The MCP protocol has achieved something remarkable over the past year: getting all AI frameworks to use the same interface for tool calling. But one problem has persisted — MCP server quality varies widely, enterprises have to maintain their own, and they are responsible when things break.
Google quietly solved this last December: officially hosted MCP servers dedicated to its own cloud services, deploy and use immediately, with Google taking responsibility for issues.
What is connected and how
The first batch includes four services: BigQuery, Google Maps, Google Compute Engine (GCE), and Google Kubernetes Engine (GKE). The choice is unsurprising — they cover the four most common scenarios for enterprise AI agents: data analysis, geospatial, cloud infrastructure, and container orchestration.
In practice: developers only need to point their AI agent client (such as Gemini CLI) to the Google-hosted MCP endpoint to directly call these services, without any installation, configuration, or maintenance.
What was the previous approach? Community members built unofficial BigQuery MCP servers that you deploy and manage yourself, and update yourself when versions fall behind. This is the common weakness of all "community MCP" solutions: useful but fragile, unsuitable for production.
The official version differs in stability and security:
- Permission management: uses the familiar Google Cloud IAM mechanism
- Audit logs: operations go through Cloud Audit Logs, meeting enterprise compliance requirements
- Security protection: Google Cloud Model Armor specifically intercepts indirect prompt injection attacks coming through MCP
- Service discovery: Cloud API Registry manages which tools are available, preventing AI agents from calling unintended interfaces
After March 17: automatic enablement
A noteworthy detail: from March 17, 2026, if you enable a supported service on GCP, the corresponding MCP server will be turned on automatically.
The implication is clear: MCP is no longer a feature requiring "extra configuration" — it is a standard interface for Google Cloud services. Just as BigQuery naturally has a REST API, it now naturally has MCP.
Additional services are being progressively added:
- Compute: Cloud Run, Cloud Storage, Cloud Resource Manager
- Databases: AlloyDB, Cloud SQL, Spanner, Memorystore
- Analytics: Looker, BigQuery Migration Service, Dataplex
- Messaging: Pub/Sub, Kafka
- Security and operations: Google Security Operations, Cloud Logging, Cloud Monitoring
- Developer tools: Developer Knowledge API, Android Management API
This essentially covers Google Cloud's full service map.
What Anthropic says
David Soria Parra, co-creator of MCP now at Anthropic, gave a clear statement on this development:
"Google's support for MCP across such a diverse range of products, combined with their close collaboration on the protocol specification, will help more developers build agentic AI applications. As adoption rates rise across major platforms, we are getting closer to the goal of 'AI agents working seamlessly between the tools and services people use every day.'"
The logic behind this: MCP's value lies in network effects — the more services connected, the more AI agents can do, and the more entrenched the protocol becomes. Google's move does something for MCP that it could not do for itself: convince developers that the protocol is serious and will not disappear.
A concrete example
Google provided a real use case: an AI agent for retailer store location selection.
- Pull sales data from the past 3 years across cities using BigQuery to generate revenue forecasts
- Call Google Maps services to identify competitor locations, surrounding amenities, and delivery radius in target cities
Previously, these two tasks required two separate SDKs. Now, the same MCP client calls both endpoints. Data flows within Google's infrastructure, no need to build your own data pipelines, and permission management uses the same IAM system.
This example illustrates Google's core logic: it is not about riding the MCP trend, but about making the Gemini + GCP combination more attractive to enterprises. You use Gemini for agents, tool calling goes through official MCP, support is available when issues arise, and billing is in one place. This is far stronger than cobbling together open-source tools.
For AI agents to be useful, tool calling must be reliable. Google's official MCP solves the question of "who is responsible for the tools."
Sources: CocoLoop, Google Launches Managed Remote MCP Servers for Its Cloud Services (The New Stack); Announcing official MCP support for Google services (Google Cloud Blog)