Is Your MCP Server Actually Good? FastMCP's Lead Maintainer Explains

Prefect
321 views •

Most MCP servers work in a demo and fall apart with a real agent. Nate Nowack (FastMCP maintainer) and Wayne Grigsby (Lead Solutions Engineer at Prefect) break down what's new in FastMCP v4, when NOT to write an MCP server, why naively mapping a REST API to MCP tools creates massive token bloat, and how middleware, evals, and auth actually work in production MCP servers. What we cover: - What's new in FastMCP v4 (the stateless protocol change) - Why MCP exists at all and when a CLI or REST API is the better choice - Token-bloat: mapping REST endpoints 1:1 to MCP tools vs. curating them - Middleware for filtering context, restricting access, and handling auth - Tool bloat and "code mode" (search + execute instead of dozens of tools) - Plugins: bundling an MCP server with skills for distribution - Evals: testing whether an agent with your MCP server actually gets the right outcome - Deployment: client proxies vs. standalone agents, and locking down a public MCP server - Live demo: adding GitHub OAuth authentication to an MCP server Timestamps: 0:00 - What's new in FastMCP v4 (stateless protocol) 2:27 - What "stateless" actually means for MCP servers 4:19 - Why use MCP at all instead of just skills 6:19 - MCP is to an agent what a website is to a human 7:47 - "Isn't this just REST APIs with more steps?" 9:34 - When NOT to write an MCP server 11:05 - CLI vs. MCP: when to add a curation layer on top 13:36 - Don't map your REST API directly to MCP (live demo) 16:18 - The token-bloat problem, measured 19:03 - Middleware: filtering context and restricting access 21:58 - Middleware for auth: hosted servers without stored credentials 26:13 - Tool bloat and code mode (search + execute) 29:58 - Demo: visualizing your agent's context/token usage 31:33 - Plugins: bundling an MCP server with skills 37:22 - Evals: testing what your MCP server should get right 43:14 - Running evals in CI on every pull request 45:19 - Deployment: client proxies vs. standalone agents 46:55 - Locking down a publicly deployed MCP server 47:36 - Demo: a Pydantic AI agent using a hosted MCP toolset 50:12 - Demo: adding GitHub OAuth authentication Resources: FastMCP on GitHub: https://github.com/PrefectHQ/fastmcp FastMCP docs: https://gofastmcp.com If this was useful, subscribe! --- X: @PrefectIO

Keyboard shortcuts

On. Switch them off if they collide with your assistive tools; ? still opens this sheet.

Go to

Press g then the letter.

  • gh Latest
  • gs Sources
  • gm Media
  • gv Videos
  • gp Podcasts
  • gc Calendar
  • gd Decoder
  • gz Dataviz
  • ga Datasets
  • gb Blog
  • gk Markets
  • gj Careers
  • gn Prompt Notebook

On this page

  • / Focus search, where there is one
  • t Back to top
  • ? This list
  • Esc Close