloud-mcp init --agent=gemini-cli

This registers the server as a Gemini CLI extension available across all projects. Verify the setup with:

gemini mcp list

Option 2: Other MCP clients (Claude Desktop, Cline, Cursor, VS Code)

Add the following block to the client's MCP configuration file:

{
  "mcpServers": {
    "gcloud": {
      "command": "npx",
      "args": ["-y", "@google-cloud/gcloud-mcp"]
    }
  }
}

The three other sub-servers can be added as needed, each as its own entry with the matching package name:

{
  "mcpServers": {
    "gcloud": {
      "command": "npx",
      "args": ["-y", "@google-cloud/gcloud-mcp"]
    },
    "observability": {
      "command": "npx",
      "args": ["-y", "@google-cloud/observability-mcp"]
    },
    "storage": {
      "command": "npx",
      "args": ["-y", "@google-cloud/storage-mcp"]
    },
    "backupdr": {
      "command": "npx",
      "args": ["-y", "@google-cloud/backupdr-mcp"]
    }
  }
}

In VS Code, the root key is servers instead of mcpServers inside .vscode/mcp.json.

Verify your installation

After setup, refresh the tools list in the AI client and send a simple test prompt, for example:

List my Google Cloud projects.

If the agent returns a project list, the server is active and correctly authenticated.

Example prompts

"Show the last 20 error log entries in project <project-id> from the last hour"
"List all objects in bucket <bucket-name>"
"Which backup vaults exist in the europe-west3 region?"
"Create a new bucket <name> in the europe-west1 region"

Security guidance

Use a restricted account — For production setups, use a service account with minimal permissions via impersonation rather than a personal account with full access.

Deliberately enable write tools — Tools like delete_bucket, delete_backup_vault, or restore_backup perform real changes. Check whether the agent actually needs these tools before production use.

Mind the preview status — Google explicitly states this project is not an officially supported solution covered by the Google Cloud Terms of Service. For business-critical automation, test with restricted accounts before a broad rollout.

Watch the data path to the model — Log, object, and configuration content is forwarded to the AI client and from there to the model provider. Do not let secrets or personal data flow uncontrolled through the agent.

Troubleshooting

Server does not connect — Check gcloud auth list for an active account. Re-authenticate with gcloud auth login if needed.

Command gets rejected — The core gcloud server deliberately blocks certain commands (such as interactive sessions). This is intentional, not a bug.

Node.js version too old — Check node --version; version 20 or later is required.