Leveraging MCP at Scale: A Case Study from Anthropic

Unlocking the Power of MCP: How Standardizing on Message Passing Can Revolutionize Your Organization's Integration Efforts

  • 1. John has spent 20 years building large scale systems and is currently a member of technical staff at Anthropic, focusing on tool calling and integration.
  • 2. Models only recently became good at calling tools, leading to a surge in excitement and rapid development of various services and custom endpoints.
  • 3. The initial euphoria led to integration chaos, as different services had their own ways of handling tool calling and authentication.
  • 4. Over time, these endpoints started to resemble the MCP (Model Call Protocol) specification, which is a JSON RPC for sending messages between providers of context and models.
  • 5. Anthropic decided to use MCP for everything involved in providing model context, standardizing on a single approach for engineers, making things faster and more efficient.
  • 6. MCP solves problems that users might not have encountered yet, such as dealing with different billing models, token limits, and usage tracking in integrations.
  • 7. Anthropic is seeing external remote MCP services emerging, which requires external network connectivity, authentication, and handling security concerns.
  • 8. The company introduced the MCP gateway, a single point of entry for engineers to connect to MCP with a simple 'connect to MCP' call that returns an MCP SDK client session.
  • 9. URL-based routing is used to route to external and internal servers, handling credential management automatically, providing a centralized place for rate limiting and observability.
  • 10. Client libraries have been created to connect to the MCP gateway with a simplified process of passing a URL, org ID, and account ID (with signed token authentication).
  • 11. The MCP SDK object allows new features to be added to the protocol and easily integrated by updating MCP packages internally.
  • 12. The same code can seamlessly connect to internal and external integrations when it comes to transports, with various transport options available based on organizational needs.
  • 13. Anthropic has set up a unified authentication model, handling OAuth at the gateway so that consumers don't have to worry about complexity in their apps.
  • 14. Adding MCP support to new services is as simple as possible, with packages available for multiple languages and operational simplicity through standardized message formats.
  • 15. Centralizing context for models provides a unified location for all the context that models are asking for and receiving, enabling better policy enforcement and content classification.
  • 16. MCP prompt injection attacks and model access to sensitive data are potential risks that can be mitigated by centralizing context and enforcing policy.
  • 17. Standardizing on something like MCP simplifies future development and allows organizations to focus on building features rather than plumbing infrastructure.
  • 18. Building 'pits of success' makes the right way to do things the easiest way, encouraging developers to naturally follow best practices.
  • 19. Centralizing shared problems at the appropriate layer enables teams to spend more time working on valuable and interesting business problems.
  • 20. MCP is essentially JSON streams, with implementation details being a minor concern; choosing something like MCP simplifies development for future selves.
  • 21. Anthropic encourages the audience to consider using MCP or another standard in their organizations, focusing on building pits of success and solving shared problems once for greater efficiency.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!