
The Model Context Protocol (MCP) is the standard way to connect an agent to tools and data. An MCP server exposes tools, each with a name, a description, and a schema for its arguments, and an agent connects as a client to list them with tools/list and call them with tools/call. For the Professional Agentic Architect exam, the key distinctions sit around where MCP fits relative to other Google Cloud agent services, how ADK connects to servers and limits tool access, where custom servers can run, and how Agent Registry catalogs them.
MCP is built for plain tool operations, such as lookups and writes against databases or APIs, that do not involve reasoning or multi-step conversations. When multiple agents need the same capabilities, exposing those operations from a single MCP server lets every agent use the same tools while keeping credentials in one place. Any change to the underlying database or API then happens once in the server rather than across every agent deployment.
Several neighboring patterns solve different problems and do not fit plain tool work:
MCP has two transports. The stdio transport launches the server as a local subprocess, so it works only for servers running on the same machine. The streamable HTTP transport connects an agent to a remote MCP server hosted somewhere else.
In ADK, an agent connects through McpToolset, using StdioConnectionParams for a local server or StreamableHTTPConnectionParams for a remote server over streamable HTTP. ADK also offers OpenAPIToolset, which builds tools from an OpenAPI specification. The two toolsets follow how the tools are described: an MCP server advertises its tools through tools/list, while a REST API is described by an OpenAPI specification.
The tool_filter parameter on McpToolset exposes only the named tools to the model, so the remaining tools on the server never appear in its list. The ADK documentation recommends always setting tool_filter to expose only the actions an agent needs. Naming allowed tools in the agent's instruction works differently, because the model still receives every tool on the server and nothing stops it from calling the rest. Limiting tools in code with tool_filter is an enforced control, while an instruction is only guidance.
Google Cloud provides its own MCP servers for services such as BigQuery and Compute Engine, which are turned on by enabling their APIs. When a team builds a custom MCP server over streamable HTTP, hosting options become a frequent distinction on the Professional Agentic Architect exam.
Agent Runtime is built for Python agents and does not support hosting custom MCP servers. Custom MCP servers that use streamable HTTP can be hosted on Cloud Run. Cloud Run hosts both MCP clients and servers packaged as containers, and Google recommends Cloud Run for containerized applications that need serverless, event-driven scaling with language flexibility, including TypeScript agents built with ADK. Cloud Run scales the number of instances with traffic and down to zero. Because Cloud Run services are stateless, session and memory data belong in a service such as Agent Platform Sessions or Firestore.
Two operational details matter on Cloud Run. First, a custom MCP server must run as a Cloud Run service rather than a Cloud Run job, because an MCP server has to keep serving requests. Second, Cloud Run sits in the middle of Google's operational overhead ladder. It is less managed than Agent Runtime, yet it leaves far less to the team than GKE, which offers fine-grained control for complex, stateful systems and needs people who can operate Kubernetes, or a Compute Engine managed instance group, where patching, scaling, and container management fall to the team.
Agent Registry provides a central catalog of agents, MCP servers, and tools. It gives security teams one place to review which servers are in use, and developers can find approved servers in the registry instead of configuring server URLs by hand.
How an MCP server enters Agent Registry depends on how it is provided and hosted:
--functional-type=mcp-server flag, and Agent Registry then registers it automatically with no separate tool specification. Automatic registration applies to Cloud Run services, not jobs, and only within the same project. On GKE, the registry.gke.io/functional-type label is the equivalent, and it has no effect on Cloud Run.gcloud agent-registry services create, along with a tool specification file that describes the server's tools.A catalog is different from simply hiding addresses. Keeping server URLs in a shared Secret Manager secret or routing calls through a forwarding proxy on Cloud Run keeps addresses out of agent code, but neither gives a security team a reviewable list of the servers agents use, and a proxy is one more custom service to run and secure.
Our Professional Agentic Architect practice exams test Model Context Protocol (MCP) alongside ADK tool configuration and Agent Registry governance, with detailed explanations that drill these distinctions.