It support · Production

Pinterest builds a production MCP ecosystem with 66,000 monthly invocations saving 7,000 hours per month

The problem

Pinterest needed a unified substrate for AI agents to access internal tools and data sources, replacing bespoke one-off integrations for every model and tool. Additionally, spinning up new MCP servers required excessive operational work before teams could write any business logic.

Workflow diagram · grounded in source
1
User initiates via surface
trigger
“A user interacts with a surface like our web AI chat interface, an IDE plugin, or an AI bot.”
2
OAuth and JWT authentication
validation
“The client performs an OAuth flow against our internal auth stack and sends the resulting JWT when it connects to the MCP registry and the target MCP server.”
3
Registry permission check
routing
“lets internal services ask "Is this user allowed to use server X?" before letting an agent call into it”
4
AI agent binds and invokes tools
ai_action
“our AI chat agent binds MCP tools directly into its agent toolset so invoking MCP feels no different from calling any other tool”
5
Human approval gate
human_review
“agents propose actions using MCP tools, and humans approve or reject (optionally in batches) before execution. We also use elicitation to confirm dangerous actions.”
6
Tool executes and returns results
output
“Presto tools let agents (including AI-enabled IDEs) pull Presto-backed data on demand so agents can bring data directly into their workflows instead of context-switching into dashboards”
Reported outcome

Pinterest's MCP ecosystem reached 66,000 invocations per month across 844 monthly active users, saving an estimated 7,000 hours per month for engineers.

Reported metrics
monthly MCP invocations66,000 invocations per month
Monthly active users844 monthly active users
Hours saved per month7,000 hours per month
Reported stack
Model Context Protocol (MCP)PrestoSparkAirflowEnvoySPIFFE
Source
https://medium.com/pinterest-engineering/building-an-mcp-ecosystem-at-pinterest-d881eb4c16f1
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Pinterest's MCP ecosystem reached 66,000 invocations per month across 844 monthly active users, saving an estimated 7,000 hours per month for engineers.

What tools did this team use?

Model Context Protocol (MCP), Presto, Spark, Airflow, Envoy, SPIFFE.

What results were reported?

monthly MCP invocations: 66,000 invocations per month; Monthly active users: 844 monthly active users; Hours saved per month: 7,000 hours per month (source-reported, not independently verified).

How is this it support AI workflow structured?

User initiates via surface → OAuth and JWT authentication → Registry permission check → AI agent binds and invokes tools → Human approval gate → Tool executes and returns results.

Related it support cases