Why it matters
MCP is the layer that gets the most attention in the agent-tooling conversation right now, but Emmanuel’s consistent position is that it’s a chapter, not the title: “MCP wraps your API. CLI wraps your API. Skills enhance your API. OpenAPI documents your API. Get the API wrong and nothing above it works. Get it right and everything follows.” Building an MCP server on top of a poorly designed API doesn’t fix the underlying problem — it just gives agents a new, faster way to hit the same bad naming, the same unhelpful errors, and the same missing list endpoints.
That said, MCP servers matter because they’re increasingly how agents discover what a product can do without a human writing custom integration code first. A well-built MCP server, backed by an AI-ready API, is one of the fastest ways to make a product genuinely usable by Claude, ChatGPT, and other agent platforms.
How to apply it
- Get the underlying API to AI-ready status first (see the AI-Ready API Checklist) — an MCP server inherits every problem in the API it wraps.
- Design each MCP tool around an outcome an agent needs, not a 1:1 mirror of every internal endpoint.
- Write tool descriptions the way you’d write API descriptions for the checklist: specific, accurate, and matched to what the tool actually does.
- Test the server the way an agent will use it — cold, without documentation, retrying from errors — not just with a known-good client.
Learn the mechanics hands-on in Build AI-Ready API Products With MCP.
FAQ
What's the difference between an API and an MCP server?
An API is the underlying interface your product exposes over the network. An MCP server is a standardized wrapper that makes a chosen set of that API's capabilities discoverable and callable by AI agents through the Model Context Protocol.
Do I need to rebuild my API to add an MCP server?
No, but you may need to fix it first. An MCP server built on top of a confusing or poorly documented API will just expose those same problems to agents faster.
Can an MCP server replace my REST API?
Generally no. Most teams keep the API as the system of record and add an MCP server as one more consumption layer for agents, alongside existing REST/SDK access for developers.
How is MCP different from just publishing an OpenAPI spec?
OpenAPI documents an API for developers and tools; MCP is a live, callable protocol an agent uses to invoke your tools directly at runtime. They're complementary — Emmanuel ranks SDKs and MCP-style direct calling above raw OpenAPI or documentation for agent usability.
Where do I see a real MCP server in action?
Emmanuel's own API Lessons content ships as a live MCP server students can install and query directly — see the course resources in Build AI-Ready API Products With MCP.