Can we set up an MCP server so Claude writes in our brand voice?

Yes, and the connecting part takes about four minutes. Claude supports custom connectors, so you point it at a remote MCP server and it pulls your positioning, your ICP and your language rules straight into a draft while it's writing. The four minutes are the easy part. What decides whether the output sounds like your company is what that server hands back when Claude asks for it, and most teams fill it with tone adjectives that nobody could have written a single sentence from.

What actually happens when Claude is connected to a brand server?

Model Context Protocol is an open standard for letting an AI assistant reach outside its own head mid-task. Your server exposes a few named tools. When a request looks like it needs one, Claude calls it, gets structured text back, and uses that text for the turn it's working on. Nothing gets retrained. Nothing gets memorized. Every draft is a fresh fetch against whatever the server says today.

That mechanic has two consequences worth understanding before you build anything. The first is good news: update the source and the next draft is already current, with no retraining and no email asking twelve people to please use the new deck. This is a different thing from fine-tuning, and we've written about how AI actually learns your messaging if that distinction is fuzzy.

The second consequence catches people out. Claude only calls a tool when the request gives it a reason to. If your tools are named vaguely and described loosely, Claude writes the email without ever fetching anything and you never see an error. The draft just comes back generic, the same as it did last month, and the team concludes the server is broken. Tool names and descriptions are part of the content job, not a technical detail you hand to whoever set up the endpoint.

What does the setup actually take?

We shipped our own on August 17, 2026: the Test Kitchen, a remote MCP server at mcp.pitchkitchen.com with three tools on it. One scores a homepage against our 19-criteria rubric, one serves the Magnetic Messaging Framework templates, one serves our agent roster. You can connect it to your own Claude right now and watch a fetch happen, which is a faster way to understand the mechanic than reading another explainer about it.

In Claude, the path is Settings, then Connectors, then Add custom connector, then paste the URL. In any chat you turn the connector on from the plus menu and ask. Claude Code takes a single command. Cursor takes one JSON entry. ChatGPT needs developer mode switched on first, then a connector created by hand. None of it is hard, and none of it is where your time goes.

Here's the honest accounting from our own build. Standing the server up and connecting it took an afternoon. The weeks went into deciding what each tool returns, in what shape, and how blunt it's allowed to be. That ratio holds for every brand server we've looked at.

Why doesn't Claude sound like us even after we connect it?

Because most teams serve voice and stop there. The file says confident, human, direct, never salesy, avoid jargon. Claude reads that, complies with it, and produces copy that a hundred other companies could publish under their own logo. You asked for a tone and you got a tone.

Voice is the surface of a position. What makes a sentence recognizably yours is what you're willing to claim out loud, who you're willing to turn away, and the argument you keep making that your competitors won't make. Strip the adjectives off a paragraph that has those things and it still reads like you. Serve the adjectives without them and you've hired a rewriter. We went deeper on this in why AI writing still doesn't sound like you.

There's a fast test. Take a draft Claude produced with your connector on, delete your logo and your product name, and hand it to someone in sales. Ask them which company wrote it. If a competitor could ship it unchanged, your server is serving tone. That's the Cover-the-Logo Test, and it's the same failure our free Brand Signal Score picks up on homepages every week.

What should the server hand back when Claude asks?

Named units, each carrying one decision, each retrievable on its own. Claude writing a cold email needs the ICP and the disqualifiers. Claude writing a launch page needs the positioning claim and the proof. Bundling all of it into one blob means every request drags in everything and the model averages it. Turning a messaging framework into something machines can serve is mostly this splitting job.

Notice how much of that column on the right is positioning work wearing a technical hat. That's the whole story of brand servers. Companies arrive thinking they have a tooling gap and find out they have five people who'd answer 'who are we for' five different ways. Deciding whether to build one at all is worth its own hour, and we laid out the readiness checks in should we build an MCP server for our brand.

How do we tell whether it's working?

Run the same brief twice, once with the connector off and once with it on, and read the two drafts side by side. If you can't reliably pick which one had the server, the server is returning descriptions rather than decisions. That test takes five minutes and it's brutal, which is why almost nobody runs it.

Then watch which tools actually get called. If Claude reaches for the voice tool every time and the positioning tool never, your descriptions are steering it to the shallow layer, and the fix is in the wording of the tool, not the content behind it.

The operational signal shows up a few weeks later, quietly. Reps stop pasting the messaging doc into chat windows. Nobody announces it. You notice because the drafts landing in review stop needing the same three corrections, and the person who used to rewrite everything gets their Thursdays back.

What this means for you

If someone asked what your server would return today and the honest answer is a brand book full of adjectives, you're a quarter early. Connecting Claude to that is a two-hour job that broadcasts vagueness to every tool your team touches, faster and more consistently than any human could manage on their own.

If your positioning is settled and your real problem is that eleven people write customer-facing copy and only four of them get it right, connect it this week. The mechanic works, it's genuinely four minutes, and the leverage is real once there's something worth serving. Want to see one running before you commit? Point your Claude at our Test Kitchen server, or come to the AI Workforce Clinic and we'll walk your team through a live fetch. The first Clinic is free.