InfoQ Architecture·April 1, 2026

Pinterest's Model Context Protocol Ecosystem for AI Agent Workflows

Pinterest engineered and deployed a Model Context Protocol (MCP) ecosystem to standardize, secure, and scale AI agent workflows, moving from ad-hoc integrations to a unified client-server mechanism. This architecture enables language models to call tools and access structured data across diverse internal systems, significantly boosting developer productivity by automating complex engineering tasks. The design emphasizes domain-specific MCP servers, a central registry for governance, and a human-in-the-loop approval mechanism for sensitive operations.

AI & ML Infrastructure Distributed Systems Security

Read original on InfoQ Architecture

Introduction to the Model Context Protocol (MCP)

Pinterest developed the Model Context Protocol (MCP) as a foundational ecosystem for integrating AI agents into engineering workflows. This system replaces fragmented, ad-hoc integrations with a standardized, scalable, and secure substrate for AI tool-calling. The core idea is to enable large language models (LLMs) to interact with internal tools and structured data through a unified client-server interface, allowing agents to perform tasks like log analysis or bug ticket inspection directly within live systems.

Architectural Design and Components

The MCP architecture is built around several key components designed for flexibility, security, and scalability:

Cloud-hosted MCP Servers: These are domain-specific, meaning separate servers exist for areas like Presto, Spark, or Airflow. This design choice prevents context bloat, isolates tools, and facilitates fine-grained access control, avoiding a monolithic service.
Central MCP Registry: Acts as the single source of truth for all approved MCP servers and their metadata. It provides both a human-friendly UI and an API for discovery, validation, and integration by internal AI clients and IDEs. Clients query the registry to check permissions and server status before making tool calls, ensuring consistent governance.
Unified Deployment Pipeline: Manages the infrastructure, scaling, and lifecycle of MCP services, allowing engineers to focus on defining tool interfaces rather than operational overhead.
AI Agent Integrations: Seamlessly integrate with developer tools, chat platforms, and IDEs, enabling agents to discover and utilize MCP servers via the registry for tasks like incident investigation or change proposal.

Security and Governance Mechanisms

Given that MCP servers can perform automated actions on live systems, robust security and governance are paramount. Pinterest implemented a multi-layered approach:

Human-in-the-Loop Approval: For sensitive operations, agents propose changes or actions, but human users must explicitly approve or reject them before execution. This prevents potentially dangerous automated actions.
Two-layer Authorization Model: End-user JWTs (JSON Web Tokens) secure human-initiated access, while service-only flows rely on mesh identities. This distinguishes between user and service contexts.
Fine-grained Authorization: MCP servers employ decorators and business-group gating to restrict high-privilege operations to approved teams. Existing internal OAuth flows are leveraged for authentication, providing full auditability.
Security Standard Compliance: Every MCP server must adhere to a defined MCP Security Standard and undergo thorough review for security, legal/privacy, and GenAI requirements before being deployed to production.

💡

System Design Takeaway

Designing a system that allows AI agents to interact with production infrastructure requires careful consideration of security, access control, and auditability. The human-in-the-loop mechanism is a critical safety net for preventing unintended consequences of autonomous actions.

Impact and Future Directions

The MCP ecosystem has demonstrated significant impact, with a north-star metric of time saved. As of early 2025, it reported 66,000 invocations per month across 844 active users, saving approximately 7,000 hours monthly. Pinterest continues to expand the fleet of MCP servers, deepen integrations, and refine governance, highlighting the value of a structured approach to enterprise AI automation.

AI AgentsLLM IntegrationAPI DesignMicroservicesCloud ArchitectureSecurityWorkflow AutomationPlatform Engineering

Comments

Loading comments...

Architecture Design

Design this yourself

Design a Model Context Protocol (MCP) ecosystem to enable AI agents to automate complex internal engineering tasks at a large tech company. Your design should include domain-specific service orchestration, a central registry for tool discovery and governance, a robust multi-layered authorization model with human-in-the-loop approval for sensitive operations, and a scalable deployment pipeline for new agent integrations. Emphasize how to achieve security, auditability, and consistent access control across diverse internal tools and data sources.

Practice Interview

Other design angles

· Design a secure API gateway specifically tailored for AI agent interactions, focusing on dynamic routing, policy enforcement, and contextual authorization based on agent identity and tool capabilities.· Design a platform for building and deploying internal AI agents that can interact with various enterprise systems. Focus on the developer experience for defining tool interfaces, managing dependencies, and deploying agents, while ensuring adherence to security and compliance standards.· Architect a distributed authorization system for AI agent workflows that can enforce fine-grained access control at the method or resource level, integrating with existing identity providers and supporting both user-initiated and service-to-service access patterns.