Google launches remote MCP server for Gemini Enterprise

Google has introduced a remote MCP server for the Gemini Enterprise Agent Platform, designed to connect external AI agents with resources inside Google Cloud.

The server lets developers link external development tools and agent frameworks to Google Cloud projects through the Model Context Protocol, or MCP. Agents built with tools such as Antigravity CLI and Claude Code can interact with Agent Platform resources from within an integrated development environment.

Through that connection, external agents can call models from Model Garden, access shared prompt templates and manage notebooks inside a project. The server is enabled automatically when customers activate the Gemini Enterprise Agent Platform API in a Google Cloud project.

Open standard

The launch builds on Google's recent MCP work, following its announcement that more than 50 Google-managed MCP servers were available. Here, the focus is on linking third-party and external agent tools to services inside a customer's cloud environment.

MCP is an open specification intended to let AI systems connect to tools, data sources and software services in a standardised way. Agents built outside Google Cloud remain compliant with the MCP specification, reducing dependence on a proprietary approach, Google argued.

Google is positioning the remote MCP server as a single interface for external agents, aiming to reduce the integration work needed to connect development environments to cloud infrastructure. The pitch addresses a common customer challenge: balancing developer speed with governance requirements around data access and system control.

Security controls

The service runs within Google Cloud infrastructure and includes controls tied to Cloud IAM Deny policies. IT teams can use those policies to restrict external developer frameworks so they interact only with authorised Google Cloud resources.

Google also highlighted Agent Registry within Agent Platform as a central library for storing, searching and governing an organisation's inventory of skills, tools and other AI assets. In practice, customers can catalogue internal resources and expose them to external agents under centrally managed rules.

To use the service, customers enable the relevant API, configure the client to point to the remote server and then use toolset endpoints to begin interacting with Agent Platform resources. Google has published a list of endpoints covering functions across model use, operations and evaluation.

Tool access

Available toolsets include endpoints for generative AI functions, prediction, notebook management, endpoint lifecycle management, model registry operations, fine-tuning, evaluation and prompt management. These endpoints are structured under paths such as /mcp/generate, /mcp/predict, /mcp/notebook and /mcp/prompts.

The breadth of toolsets suggests Google wants the remote MCP server to be relevant across several stages of AI development rather than limiting it to inference alone. External agents can connect not just to model outputs but also to workflows involving prompt engineering, notebook execution, model deployment and quality testing.

That matters because many organisations are building AI software in mixed environments, with developers using external coding tools while data, models and governance controls remain inside a cloud provider's platform. A standard connection layer could make it easier for those teams to use their preferred tooling without moving sensitive resources outside managed infrastructure.

Competition among cloud providers and model companies has increasingly centred on how easily customers can integrate third-party agents and coding assistants into existing systems. Google's move adds to a broader industry push to support open protocols while keeping data access and operational controls within the provider's cloud estate.

The release also reflects the growing importance of agent management in enterprise AI. As businesses move beyond single chatbot deployments towards systems that call tools, chain tasks and access internal assets, the infrastructure used to govern those actions is becoming a larger part of the market.

Google said external IDEs and frameworks can interact with cloud environments without being locked into a proprietary ecosystem, while IT teams can apply native controls over what those agents are allowed to access. Supported toolsets span core generation features, inference and raw prediction, notebook runtime management, model upload and deployment, fine-tuning job tracking, automated quality evaluation and prompt versioning workflows.