Connecting Claude to a headless CMS: how our content pipeline works

By Max Ikaheimo

October 1st, 2026

Content-heavy platforms need a lot of accurate content, and keeping it up to date is a job of its own. For Laturikotiin.fi, a Finnish EV charger comparison and installation platform, we built a pipeline where Claude does the heavy lifting on drafts and fact-checking, and the team stays in charge of what gets published.

What MCP does here

The Model Context Protocol (MCP) gives an AI assistant a defined set of tools for working with another system. Instead of copying text between a chat window and the CMS, Claude connects to Sanity through MCP and works with the content directly: it can read the content model, create drafts and fill in structured fields.

That structure matters. Chargers, guides and FAQs each have their own content type with defined fields, so Claude fills in specs, prices and answers in the right places instead of producing one block of text.

The workflow

• Drafting: Claude creates and fills drafts for chargers, guides and FAQs in Sanity.

• Fact-checking: Claude checks specs, prices and regulations against manufacturer pages and Finnish authorities such as Traficom and Verohallinto.

• Review and publishing: the team makes the final edits in Sanity Studio, reviews everything and publishes it.

Why drafts only

Claude fills the drafts and checks the facts. People make the final edits and hit publish. That publishing gate is the most important part of the design: nothing reaches the live site without a human decision, and every change is visible in the CMS before it goes out.

Fact-checking as a step, not an afterthought

Prices change, product specs get updated and regulations are revised. Making fact-checking an explicit step, with named sources, is what keeps the content trustworthy over time.

It also helps with AI search. Assistants like ChatGPT, Perplexity and Google's AI Overviews prefer sources with concrete figures, named references and visible update dates. Our post on Generative Engine Optimization covers this in more detail.

What you need for a setup like this

• A structured content model, so the assistant writes into defined fields instead of free text.

• Clear editorial rules: tone, required sources and what must always be checked.

• Scoped permissions, so the assistant can create and edit drafts but not publish.

• A review step owned by people, with time reserved for it.

The same approach works for any platform that combines a real backend with a lot of content. Read more about our AI development services, or the full Laturikotiin case study.