5 proven Claude + Hygraph MCP workflows that actually save you time

Can AI structure noise into knowledge? We tested Hygraph MCP on one real question

Fourteen messy files, two unrelated data sources and an empty Hygraph project. How our CMS Camp Mallorca workshop turned noise into a queryable content graph.
Bohdan Zatirko

Written by Bohdan 

Nov 16, 2026
Can AI structure noise into knowledge? We tested Hygraph MCP on one real question

I've been a fan of MCP since the first time I connected an AI assistant to a Hygraph project and watched it read our schema back to me. Demos are easy to love, though. The sample data is clean, you already know the happy path, and the prompt works on the first try. What I wanted to know was how far I could trust it with the kind of mess every company actually has.

With AI, unstructured content goes from a content ops headache to the thing that blocks everything else, since an agent can only act on content structured enough to reason over. So I couldn’t help but ask: can MCP turn that noise into knowledge without someone babysitting every field?

Content is only AI-ready for a specific use case, as Gartner points out for data. So we picked one question that needs two systems to answer today.

At CMS Camp Mallorca, Daniil Tykheev, our Director of Product, and I tested it in a two-hour, hands-on workshop. A passenger at Palma airport asks: “My flight is disrupted. What should I do, and what’s happening at my gate?” The answer sits in a help center and a set of live boards that were never connected. Here’s how it went, and how you can apply the same pattern.

#Where noise lives, and what it costs

Noise looks different depending on who you ask. Before we started, we asked the room what noise meant to them, and nobody gave the same answer. Some talked about editors drowning in content they can't find or manage. Others talked about end users who can't get an answer, or teams that can't make a decision because knowledge is scattered everywhere.

In practice, noise lives in more places than most teams expect:

  • Legacy CMSs
  • PDFs and Word or Google Docs
  • Spreadsheets
  • Support tickets and call transcripts
  • Processes
  • The knowledge that only exists in people's heads

Some of it isn't content at all. Nobody knows who owns a status value, who updates the board, or what happens to a draft nobody finishes.

That noise has a price. An agent gives a wrong answer with full confidence. Two teams answer the same customer question from different sources. A third-party AI assistant describes your brand using a page you retired last year.

The gap is between content that works for people and content an agent can use. Closing it is what our mission at Hygraph is about: structuring noise into knowledge. It's also why I keep coming back to a line we use a lot internally. AI doesn't need better prompts. It needs better context.

My assumption going in was that MCP could do most of the heavy lifting under two conditions. The content model has to be right, and a human has to stay in the loop at the moments that matter. The workshop was a chance to test both.

#The project: one airport, two teams, no shared graph

We gave everyone the same brief. You're the implementation partner. Palma de Mallorca airport has brought you in because their content operation is a mess. Passengers, operations and communications all need the same information on the same morning, and it lives in two places that never meet.

On one side sits a help center full of drafts. One article still carries a heading marked "do not show this heading to passengers". Somewhere in a baggage thread, someone pasted an internal ops comment that was never meant for travelers.

On the other side sits live operational data. A belt-number column says HOTEL, and a blank cell means nobody has updated it yet.

So we had two use cases owned by two teams in two different tools, and nothing connecting them. If that sounds like an airport problem, swap in your own: product documentation and inventory feeds, support articles and incident status, policies and store data. Knowledge and operations sit in separate silos in almost every company I work with.

The job was to build one content graph you can query, then a surface a passenger could actually use. Screenshot 2026-11-10 at 18.19.58.png

#Step by step: from messy files to one content graph

Here's how the two hours went. Every step ran through the Hygraph MCP Server, using whichever AI tool people preferred. Claude and Cursor were the usual picks. We didn't expect anyone to bring Codex, but for the people who did, it worked well too.

1. Look at the noise before touching anything

Everyone started with an empty Hygraph project and connected it to their assistant. Then we attached the files and asked it to walk through them without writing anything: “What is noisy, missing, or unqueryable?” The list it came back with made it much easier to decide what the content model needed.

2. Propose the model, then build it around one bridge

We asked for the smallest model that could hold both use cases, and people reviewed it before anything was created. It had a knowledge article, an operational record, a topic, a facility for gates and zones, and one shared tag. Put service-disruption on a help article and on a live record, and one query finds both.

The review step exists because of our dry run. Cursor proposed HelpArticle, PublicCue and BoardRow instead of the names we’d asked for, and the service-disruption tag quietly disappeared. That tag is the one thing that connects the two domains, and it’s the first thing an assistant drops, because each model looks fine without it.

So I’d treat it as a rule: before anything gets built, write down the contract (the model names and the shared bridge) and check the proposal against it. That check is a few minutes of review, and it’s what stops a graph that looks right from falling apart.

3. Ingest, and make the assistant show its guesses

MCP turned each file into structured entries and created facilities for gates and zones where they were missing. Because we asked it to list what it had to guess, every assumption was visible, and anything it couldn’t place, like a belt number that read HOTEL, came back flagged for review. We also added one article by hand, so human-made and AI-ingested content sat in the same model.

4. Ask one question across both worlds, then ship a page

After merging duplicate tags, we asked for everything tagged service-disruption, grouped into what a passenger should do and what’s live on the boards. Then we tried “What’s happening at gate B14, and is there a help page I should read?”, and the agent hopped from the gate to the tag to the right article. To finish, everyone built a one-page passenger search on the Hygraph Content API, which was quick because the structure was already there. Screenshot 2026-11-10 at 18.23.56.png

#What happened in the room

By the end, people had gone from a pile of mixed-format files to a connected, queryable content graph with a working search page on top. The internal notes, stray comments and half-finished drafts were cleaned up or flagged. A help center and a set of live boards were finally answering the same question.

My favorite moment came from one attendee who completed the entire workshop on his phone. He didn't need a laptop or a terminal, only an AI assistant connected to the MCP Server.

The room was mixed, with developers sitting next to people who had never configured an MCP connection, and that showed. One non-technical participant needed our help to connect MCP, and later got stuck reviewing the content model. Daniil and I walked her through it and she got there. I'm glad it happened, because it pointed straight at where the real learning curve is.

#What I learned

You can trust MCP when you give it guardrails

That's my answer to the question I started with. Prompts like "do not write yet", "propose, don't build" and "tell me what you had to guess" turn the assistant into a colleague whose work you can review. We saw and agreed to each write before it happened.

The content model is the real work

Ingestion was fast. The value came from what we decided before importing anything: keeping two content types, linking them with one shared tag and giving locations their own model.

Structure first, surfaces second

Once the graph existed, the frontend was cheap, and chat and web were querying the same thing. If an agent can't query your content cleanly, a frontend can't either.

People need to understand the model

Connecting MCP and reading a content model were the two places people asked for help. AI speeds up the work, and someone still has to judge whether the structure makes sense. Teams that build that skill will get the most out of it.

#How to apply the pattern in your own team

The airport was a stand-in. The same pattern works wherever two teams hold different halves of the same answer, and this is how I’d run it on a real project.

  1. Pick one question that needs two systems today. Look for something customers or colleagues ask that nobody can answer from one place, like “Is this product in stock, and what’s the return policy for it?” or “Which stores are affected by the outage, and what should staff tell customers?” A concrete question keeps the scope realistic.
  2. Name the two domains and who owns them. Write down which team owns each half, which tools it lives in and what shape it arrives in. At the airport, that was the help center on one side and the live boards on the other.
  3. Choose the bridge. Find the one concept both sides already talk about. It can be a tag like service-disruption, a location like a gate or a store, or an entity like a product or customer account. If you can’t find one, the two domains probably don’t belong in the same question yet.
  4. Model the bridge first, then everything around it. Put the shared concept into the schema before anything else, and check the proposal against your contract before anything gets built.
  5. Ingest one domain at a time and ask what the assistant guessed. Review those guesses before you load the second side, and make sure both sides use the same bridge values.
  6. Ask your question across both domains. If one side comes back empty, check the tags and the publishing stage before you question the model.
  7. Give the bridge an owner. A workshop can skip this step, but a real team can’t. Someone has to own the tag list, decide when two tags mean the same thing and keep live data fresh. Otherwise, the noise comes back one unowned tag at a time.

The same approach works for migrating eCommerce product data with Hygraph MCP and Claude, too.

We went to Mallorca asking whether AI can structure noise into knowledge. With a clear model and a human in the loop, it can get you a first working graph in a few hours. Keeping that graph healthy is the work that comes after.

Want to see what it does with your own content? Read the Hygraph MCP Server docs and start for free.

Blog Author

Bohdan Zatirko

Bohdan Zatirko

Principal Solutions Architect

Bohdan Zatirko is a Principal Solutions Architect at Hygraph. After years in content platforms and observability at Contentful, Uniform and Datadog, he now helps enterprise teams design systems that are structured, governed and ready for what comes next.

Share with others

Sign up for our newsletter!

Be the first to know about releases and industry news and insights.