How do we get every AI tool our team uses to follow our messaging framework?

You stop handing it to people and start serving it to machines. Concretely: take the positioning, the ICP with its disqualifiers, the proof points and the banned language out of the deck, and put them behind an MCP server, a live endpoint every AI tool your team uses can query mid-draft. The document stays for humans. The server is the copy the tools read, and it's the only version that gets followed at 4pm on a Thursday when someone's rushing an objection-handling email out the door.

What's actually going wrong when five people prompt five different ways?

Every person on your team is quietly re-summarizing your company each time they open a chat window. The rep types a version of what you do. The product marketer types a different one. The founder types the good one, because he's the only one who's said it out loud four hundred times. Each model fills whatever's missing with the industry average, which happens to be the same industry average your competitors get handed.

The left side of that picture is what most $5M-$75M companies are running right now, and nobody chose it. It happened because the framework is a 40-page document and the tool asked for context in a text box. Nobody's being lazy. They're compressing your narrative identity down to whatever fits, forty times a week, and the compression is different every time.

The cost shows up as rework. Your team writes with AI to go faster, then spends the saved hour dragging generic copy back toward something true. That's the tax you're already paying, and it's the number worth measuring before you scope anything.

There's a second cost nobody logs. Outdated specs travel. A pricing number from two quarters ago, a customer count that grew, a feature you sunset in March: each one sits in somebody's saved prompt and keeps shipping. Buyers catch it before you do, usually on a call, usually in the worst possible sentence.

What does a messaging framework have to become before a machine can use it?

A model can't act on atmosphere. "Bold, human, approachable" gives it nothing to do. Three changes turn a framework into something queryable.

First, decisions replace descriptions. Who you're for, and just as load-bearing, who you're not for. The named alternative buyers actually compare you against and the honest reason they pick you. The villain you name. Proof points with real numbers and the dates they were true. A model can execute a decision. It can only paraphrase a description.

Second, prose becomes retrievable units. One claim per unit, each under a name a tool can ask for: positioning, ICP, disqualifiers, competitors, proof, vocabulary, offers. A writer drafting a cold email needs three of those, not all thirty-five. Retrieval beats recitation, and it's cheaper on context.

Third, the whole thing gets versioned and owned. Dated, single editor, one place. The moment two edited copies exist, you've rebuilt the original problem with better technology. This is the same discipline behind a Voice Spec your team and your AI will both follow, applied one layer down at the source.

How do we turn our framework into an MCP server?

Four steps, and only one of them is engineering.

We built ours, so this isn't a diagram we admire from a distance. PitchKitchen's Test Kitchen went live on August 17, 2026 at mcp.pitchkitchen.com. It exposes three tools any AI client can call: score a homepage with the Brand Signal Score, pull the MMF template, and run THE LINE. Sprint clients get a private server carrying their own AI Brand Twin instead of ours. The engineering took days. What took real time was pinning down which sentence was actually true, which is exactly where it should hurt.

What does this actually change on a normal Tuesday?

A rep is drafting a follow-up to a prospect who just said the dreaded "we'll keep doing what we're doing." Without the server, he pastes in whatever he remembers, the model writes a competent paragraph about efficiency and scale, and he sends something a hundred other vendors could have sent. Nothing is wrong with it. Nothing in it is yours either.

With the server, the same tool pulls the named alternative that prospect is actually comparing against, the disqualifier that tells the rep this deal is real, the villain you name and the two proof points with dates attached. The draft comes back arguing your position. He edits a line and sends it. That gap, repeated across everyone writing customer-facing language every week, is the entire return on this.

What belongs on the server, and what stays off it?

Put on it anything a writer would otherwise guess at: the positioning sentence, the ICP and its disqualifiers, named competitors with the honest reason buyers choose you, proof points carrying numbers and dates, the vocabulary you refuse to use, and format rules for the pieces your team writes weekly.

Leave off the mission statement, the visual identity rules, and the adjective cloud. A model drafting a follow-up email has no use for your hex codes, and aspirational adjectives give it permission to generate the same beige paragraph every other company gets. Everything you add that isn't a decision dilutes the things that are.

How do we know it's working?

Watch four things, none of which are traffic. First-draft acceptance: how often a piece of AI-assisted copy survives review without a rewrite. New-hire ramp: whether someone in week one produces language a founder recognizes. The two-room test: two people, separately, writing one sentence on who you're for and why you win, and the sentences matching. Stale-claim incidents: how often a number from last year escapes into a live deck.

Baseline them before you build. Ask three people to draft the same email today and keep the results. Run the two-room test cold and write down both sentences. Those two artifacts take an hour, and they're the only honest before-picture you'll get, because once the server is live everyone remembers the old way as worse than it was.

One honest boundary. A brand MCP server is internal plumbing your team's tools query. It doesn't change what ChatGPT tells a stranger about you, because buyers asking for recommendations never touch it. That's a public-surface problem, and it comes back to why AI engines describe you the way they do. Any vendor who blurs those two together is selling you one thing and pricing it as both.

What this means for you

The picture at the top of this page has a left side and a right side, and moving between them is one decision plus one afternoon of configuration. What separates the two isn't the protocol. It's whether there's a settled narrative worth broadcasting, because the server will distribute whatever you hand it with perfect consistency, including confusion.

If your framework is real and your team still writes off it half the time, you have an enforcement problem and the server fixes it. If two people in separate rooms would write two different companies, build the framework first. The decision guide for that call sits in should we build an MCP server for our brand and messaging guidelines.