Publish a blog post without opening the CMS?

Earlier today I created a post for this blog without opening the headless CMS I've been using to run it. I didn't login to the Contentful Platform, nor did I draft any copy within it.

Instead, I had a conversation with Claude, picked a few options from menus, and then the entry appeared in my space, fully populated and ready to publish.

This is what working with an MCP server looks like in practice, and it's worth writing about because I think it changes the shape of content production in ways both meaningful and mundane.

What's an MCP then?

Model Context Protocol is, in the dullest possible description, a way for AI models to call external tools.

Anthropic publishes the spec; Contentful (and others) ship MCP servers that expose their APIs through it.

When I enable the Contentful connector in Claude, the model gets access to a set of functions — list_content_types, search_entries, create_entry, and so on — that it can call on my behalf during a chat.

The interesting thing isn't the protocol, though. It's that it makes a content platform addressable from a conversation.

What I actually did

After an initial setup which was trivially easy, I opened a chat and asked Claude to check whether the connector was working.

It tried to fetch the initial context, hit a "no approval received" error, and told me to look for an approval prompt or toggle the connector off and on.

I toggled it; the second attempt succeeded, returning the space ID, environment ID, and a long set of instructions that the Contentful MCP server provides to every model that connects.

From there, I wrote: "Give me a list of content types." Claude returned two: Post and Author. Post has seven required fields, including a cover image asset and an author reference.

"Show me entries." Thirteen posts, three authors — a small space, recognisably mine.

Then, the moment of truth: "Shall we create a test post and preview it?"

Chatting with Claude. What could possibly go wrong?

And that's exactly what we did. Claude asked which author to attribute it to, what the post should be about, what to do for a cover image.

When I said "reuse an existing asset," it paginated through the 22 assets in my space — Contentful's MCP server returns a maximum of three per call — and offered me a shortlist, then a fuller list when I asked for more options.

I picked the key visual, told it to attribute the post to Genial Yeti, and asked for a short note about the MCP connector test itself.

Claude constructed the entry, including a rich-text body with headings and a bulleted list. The entry appeared as a draft, with an ID and a deep link to the entry in the web app. I previewed it, made some tweaks, hit publish, and now it's live.

End to end, maybe twenty minutes of conversation, and most of that was me being indecisive about cover images.

What worked, and one thing I want to flag

Reading the space — content types, entries, assets, fields, references — felt natural. Claude inferred what I meant from short prompts.

Linked references for the cover image and the author worked first time. The conversational picker for the cover image, where Claude paginated through assets and offered me a shortlist, was a genuinely nice pattern: more like asking a colleague "what have we got?" than scrolling a media library.

The one bit of friction worth being honest about: I'd left the MCP connector installed in read-only mode by default, which was the right safety choice but meant the first attempt to create the entry failed.

I had to leave the chat, flip the permission in Contentful's settings, and come back. That's on me for not configuring it before we started, not on the connector. If you're trying this yourself, set the write permissions up front.

The joy of a structured content model

One theory as to why the process was so smooth is that this blog's content model is deliberately simple.

The Post type has seven fields, all top-level, with two references — one to an Author entry, one to an Asset for the cover image. No nested components. No modular page sections. One locale.

That simplicity meant Claude could understand the whole schema in one look, ask three clarifying questions (author, topic, cover image), and have everything it needed.

I suspect — but haven't tested — that a more complex model would change the experience.

A page built from a tree of modular components, each with their own references and validation rules, would mean more clarifying questions, more chances for the model to guess wrong, and more cases where the right answer requires seeing the layout rather than describing it.

My hypothesis is probably rubbish, though. Down the line, I will tinker with the content model and see how Claude handles the changes.

Fields populated with content, all present and correct.

What does it all mean?

The temptation is to frame this experience as "AI will replace your CMS." That's the shallow reading, and I don't think it's correct.

What it actually changes is the cost of small, structured tasks. Creating a draft entry from scratch — title, slug, excerpt, body, references, cover — is something that can be done with a prompt rather than manual entry.

The fields are simple; finding them, tabbing between them, picking the right asset from a grid of 22 thumbnails, is what takes the time. A chat interface compresses that into "make me a post with these properties and this cover," and the model handles the field-by-field work invisibly.

This is genuinely useful for repetitive tasks. It's less useful for the work where I'm actually thinking — where I want to see the page laid out, see the cover image next to the headline, judge whether the excerpt has the right rhythm. The UI of a content platform exists for a reason, and that reason isn't going away.

The right framing, I think, is that MCP connectors give you a second mode for content work. The web app for considered editing. The chat for everything else.

Closing thought

The thing that surprises me, sitting here writing this, is how quickly twenty minutes of MCP-driven work has stopped feeling novel. It's already just how this kind of task gets done. That's the real signal. The demo works, and you barely notice the seams.

Prompt for key visual

Midjourney: "Editorial illustration, close three-quarter view of a vintage wood-cased CRT television sitting on a sideboard in a dimly lit mid-century living room. The television is switched on; its slightly curved glass screen glows warmly, displaying a jigsaw puzzle in progress — most of the puzzle is assembled, showing a soft abstract image, but one piece-shaped gap remains in the lower right. A single human hand enters the frame from the right, fingers gently pressing the missing puzzle piece into place against the glowing screen. The hand is mid-toned, ordinary, unadorned. The room around the television is in deep shadow, with hints of patterned wallpaper, a doily under the set, a framed picture just visible on the wall. Warm amber glow from the screen as the dominant light source, casting soft light onto the hand and the front of the cabinet, with deep falloff into the rest of the room. Gouache illustration with visible brushwork, muted earth-tone palette with the warm screen glow as the focal point, 1940s–1950s American magazine illustration style. --ar 16:9 --s 350 --no deformed hands, extra fingers, text, photorealism, modern computers, modern flat-screen television"