Claude Code on Agent Platform: Model Selection and MCP Server Scopes

GCP Study Hub
November 20, 2026

Running Claude Code with Claude models on Agent Platform involves two separate configuration layers: selecting Agent Platform as the model provider with CLAUDE_CODE_USE_VERTEX, and connecting MCP servers through configuration scopes. Keeping those layers distinct matters on the Professional Agentic Architect exam, because several nearby files and Google Cloud components look related to tool sharing without actually connecting an MCP server to Claude Code.

Selecting Agent Platform as the model provider

CLAUDE_CODE_USE_VERTEX selects Agent Platform as the model provider for Claude Code. That setting controls only which platform serves the Claude models. It has no effect on which MCP servers Claude Code loads, so setting CLAUDE_CODE_USE_VERTEX does not expose an MCP server's tools or distribute server definitions across a team.

Sharing MCP servers through project scope

Claude Code connects to MCP servers through scopes, and project scope is the mechanism built for sharing a server across everyone working in a repository. Project-scoped servers live in a .mcp.json file at the repository root. Because .mcp.json is committed with the rest of the code, Claude Code reads those server definitions automatically whenever a developer opens the repository, giving the team access to the server's tools without each developer configuring it by hand.

Sharing a server through .mcp.json still keeps an explicit approval step in place. Before Claude Code uses any project-scoped MCP server, it asks each developer to approve that server.

User scope, set with --scope user, behaves differently. It stores the MCP server in one developer's own configuration instead of in the repository, so the entry stays local to that developer and is not shared with anyone else. The difference between the two scopes is where the configuration lives, and only project scope travels with the repository.

Remote HTTP servers versus local stdio servers

Project-scoped entries in .mcp.json fall into two types depending on how the MCP server runs and communicates:

  • Remote servers: A remote MCP server uses the HTTP transport and specifies the server's URL.
  • Local servers: A local MCP server, such as a Python script, uses the stdio transport and lists the command and arguments that launch it.

You can write either project-scoped entry to .mcp.json using the matching transport flag with claude mcp add:

# Add a project-scoped remote MCP server over HTTP
claude mcp add --transport http --scope project

# Add a project-scoped local MCP server over stdio
claude mcp add --transport stdio --scope project

In both cases, --scope project writes the configuration to .mcp.json at the repository root so it can be committed with the codebase. The transport flag matches how the server is reached, using http for a remote server URL and stdio for a locally launched command.

How CLAUDE.md and Agent Registry fit in

Two other pieces often come up alongside Claude Code configuration, and each has its own job:

  • CLAUDE.md: Holds instructions and project context for Claude Code. It shapes how Claude Code works in a repository, while MCP connections come from the scoped configuration described above.
  • Agent Registry: A central Google Cloud catalog of agents, MCP servers, and tools. Claude Code takes its MCP servers from its own configuration rather than from Agent Registry.

Keeping the layers apart makes Claude Code setups easier to reason about. CLAUDE_CODE_USE_VERTEX picks the model provider, scoped configuration connects MCP servers, and CLAUDE.md carries instructions.

Our Professional Agentic Architect practice exams test Claude Code configuration on Agent Platform alongside MCP server scoping and Agent Registry boundaries, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow