On 30 June 2026 Google Cloud shipped a fully managed remote MCP server for the Gemini Enterprise Agent Platform. Instead of local-only setups or bespoke point-to-point integrations, an external agent — one you build in Claude Code, Antigravity, or any MCP-compatible framework — reaches your Google Cloud resources through a single managed HTTP endpoint. Here's what it exposes and how to switch it on.
What the managed server gives you
The point is to let a tool you're already using reach into a project without leaving the IDE. An agent can call models from Model Garden, run predictions, pull down shared prompt templates, and manage notebooks directly — all over MCP. The toolsets map to clear paths: /mcp/generate for generative calls, /mcp/predict for prediction, /mcp/notebook for notebook management, and /mcp/prompts for prompt management, alongside endpoints for model registry operations, fine-tuning, and evaluation.
Because your agent stays fully compliant with the open MCP specification, external IDEs and frameworks interact with the cloud environment without locking you into a proprietary path — the same agent code that talks to our directory servers talks to Google Cloud.
Turn it on
There's almost nothing to install. The remote MCP server is automatically enabled when you enable the Gemini Enterprise Agent Platform API in your Google Cloud project. Once the API is on, the endpoint is live and ready for a client to connect.
Authentication and access control
The server uses OAuth 2.0 with Identity and Access Management (IAM) for authentication and authorization, and all Google Cloud identities are supported. That means the same principals and roles you already govern apply to agent traffic — no separate credential store. IT teams can go further and use native Cloud IAM Deny policies to guarantee that external developer frameworks only reach authorized resources, so a misconfigured agent can't wander outside its lane. If OAuth 2.0 resource-server terms are new to you, our MCP authorization guide covers the model.
Connecting a client
Point your host at the managed HTTP endpoint and supply an OAuth token. In Claude Code or Antigravity you add it as a URL-based server; the client discovers the toolsets and surfaces them like any local server's tools. This is a remote, Streamable-HTTP-style connection, so if you're deciding between transport styles, read Streamable HTTP vs SSE vs stdio first.
How it fits the rest of Google's MCP story
This managed server is the enterprise counterpart to the developer-facing tooling. If you work at the terminal, the Gemini CLI MCP setup and Antigravity CLI MCP setup cover client-side configuration, while Google ADK MCP servers show how to consume MCP tools from agents you build on Vertex AI. Together they're a full loop: build the agent, give it managed cloud tools, and keep everything inside one IAM boundary. Browse the ML engineer loadout for a vetted starting stack.