[mercari GEARS 2025] Backend Standardization with MCP

Transcript

1. Backend Standardization with MCP
 Katie Hockman
 Mercari Platform Software Engineer
 Tae Hoshino 
 Mercari Tech PM

2. What we do at Mercari

3. Worked on the Go team at Google and as an Engineering Manager at Datadog before joining Mercari in May 2024. Working on core foundational services that support the company..
 @Katie
 Platform Software Engineer
 
 @tae 
 
 Platform Product Tech PM
 
 
 Worked for Dropbox and Postman before joining Mercari in January 2025.
 Leading cross-organizational projects including backend standardization for the entire Mercari group. 

4. Challenges with Standardization

5. Problems with Lack of Standardization
 Domain Siloes
 Each engineering team develops differently
 Inconsistent Structure
 Maintenance Cost
 Different coding patterns used in different services
 Each team needs to maintain their own assets beyond business logic

6. Real-World Example at Mercari
 • Inconsistent logging and tag standards, along with observability gaps, hinder efficient cross-service debugging.
 • Service ownership changes following re-organization lead to substantial onboarding costs.
 • Teams spend hours or days in alignment discussions debating what libraries or patterns they should use

7. Why Standardization Matters
 Faster Onboarding
 
 Learn once, code anywhere
 Reduced Cognitive Load
 Focus on what brings business impact
 Consistency
 Enable AI driven development flows across service boundaries 

8. Why is standardization hard?
 1. Significant Cost: Requires lots of manual code changes (1 FTE +7days) and testing (QA) to ensure nothing breaks
 2. No Clear ROI: Hard to justify investment or quantify benefits 
 3. Cold Start Problem: Certain benefits can only be realized after majority has adopted (Network Effect)

9. Paradigm Shift with AI

10. 84% Programmers are currently using, or plan to use, AI for coding this year
 
 According to
 StackOverflow 2025 Developer Survey 

11. Mercari’s Shift to AI-Native
 Bold investments to AI: • AI Taskforce • Group wide access to AI tools (e.g. Cursor, Devin, Claude Code, Gemini, ChatGPT, n8n, dify, etc) • AI Hackathon and Knowledge Sharing (Lightning talks, Demo days, etc) Image(top): https://www.watch.impress.co.jp/docs/news/2037080.html Image(bottom): https://forbesjapan.com/articles/detail/82714
12. Platform’s AI Hackathon
 • Held 2 day Platform Hackathon in June
 • Weekly FriAI on Fridays! 
 
 → Nurture the culture of exploration and sharing learnings

13. Why AI is a game changer
 
 AI in Development Workflows
 • AI as a coding assistant: “Add a new function to do x…”
 • AI as an autonomous agent: “Help me debug this problem…”
 • AI as knowledge bridge: “Tell me about this code…”
 
 → Game changer for driving standardization
 • Autonomous agent understand company standard and migration steps
 • Knowledge bridge to help in onboarding
 
 
 

14. Intro to MCP

15. AI Tooling Limitations
 Inaccurate
 AI can hallucinate or give inaccurate data, creating a worse result
 Too General
 The information may be too broad to be useful, missing domain-specific context
 Limited Power
 It can only perform a small subset of actions such as searches or terminal commands

16. Logos are trademarks of their respective owners.
17. Enter MCP Servers!
 Logos are trademarks of their respective owners.
18. What does the AI Agent need?
 • Relevant context
 • Powerful tooling

19. Introducing the Model Context Protocol (MCP)
 “MCP is an open protocol that standardizes how applications provide context to LLMs.”
 Source: https://modelcontextprotocol.io/introduction

20. In simple terms…
 
 MCP is a protocol for AI models to interact with external tools.

21. Logos are trademarks of their respective owners.
22. Logos are trademarks of their respective owners.
23. MCP in
 Developer Environments

24. Logos are trademarks of their respective owners.
25. Logos are trademarks of their respective owners.
26. Logos are trademarks of their respective owners.
27. Logos are trademarks of their respective owners.
28. LLMs empower users to perform actions using human language
 MCP Servers empower LLMs to be more accurate and targeted

29. Demo

30. Where can AI tools go wrong?
 • Struggles with large, spread out codebases
 • AI may not choose the most efficient strategy
 • Strength of the AI model plays a huge role

31. Looking to the Future

32. New Challenges
 
 
 • Implicit knowledge to Explicit knowledge
 • Enable AI to autonomously take actions while ensuring security and quality
 →Complex, cross-domain problems require us to collaborate effectively 
 → AI helps us to step out of comfort zone
 

33. Wrap Up

34. Conclusion
 Context is Key
 
 • Feeding AI organizational context
 • MCP is one approach
 Win-Win
 AI-Native Culture
 • AI can scale standards
 • Standards enable AI
 • Experiment
 • Learn from failures
 • Share findings

35. @Katie
 Platform Software Engineer
 
 @tae 
 
 Platform Product Tech PM
 
 
 Thank you!