There have been many articles about MCP, but most stop at protocol introductions or how many servers have been connected.
Pinterest's engineering team recently shared their MCP architecture running in a real production environment, with specific data and enough detail to warrant a close look.
Architecture choice: multiple dedicated servers, not a monolith
Pinterest did not build one giant MCP server to handle everything. Their approach: each core data system corresponds to an independent MCP server.
Currently in production:
- Presto server: data queries
- Spark server: job debugging and log analysis
- Airflow server: workflow management
- Knowledge server: internal documentation and knowledge base
Why not a unified server?
Their reasoning: multiple servers allow independent access control, so different teams can only access their corresponding tool domains, avoiding excessive permissions on a single server. At the same time, each server's toolset remains lean, preventing the model from getting tangled in a large set of irrelevant tools.
Registry: letting agents know what is available
Pinterest built a central registry as the authoritative source for which servers are approved.
Before calling a tool, the AI client checks the registry: does this server exist? Do I have permission to use it?
This solves a real problem: not all data tools should be open to every employee. For example, Presto (large-scale data queries) is only accessible via AI tools to members of specific business groups.
Security design details
Two layers of authentication:
- User JWT token: verifies who is operating (human-in-the-loop scenarios)
- Service mesh identity: service-to-service calls use mesh identity
They chose not to adopt the MCP standard OAuth flow, instead integrating with their existing internal authentication system — the reason being to reduce friction from requiring separate authorization for each server.
There is also a design called elicitation: before executing high-risk or high-cost operations, the agent must request confirmation from the human user; if the person does not approve, it does not execute. This strikes a balance between automation and security.
Numbers in production
Data as of January 2025:
| Metric | Number |
|---|---|
| Monthly calls | 66,000 |
| Active users | 844 |
| Estimated monthly time saved | ~7,000 engineer-hours |
7,000 hours per month — based on roughly 160 working hours per engineer per month, this equates to saving the workload of about 44 people. This is an estimate derived from summing the estimated time saved per tool call, not an exact figure, but the magnitude is informative.
Notably
One sentence from Pinterest's article stands out: this system is integrated into the IDE engineers use daily, internal chat platforms, and AI workflows — not a separate AI portal that engineers have to deliberately go to.
This is the key to whether enterprise AI deployment succeeds: the tools need to appear where workers work, not require them to go somewhere else.
Sources: Pinterest Deploys Production-Scale Model Context Protocol Ecosystem for AI Agent Workflows (InfoQ); CocoLoop, Building an MCP Ecosystem at Pinterest (Pinterest Engineering Blog / Medium)