Skip to content
Home › Tool Design & MCP Integration

MCP Connector

Tool Design & MCP Integration

Definition

The MCP connector lets the Claude API call tools on a remote Model Context Protocol server directly, without you running a separate MCP client.

Key points

  • Beta header: anthropic-beta: mcp-client-2025-11-20 (previous mcp-client-2025-04-04 is deprecated — config moved out of mcp_servers into tools[] as an MCPToolset).
  • Only tool calls are supported (not the full MCP spec — no prompts/resources).
  • Server must be publicly exposed over HTTP: Streamable HTTP + SSE transports. Local STDIO is not supported by the API connector.
  • Not ZDR-eligible.
  • Availability: Claude API, Claude Platform on AWS, Microsoft Foundry ("Hosted on Anthropic" deployment). NOT on Amazon Bedrock or Google Cloud/Vertex.
  • Two components: mcp_servers array (connection) + mcp_toolset entries in the tools array (tool config) — see MCPToolset Allowlist and Denylist.
  • mcp_servers entry fields: type (only "url"), url (must start https://), name (unique; referenced by exactly one MCPToolset), optional authorization_token (OAuth Bearer; caller handles flow + refresh).
  • Response blocks: mcp_tool_use (id, name, server_name, input) and mcp_tool_result (tool_use_id, is_error, content).
  • Works in the Message Batches API at the same pricing.

Why it matters for the exam

  • Domain 2 tests the beta header string, the HTTP-only transport rule, and the platform availability matrix.
  • ZDR ineligibility contrasts with prompt caching (which qualifies).

Common gotchas

  • STDIO is not supported by the API connector — this is a favourite distractor.
  • Not available on Bedrock or Vertex.
  • url must be https://; not ZDR-eligible.
  • This is the API connector; Claude Code's MCP config uses files — see MCP Config Scopes.

See also

Sources

Referenced by