Remote MCP Servers

While stdio servers run locally, Remote MCP Servers are cloud-hosted endpoints that any authorized client can connect to over the internet. Anthropic's 2025 specification formalizes these as OAuth 2.1-secured HTTP endpoints.

How Remote MCP Works

1. The MCP Server is deployed as an HTTP service (e.g., on AWS, Vercel, or Cloudflare).
2. The Client discovers the server's capabilities via a /.well-known/mcp manifest.
   • Authentication uses standard OAuth 2.1 with PKCE (Proof Key for Code Exchange) and DCR (Dynamic Client Registration) to smoothly onboard new agent clients without manual admin intervention.
   • Communication uses Streamable HTTP with optional SSE for real-time push events.

   // Remote MCP Server manifest (/.well-known/mcp)
   {
     "name": "acme-crm",
     "version": "2.0.0",
     "endpoint": "https://mcp.acme.com/v1",
     "auth": {
       "type": "oauth2",
       "authorization_url": "https://auth.acme.com/authorize",
       "token_url": "https://auth.acme.com/token",
       "scopes": ["read:contacts", "write:deals"]
     }
   }

   MCP Connector

   MCP Connector is Anthropic's first-party integration that lets Claude connect to remote MCP servers directly via the API - no Host application needed!

   // Using MCP Connector in the Messages API:
   {
     "model": "claude-sonnet-5",
     "mcp_servers": [{
       "type": "url",
       "url": "https://mcp.acme.com/v1",
       "authorization_token": "Bearer eyJ..."
     }],
     "messages": [...]
   }

   💡 Key Insight: MCP Connector eliminates the need for client-side MCP infrastructure. You just pass server URLs in your API call, and Claude handles the MCP handshake, tool discovery, and execution automatically.

   Tool Search & Discovery

   When connecting to many MCP servers with hundreds of tools, Claude's Tool Search automatically discovers the most relevant tools for each request - saving tokens and improving accuracy.

   Instead of loading all 200 tools into context, Tool Search indexes your catalog server-side and injects only the 5-10 tools relevant to the current query.

   Fine-Grained Tool Streaming

   Standard streaming returns text tokens. Fine-grained tool streaming streams individual tool input fields as they're generated - enabling real-time UI previews of tool arguments before execution completes.