Sometimes you don't need an MCP

You can teach Claude to use any public API directly - no local server required. Here's how I post to Ghost from an iPad in an airport, and when you should take a more secure approach instead.

Hand-drawn illustration of a key

I'm writing this on an iPad in Chania airport, about to board a flight, and I'm going to push it straight to Ghost as a draft the moment I'm done. No laptop, no local server running, no special setup. Just Claude and a skill file.

That wasn't always possible.

I use Claude to help me draft content for this blog. Not to generate it - the ideas and experience are mine - but to shape rough notes into something readable. The natural next step was having Claude push that draft directly to Ghost so it's waiting for me when I'm back at a desk. The Ghost MCP on GitHub exists for exactly this purpose, but it faced so many issues I lost count, and the fundamental one is that it's local. It runs on your machine. Which means the moment you're out and about with only an iPad, it stops being useful. Yes, you can set up a local MCP and use your laptop as a proxy - but that assumes your laptop is on, at home, connected, which is a lot of assumptions to make when you're at an airport with an idea and ten minutes to spare.

So I went a different route.

Teach Claude to use the API directly

Any public API is fair game. Claude can read documentation, understand authentication flows, and generate working code. The process I landed on takes maybe 30 minutes the first time:

  1. Find the API you want to use
  2. Ask Claude to write a report on it - endpoints, auth method, key operations
  3. Feed that report back to Claude and ask it to write a skill file
  4. Test it

That's it. For Ghost specifically, there were a couple of gotchas the documentation doesn't shout about - JWT tokens that expire every five minutes, and a ?source=html parameter you have to append to POST requests or Ghost silently creates a blank document. Neither of those came out of the docs. They came out of testing. Once the skill had those details, it just worked.

The credential question

The obvious trade-off is that your API key lives in the skill file. For a blog CMS, I can live with that. The worst realistic outcome is unwanted content getting posted. Not catastrophic.

But that ceiling is real. Anything touching billing, financial records, or personal data needs a different approach - and that's where hosted MCPs make sense. Rather than running an MCP server locally on your laptop, you deploy it somewhere you control: a small VPS, a cloud function, whatever fits your setup. Claude connects to it over the network the same way it would a local instance, except it's always available, works from any device, and you can manage credentials properly through environment variables rather than a text file. You also get proper access controls and logging if you need them. More work to set up, but for integrations where data sensitivity actually matters, it's the right call.

Which tool for which job

Skills with embedded credentials work well for low-stakes personal integrations. Hosted MCPs are the right answer when security matters. Local MCPs are fine for development and testing if you're happy being tethered to your computer.

MCP is a good abstraction and it'll keep getting better. But the ability to just describe an API to Claude and have it work with that API directly is genuinely useful right now - particularly when you need something that works from wherever you happen to be.