← Blog

August 20, 2026 #ai #automation

Everyone's talking about MCP. I keep reaching for the API.

Over the last couple of months I’ve connected a lot of tools to my AI setup. Shopify, Stripe, Klaviyo, Circle, Amazon, Google Analytics, QuickBooks. In every case, I had the option of going the MCP or API route.

MCP is getting a lot of hype right now, and for good reasons. You can connect a service in seconds with no code, the AI can use it live in the conversation, and new connectors are showing up for just about everything. It’s all over my feeds. And yet, more often than not, I ended up walking through the plain old API door instead.

Every time, I’d start with the MCP because it was the quick, shiny option. Partway in I’d hit a wall, or just notice the API route gave me more of the service to work with and more room to change things later. So I’d switch.

As someone still understanding the nuance of working with both, here’s what I learned:

MCP (Model Context Protocol) plugs a service straight into the AI chat, so the model can use it live while we’re talking. Think of it as handing the assistant a remote control for an app. Flip it on, no code, and it just works in the conversation.

An API (Application Programming Interface) connection is a little script that goes and fetches the data on command. It gets written once. After that it runs on its own, no person and no AI needed.

You can schedule either one. The difference is that a scheduled MCP still has to wake the AI up to run. A script is just code, it runs on its own.

Both can be automated, but what happens at run time is completely different. My API scripts run on a scheduler on my Mac with no AI in the loop at all. The AI shows up once, to write the script, and after that it’s plain code firing in the background (I could even move it onto an always-on server so my laptop could be shut). An MCP runs every request back through the AI, so it costs tokens every single time, and just being connected loads its instructions into the model’s memory before I’ve done anything. For a report I’ll run every month for years, that gap adds up.

The token savings were just a bonus. The real driver was that the API almost always gave me access to more of the source’s data than the MCP did.

Take the monthly report I keep. It pulls key numbers from Shopify, Stripe, Amazon (which has no MCP at all), Google Analytics, Search Console, and QuickBooks into one Google Sheet, every month, without me babysitting it. Done through MCPs, that job would mean running an AI session every month, hoping each connection behaves, paying tokens for every pull, and settling for whatever slice of each tool’s data the connector exposes. As scripts it just happens on a schedule, the same way every time, for free, pulling exactly the fields I asked for.

Here’s how the rest shook out, tool by tool:

  • Circle I started on MCP, but it was less accurate than going through the API, and there were certain things I just couldn’t count on it for. Since switching to the API it’s been much more reliable.
  • Klaviyo stayed on MCP. It’s great when I want to poke at email campaigns mid-conversation and ask follow-up questions.
  • Slack, Gmail, Drive, and Calendar are all MCP too, and they’re a joy, because I’m always right there in the chat when I use them.

For me, choosing between the two has come down to what the job needs.

MCP (the cockpit)API connection (the autopilot)
Best forSitting with the AI, exploring, asking follow-upsA job that has to run on a schedule, alone
SetupFlip it on, no codeWrite a small script once (minutes for the AI to code)
Needs Claude open to run?Yes, it works only inside a live chatNo, it runs on its own in the background
Token costSpends tokens every time it’s usedFree to run once it’s built
When it breaksYou wait on whoever built itYou can open it up and fix it yourself
Picture it asA power tool in your handA machine on the factory floor

These days I reach for an MCP when I want to explore something with the AI right there in the conversation, and I write a small API script for anything that needs to run on its own every week or month.