MCP, API and browser-use, explained

Most people use AI agents for two things: asking questions and building stuff. The bigger shift is that agents can now operate your other software. Your email platform. Your video host. Anything with a login.

Last week I handed off two projects I'd been putting off for months: a full email marketing overhaul, and pulling text transcripts from over a hundred videos.

Every time an agent connects to another app, it takes one of three doors. You don't need to be technical to use any of them. But once you know the three, you'll know exactly what your agent is doing on your behalf, and you'll start spotting projects you can hand off completely.

Door 1: MCP, the door built for agents

An MCP is something you install into your agent, whether that's Claude Code, Codex, or the Claude app. You install it once, and from then on your agent walks into that app already knowing what it can do there. The capabilities come baked in.

The catch: the provider has to offer one. And even when they do, it might not cover everything the app can do. That gap matters, as you'll see in a minute.

Door 2: API, the door built for developers

APIs are how apps have talked to each other since long before AI agents. An API usually requires a credential, which means a bit more legwork to set up. This is the door that comes into play when you're building your own custom apps or custom agent skills that need to interact with a service.

The key difference from an MCP: with an API, your agent comes in knowing nothing. It has to get itself up to speed. You can hand it a link to the API docs, or let it run its own probing tests to figure out what the API can do.

Door 3: the browser, your own front door

No integration at all. The agent drives the same logged-in browser you use, clicking buttons, filling forms, and downloading files like a human would.

Browser use gets treated as the last resort, but these agents have gotten shockingly good at it. When there's no MCP, or the MCP doesn't cover something, or there's no API, or you just don't want to build a custom integration, the front door is always there.

And you don't have to memorize any of this. Ask your agent "how can I connect to X?" and it will pick the route. Knowing the doors is the difference between watching your agent work and directing it.

All three doors in one project

My email flows in Kit were due for a big overhaul: sequences, tagging, onboarding, all of it. Anyone who's done email automation knows how tedious that gets. I delegated it to Claude Code, and the project ended up using all three doors:

• The API. Every email form on my website hits Kit's API. Someone grabs a free tool or joins a workshop, and the API subscribes them and tags them so they route into the right flow automatically.
• The MCP. To plan and write the emails and architect the flows, I built a Claude Code project and installed Kit's new MCP. Claude could read my sequences, push drafted emails, and configure settings. Most of the capabilities, but not all.
• The browser. The most important piece, Kit's visual automation flows, can't be touched by the MCP or the API. So Claude used a skill called agent-browser to open Kit and assemble those flows by clicking through the editor, just like I would.

One project, three doors, and the agent picked the right one for each piece.

When the browser is the only door

Members of my community wanted text transcripts for my course videos. I host over a hundred of them on Vimeo, and getting transcripts the manual way means opening every video, navigating to subtitles, and downloading files one at a time. I pushed it off for months.

No MCP for this. No API set up either. So I asked ChatGPT how it would handle it, and the answer was: just log into Vimeo in Chrome. It took over the tab, worked through the videos, and about 20 minutes later there was a folder on my desktop with every transcript. Claude Code then built a transcripts feature into my member library and pushed them to the site.

Days of tedium, done in about 45 minutes.

The leverage math

If you sit and watch an agent click through a browser, it looks slow. You could do any single step faster than it can.

But you're not supposed to watch. You hand off the project, go work on something else, and an hour later the thing is done. The agent's 20 minutes cost me about 2 minutes of attention. The leverage isn't that the agent is faster than you. It's that it works without you.

So pick the tedious, click-heavy project that's been sitting on your list for months, and ask your agent one question: "How would you connect to this?" It will tell you which door it can take. There's a good chance you can delegate the whole thing this week.

This kind of delegation is the heart of working AI-native. If you want the full method, my free AI-Native workshop is at buildermethods.com/workshop.