KOEN
Read Original
#AI

Building an MCP Ecosystem at Pinterest

Building an MCP Ecosystem at Pinterest
01

Summary

How Pinterest Standardized the 'Toolbelt' for AI Agents

Cutting 7,000 Engineering Hours Monthly with a Scalable MCP-Based Infrastructure

This article explores how Pinterest addressed fragmented AI tool integrations by adopting the Model Context Protocol (MCP) enterprise-wide. It details their journey in building a secure, governed, and efficient AI agent ecosystem through a central registry and robust auth models.

  • 01Adoption of domain-specific micro-MCP servers
  • 02Two-layer security using JWT and SPIFFE identities
  • 03Centralized discovery and governance via the MCP Registry
  • 04Successful integration with high-traffic systems like Presto and Spark
  • 05Ensuring safety through Human-in-the-loop design for sensitive actions

RECOMMENDATION

Essential reading for platform engineers looking to securely connect LLMs to internal infrastructure or teams building enterprise-grade AI agents.

The Problem

Pinterest faced challenges in connecting LLMs to diverse internal tools and data sources due to the need for bespoke, one-off integrations and complex security management.

The Solution

They implemented an ecosystem based on the open-source Model Context Protocol (MCP), featuring cloud-hosted domain-specific servers and a central registry for discovery and governance.

The Result

The system reached 66,000+ monthly invocations, saving approximately 7,000 engineering hours per month for over 800 active users.

Trade-off

Opting for cloud-hosted servers over local ones required building a unified deployment pipeline, and strict security reviews are mandated for all production-grade servers.

03

Key Concepts

Concept · 01

Model Context Protocol (MCP)

An open-source standard enabling LLMs to communicate with data sources and tools through a unified protocol.

  • Acts as the substrate for Pinterest's AI agents to automate engineering tasks safely.
Concept · 02

MCP Registry

A central source of truth for approved MCP servers, providing discovery UI and authorization APIs.

  • Allows AI clients to validate servers and check user permissions before tool execution.
Concept · 03

Human-in-the-loop

A design pattern where a human must approve sensitive or high-cost actions proposed by an AI agent.

  • Mandated for data-altering operations to mitigate the risk of automated tool execution.