Get in touch
Glossary

Agent Access Stack

The Agent Access Stack is the layered model for how an AI agent actually reaches a product: your API as the foundation, with MCP, CLI, and Skills as layers built on top of it, and OpenAPI/documentation describing all of it. Emmanuel Paraskakis teaches this stack, building on Aakash Gupta's public analysis of how companies like Stripe, Cloudflare, GitHub, and Shopify ship for AI agents, with a consistent conclusion: get the API layer wrong and every layer above it inherits the problem.

Why it matters

Emmanuel’s summary of the stack: “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.” Stripe is the example he points to most: “Stripe didn’t retrofit for agents. They just exposed what was already there. A solid API foundation.”

The practical implication is a sequencing decision, not just a technical one. Teams often ask “should we build an MCP server or a CLI first?” when the real first question is whether the underlying API is agent-ready at all. Every layer on top — MCP, CLI, Skills — inherits whatever naming, error design, and structural problems exist in the API. Fixing the foundation first is cheaper than fixing it four times, once per layer.

How to apply it

  • Start at the bottom: run the AI-Ready API Checklist against your core API before building anything above it.
  • Decide which upper layers you actually need — not every product needs all four (API, CLI, MCP, Skills); pick based on where your users’ agents will actually look.
  • Treat OpenAPI/documentation as a description layer, not a substitute for the other layers — it explains the stack, it doesn’t replace any part of it.
  • Audit the stack the way an agent would use it: cold, without prior knowledge of your product.

FAQ

Who came up with the Agent Access Stack?

Aakash Gupta mapped it out by analyzing 10 companies shipping for AI agents (Stripe, Cloudflare, GitHub, Shopify, and others). Emmanuel Paraskakis teaches the underlying principle — that the API is the irreducible foundation — as part of his API-first curriculum.

Do I need every layer of the stack?

No. Most products need a solid API first; whether you also need an MCP server, CLI, or Skill depends on where your specific users' agents are going to look for you.

What happens if I build MCP before fixing my API?

You end up exposing the same naming, error, and structural problems to agents through a new channel, which doesn't fix anything, it just multiplies the surface area of the same bugs.

Is OpenAPI part of the stack or separate from it?

It's part of the stack, as the documentation layer — it describes the API but doesn't itself let an agent take action the way an MCP server, CLI, or SDK does.

Where should I start if I'm building all of this from scratch?

With the API. Every layer above it, in every company Aakash Gupta analyzed, was a wrapper or enhancement on top of a working API foundation.