What happened
Google Cloud’s November 20, 2026 announcement describes three production challenges: keeping an MCP server off the public internet, verifying both the Gemini Enterprise service and the signed-in human user, and avoiding DNS collisions in large Shared VPC environments. Its reference architecture uses an Egress Agent Gateway, Private Service Connect (PSC), Private Cloud DNS, Cloud Run IAM, and OAuth 2.0.
The network path has two bridges. First, a regional Egress Agent Gateway in Agent-to-Anywhere mode connects to a PSC Network Attachment in a regional subnet such as us-central1 or europe-west1. This puts a managed network interface in the VPC without public IPs or inbound firewall rules. Second, traffic from the VPC reaches Cloud Run privately. Google offers two options: Subnet Private Google Access, using 199.36.153.8/30, or a global PSC endpoint for Google APIs at 240.0.0.10. The latter injects a dedicated /32 route, allowing private Cloud Run routing in locked-down VPCs without a default internet route.
For DNS, the post recommends a private zone for the MCP service’s exact fully qualified domain name rather than a wildcard *.run.app zone. The gateway resolves that hostname before sending traffic into the VPC. With longest-suffix matching, other Cloud Run services continue to resolve normally. The gateway connects to a private VIP while preserving the Cloud Run hostname in TLS SNI and the HTTP Host header, so Cloud Run can route the request to the intended container.
Authentication is split across two headers. X-Serverless-Authorization carries a Google-signed OIDC ID token for the project’s Gemini Enterprise service agent; Cloud Run IAM checks its Invoker role and strips the header before the request reaches the container. Authorization: Bearer carries the signed-in user’s OAuth 2.0 access token, which application middleware can validate to enforce individual permissions. Google’s example middleware exempts /health and can reject missing or invalid tokens when REQUIRE_OAUTH is enabled.
The five setup steps cover locking down Cloud Run and granting invoker access; provisioning the VPC, network attachment, and gateway; configuring FQDN-scoped DNS and peering; binding the gateway and registering the MCP server in Gemini Enterprise; and testing calls in Cloud Logging. Google says its companion GitHub repository includes a Python server, OAuth middleware, deployment scripts, console instructions, and local smoke tests for both routing options.
What the archive adds
New Runtime’s earlier coverage described MCP as a security boundary, including risks from untrusted dependencies, network access, and exposed tokens. A separate earlier note covered connecting Gemini through Cloud Run. This is a new implementation detail on those threads: Google lays out private routing and distinct infrastructure and user identity checks. Related coverage: DeployHQ’s hosted, read-only MCP server.
Trend signal
This supports agent protocols becoming an interoperability layer: MCP connectivity is being treated as production infrastructure, with a managed gateway, scoped network access, DNS, and identity controls around the protocol. The important signal is not merely that an MCP endpoint exists, but that enterprise teams need a governed path from agent platform to internal capabilities.
Why now
Google frames this as a production problem arising as enterprise teams build custom MCP servers to connect agents with internal APIs, databases, and workflows. The design addresses the gap between a working prototype and an internal service that must stay private and enforce user-specific permissions.
Why it matters
An MCP server can expose actions backed by sensitive company systems, so network reachability and user authorization are consequential design choices. This architecture separates the right to reach the Cloud Run service from the right of an individual employee to use a particular operation. That distinction can prevent an agent integration from collapsing into one broad service identity with no per-user check.
The DNS guidance also matters at scale: a wildcard private zone can disrupt unrelated Cloud Run services in a Shared VPC, while an exact-hostname zone narrows the override. Teams can use the two routing choices to match their VPC posture, then validate actual tool discovery and execution in Cloud Logging. The recipe is specific to Google Cloud, but the separation of network, service identity, and human identity is a reusable way to reason about internal agent integrations.
Limits and uncertainty
Google describes an architecture and provides a reference implementation, but the announcement does not report independent security review, operational measurements, or deployment results across customer environments. The example’s OAuth middleware shows token validation, while teams still need to define which user claims map to which MCP actions and how credentials are managed. The post also does not quantify latency, cost, or the operational burden of maintaining gateway, DNS, and IAM configuration. Its mechanisms are tied to Gemini Enterprise and Google Cloud services; portability to other agent platforms is not established.
Skill to build now
Map one internal MCP integration as three separate controls: who can reach its network endpoint, which platform identity may invoke the service, and which user identity authorizes each action. For a Cloud Run deployment, test the exact service hostname in a dedicated private DNS zone before considering any wildcard, then verify tool initialization, discovery, and execution in Cloud Logging. Keep a negative test for missing or expired user tokens alongside a successful service-authenticated call.
Reader takeaway
Treat private MCP connectivity as a layered system, not a URL plus a token. Google’s design keeps the endpoint private, scopes DNS to one service, and checks both the Gemini service identity and the human user—controls that matter when agents can reach internal business systems.