Wrong Reasons to Build an MCP Server (and What to Do Instead)

Daniel Oh, CNCF Ambassador and Java/Cloud AI lead at IBM, makes the contrarian case at the Agentic AI Foundation Summit: most teams shipping MCP servers do not need one. He walks through the five wrong reasons engineers reach for MCP, the mega-server architecture mistake almost everyone makes in month six, and the five-point checklist that tells you whether MCP is actually the right call for your agent. If you are early on your agentic AI journey, this is the talk that saves you from a year of token bloat, cascading failures, and 10x latency you did not budget for. Wrong Reason 1, Single-Consumer Trap: when one agent talks to one tool, function calling is faster, simpler, and cheaper than MCP Wrong Reason 2, Performance Overhead: a direct call lands in under 20ms. The same call through MCP can run 10x slower across multi-hop networking, JSON-RPC, and added observability layers Wrong Reason 3, Security Bloat: MCP is a transport layer, not a security layer. Cramming auth, rate limits, and zero trust into the MCP server itself makes it heavier and slower Wrong Reason 4, Garbage In, Garbage MCP: if your REST API has a messy JSON schema, exposing it through MCP only amplifies the mess Wrong Reason 5, Already Have a Great API: if your team already knows the path to the data and the API is well documented, MCP is overhead, not value Mega-Server vs Micro-MCP: why the giant all-in-one MCP server is the new monolith, and how to apply microservice thinking to break it apart Token Math That Hurts: a direct function call costs under 50 tokens. A simple MCP call costs 500 to 1,500. A complex one costs over 2,000. Before your user types anything The 5-Point MCP Checklist: distribution, discoverability, multi-team delivery, context-window pressure, and clean API design. Three or more yeses means go. Fewer means rethink Live Demo: a help-desk agent built on Quarkus calling a custom Postgres-backed MCP server, end to end Plan the Sunset: why every MCP server you build should ship with a retirement plan If you are an engineering lead, architect, or platform team thinking about adopting MCP in 2026, watch this before your next sprint. Links and Resources: Agentic AI Foundation: https://agenticaifoundation.com Daniel Oh on GitHub: https://github.com/danieloh30 Quarkus framework: https://quarkus.io Model Context Protocol: https://modelcontextprotocol.io CNCF: https://www.cncf.io Timestamps (approximate, may need adjustment): 00:00 Intro: Daniel Oh, IBM, CNCF 01:00 The MCP gold rush: 5,000+ servers and counting 02:30 Why standardization is not the same as enterprise-ready 04:00 Wrong Reason 1: the single-consumer trap 06:30 Wrong Reason 2: latency and the 10x overhead 09:00 Wrong Reason 3: security does not belong in the transport layer 12:00 Wrong Reason 4: garbage API schemas turn into garbage MCP 14:30 Wrong Reason 5: you already have a great REST API 17:00 Mega-server anti-pattern and the move to micro-MCP 19:30 Token math: 50 vs 1,500 vs 2,000 22:00 The 5-point MCP checklist 24:00 Live demo: Quarkus, Postgres, and a help-desk agent 26:30 Plan the sunset, close, and Q&A #MCP #AgenticAI #ModelContextProtocol