“Don’t build MCP servers from your APIs.”
I keep hearing this. I disagree.
You’ve spent years building your API infrastructure and process. Those APIs aren’t going away.
What people mean is: don’t blindly wrap your APIs in MCP and call it done. You’ll end up with a chatty, bloated server that agents struggle with.
Fair. But you have to start somewhere.
Generate your first MCP server from your existing APIs. Try it out. Eval it against real prompts. Iterate until you hit a high success rate.
Along the way you’ll evolve & redesign your APIs. Or build new ones for agent use cases.
This serves double duty:
Good APIs for coding agents like Claude Code that integrate directly, low token overhead.
Better MCP server for chatbots like ChatGPT, where users discover your product through conversation.
You may drop some tools. Split the server into two or three. That’s good for agents and probably for your human developers too.
Three key things:
- Build an Eval harness you can run with many models
- Have a set of Golden Prompts that match the use cases you’re going after
- Track task success rates across iterations.
Continuous Evals are what makes a highly usable MCP product, not ditching your existing APIs.
If you want to get started down this path, I’ll show you how. Join me this Wednesday through Friday for @Build AI-Ready API Products with MCP: