← back
Remote MCPs: What we learned from shipping — John Welsh, Anthropic
Takeaway
Inside a large org, every custom tool API converges on MCP — adopt it as the default context interface even before you need all features.
Summary
- Anthropic's internal integrations all converged on MCP because every custom tool endpoint kept reinventing get-tools, get-resources, auth, and elicitation.
- Distinguishes MCP's two halves: the JSON-RPC message spec (valuable, well-debated) vs the global transport standard (streamable HTTP, OAuth 2.1, session management — the painful interop layer).
- 'MCP for everything' is the right default for anything that provides model context — even if you don't use the full feature set, you'll grow into it.
- Shipping remote MCP at scale forced solutions for centralized auth, session resumption, and elicitation across heterogeneous clients.
mcpanthropicintegrations
Original description
We recently released remote MCP support for both claude.ai and the Anthropic API. This talk will cover architectural decisions we made in our implementation, remote MCP authentication, supporting engineers who are building out agentic AI tools, implementing custom internal transports, and whatever else we can fit into 18 minutes of your time. About John Welsh I'm John Welsh, a software engineer who's been building large scale systems for the past 20 years. I'm currently at Anthropic, where I've been building our public API and defining how internal systems communicate with MCP servers and other integrations. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter