Agent Registry on Google Cloud: Cataloging Agents, MCP Servers, and Tools

GCP Study Hub
November 18, 2026

Agent Registry is the central catalog of approved agents, MCP servers, tools, and external REST endpoints on Google Cloud. It serves three functions in an agentic architecture. Agent Gateway checks outbound destinations against the registry, the security team gets one place to see what is in use, and developers can browse approved servers instead of configuring server URLs by hand. On the Professional Agentic Architect exam, Agent Registry comes up both in catalog setup and in outbound traffic governance where every destination must be approved before an agent can call it.

Automatic registration for Google Cloud and same-project servers

Registration happens in two ways, automatically or manually. Automatic registration covers only Google Cloud MCP servers and custom MCP servers deployed on Cloud Run or GKE in the same project.

  • Google Cloud MCP servers: Google Cloud MCP servers, such as BigQuery and Compute Engine, register automatically in the global location when their APIs are enabled. Writing tool specifications for these servers duplicates work that Agent Registry already does on its own.
  • Custom MCP servers on Cloud Run: A custom MCP server deployed as a Cloud Run service in the same project with the --functional-type=mcp-server flag registers automatically, with no separate tool specification file. This works for Cloud Run services and not jobs, because an MCP server must keep serving requests, which a Cloud Run job does not do.
  • Custom MCP servers on GKE: On GKE, the label registry.gke.io/functional-type with the value MCP_SERVER is the equivalent for automatic registration in the same project. That label has no effect on Cloud Run.

Hosting choices matter here as well. Agent Runtime hosts agents and does not support hosting custom MCP servers, so custom servers belong on platforms such as Cloud Run or GKE. When a custom MCP server runs on Cloud Run or GKE in the same project, we would lean toward automatic registration rather than manual registration so the team does not have to maintain a separate tool specification file.

Manual registration and tool specification files

Any server outside those automatic paths, such as a third-party hosted MCP server running outside Google Cloud, is registered manually with gcloud agent-registry services create. The automatic mechanisms depend on the server being deployed on Cloud Run or GKE in the same project, so a server that a vendor hosts elsewhere always goes through manual registration.

Manual and automatic registration differ in what the catalog learns about a server. A manual registration records the endpoint, and Agent Registry does not introspect the server to discover its tools. Everything developers can see about a manually registered server's tools therefore comes from the tool specification supplied with the registration.

That specification is a toolspec.json file that follows the MCP specification for tool schemas. The file can describe each tool with annotations such as readOnlyHint or destructiveHint, and it is limited to 10 KB. Including it when an external server is registered is what makes the server's tools discoverable in the catalog.

Catalog contents and IAM tool permissions are separate concerns. The MCP Tool User role applies to Google Cloud MCP servers, where it controls permission to call tools, while the catalog's tool listings describe what a registered server offers.

Enforcing default-deny egress with Agent Gateway

Agent Registry provides the catalog, and Agent Gateway provides the enforcement point between agents and the tools, APIs, and agents they call. When an agent on Agent Runtime is bound to an Agent Gateway in Agent-to-Anywhere mode, also called egress mode, the gateway checks each outbound destination against Agent Registry and evaluates the access policies for the agent's identity.

When no matching access policy exists, the gateway denies the request by default. Because Agent Registry catalogs MCP servers, agents called over A2A, and external REST endpoints, an Agent-to-Anywhere gateway gives security and compliance teams one control point across all of those outbound destination types. Any new destination stays blocked until it is approved, cataloged in Agent Registry, and covered by an egress policy for the agent's identity, with no changes to the agent's code.

In practice, this makes registration and policy grants part of the change process for every new destination. The gateway evaluates the registry and the identity's egress policies on each request, so the catalog entry and the policy have to be in place before an agent starts calling a new server, not after.

Distinguishing Agent Registry from adjacent controls

Scenarios on the Professional Agentic Architect exam often place Agent Registry and Agent-to-Anywhere gateways alongside controls that solve different problems.

  • Client-to-Agent gateway mode: Governs traffic coming into an agent from users, so it never sees outbound calls to MCP servers, REST APIs, or partner agents.
  • VPC Service Controls: Governs access to Google Cloud APIs through perimeters and egress rules, and does not catalog third-party MCP servers or partner agents.
  • Firewall rules behind Cloud NAT: Operate on IP addresses only, cannot recognize MCP or A2A traffic, and break whenever a provider changes its addresses.
  • Secret Manager secrets and Cloud Run forwarding proxies: Storing server URLs in a shared Secret Manager secret or routing through a forwarding proxy on Cloud Run hides the servers instead of giving the security team a catalog to review, and a proxy adds a custom service to run and secure.
  • In-code allowlists: Checking approved destinations inside the agent's own code requires code changes and depends on the code always running the check.
  • Model Armor and network attachments: Model Armor on the gateway screens the content of tools/call and prompts/get requests, not tools/list, so it filters what a request contains rather than cataloging or authorizing destinations. A network attachment provides private connectivity into a VPC network, a networking concern that is separate from catalog membership and egress policy.

Our Professional Agentic Architect practice exams test Agent Registry alongside Agent Gateway egress modes and MCP server governance, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow