Upgrade to Pro — share decks privately, control downloads, hide ads and more …

The Path Paved by Flava MCP Hub Piping Work for...

The Path Paved by Flava MCP Hub Piping Work for the AI ​​Era

As AI agent adoption accelerates, MCP servers have emerged across the company, but each team has handled authentication, security, and connection patterns differently. In this session, we share how Flava MCP Hub solved this fragmentation as a platform.

We will cover the four-layer architecture built through operating MCP Hub, the Connector that bridges the protocol gap between internal authentication and the MCP standard, and lessons learned from efforts such as bringing the Gateway in-house. Based on these experiences, we will also share where Flava MCP Hub is headed next.

Recommended For

Anyone who wants to safely connect AI Agents with internal systems
Anyone thinking about how to bridge the gap between standard protocols and internal authentication systems
Anyone interested in real-world MCP operations, including Gateway, OAuth, and Monitoring
In One Line

A practical session on how we designed and operated authentication, routing, authorization, monitoring, and more to turn MCP into company-wide AI infrastructure.

More Decks by LINEヤフーTech (LY Corporation Tech)

Other Decks in Technology

Transcript

  1. 2026.06.29 LY Corporation Minhyeok Park | Cloud AI Platform Inbae

    Kang | Cloud AI Platform The Path Paved by Flava MCP Hub Piping Work for the AI Era
  2. Minhyeok Park Tech Lead of Flava MCP hub Cloud Engineer

    at LINE Plus Corporation Inbae Kang Backend engineer of Flava MCP hub Cloud Engineer at LINE Plus Corporation A brief introduction About Us
  3. Agenda 01 Why Flava MCP hub? 02 4-layers plumbing model

    03 Bridging the protocol gap 04 Building Our Own Gateway 05 Values & Roadmap
  4. Why Flava MCP hub? What is the MCP External Service

    Examples Internal Data/Platform MCP
  5. Why Flava MCP hub? What is the MCP The MCP

    platform for all of LY Corporation
  6. Why Flava MCP hub? See how simple it really is

    Deploy Kubernetes MCP Server and lookup workloads Search Tech Week 2026 Schedule using Confluence MCP Demo 01 Demo 02 1 min 1 min
  7. The Problem of MCP Use The hidden walls in the

    AI agent era 01 MCP Servers Sprawling Out of Control 02 Auth Chaos Across Every Client & Server 03 Internal Auth Can't Keep Up 04 No Visibility and Insight
  8. 4-layer plumbing model How can we solve the problem Client

    Gateway Connector MCP Server 01 Connection Connector handles Client‒Hub auth automatically via stdio+OIDC 02 Proxy Gateway enforces policies and standardizes the protocol 03 Authorization OAuth / OIDC middleware validates permissions on every request 04 Monitoring Logs, audits, and collects usage patterns across the entire request flow
  9. L1 Connection: MCP Connector Connector handles Client‒Hub auth transparently Connector

    owns the full OIDC flow. Client stays completely auth-unaware. Client Gateway Connector MCP Server
  10. L2 Proxy: MCP Gateway Where policy meets protocol Deploy an

    MCP. Gateway appears automatically ̶ enforcing auth, CORS, header transforms, and transport protocol under one standard. Client Gateway Connector MCP Server
  11. L3 Authorization: OAuth/OIDC Middleware Centralized auth validation at the Gateway

    Connector issues the token. Gateway middleware validates all auth types in one place. Client Gateway Connector MCP Server
  12. L4 Monitoring From blind spots to full observability One pipeline

    captures every call, auth result, and latency ̶ turning usage patterns into design decisions. Client Gateway Connector MCP Server
  13. Protocol Gap Follow the spec or fit the system ̶

    you couldn't do both Protocol Gap Internal Auth MCP Spec
  14. Protocol Gap Follow the spec or fit the system ̶

    you couldn't do both Client MCP Server Protocol Gap
  15. Bridging the Protocol Gap stdio proxy that absorbs all auth

    complexity behind the scenes Connector stdio Connector proxies over stdio ̶ so auth complexity never reaches the client. Authenticated Client MCP Server
  16. Bridging the Protocol Gap stdio proxy that absorbs all auth

    complexity behind the scenes 1 Week 5 minutes Build the logic. Flava MCP Hub handles the rest.
  17. • Only 2 Months • Avoid reinventing the wheel •

    Focus on internal requirements Why We Chose an OSS Project Building Our Own Gateway
  18. Operational Gaps We Could Not Ignore Building Our Own Gateway

    JWKS Compatibility Gap Middleware Ordering Issue Memory Instability © LY Corporation ES512 Not Supported OOM occurred CORS & Auth
  19. What We Improved Building Our Own Gateway • Design by

    internal requirements • Full control over core behavior • Easier troubleshooting • Reduced patch burden • Improved stability and operability
  20. AI reduces migration effort OSS helps initial phase Rebuilding can

    improve stability Lessons Learned Building Our Own Gateway
  21. What the Plumbing Showed Us Once the platform plumbing was

    in place, MCP Hub became more than a connection layer.
  22. • Remove Duplicate Work • Standardize Authentication • Unify Governance

    • Provide Platform Support What the Plumbing Was Built For What the Plumbing Showed Us
  23. What the Plumbing Showed Us What the Plumbing Revealed Tool

    Usage Visibility Cross-Team Reuse Feedback Loop Observation point of AI tool usage
  24. The Road Ahead Automating more of the auth, tooling, and

    sharing flows OpenAPI → MCP Internal Auth MCP 2026-07-28 CLI & Skills ID-JAG
  25. Turning existing REST APIs into MCP servers The Road Ahead

    Existing APIs are not directly accessible by MCP clients Team Existing APIs MCP client
  26. Turning existing REST APIs into MCP servers The Road Ahead

    OpenAPI specs become MCP tools No custom MCP server required Existing APIs MCP hub MCP client
  27. Automating token exchange for secure MCP access The Road Ahead

    Token for A Token for B MCP client MCP server A MCP server B
  28. Automating token exchange for secure MCP access The Road Ahead

    MCP client MCP server A MCP server B ID-JAG Token Token for A Token for B
  29. Standardize auth & governance Remove duplicate work Turn usage into

    platform feedback Teams build and reuse MCPs faster Flava MCP Hub handles the plumbing
  30. Trademark Notice Slack is a trademark and service mark of

    Salesforce, Inc. Jira and Confluence are trademarks of Atlassian Pty Ltd. Model Context Protocol, MCP, and the MCP logo are trademarks or properties of their respective owners. All other trademarks, service marks, logos, and trade names are the property of their respective owners.