What an MCP server does for your company's rules

If you search for “MCP server” you’ll land in a world of GitHub repos, config files and lists with hundreds of entries. Servers for Postgres, for Slack, for the file system, for browsers. Useful if you write code. Confusing if you run a company and someone told you that MCP is how you get your AI to “know your business”.

Both are true. The difference is what the server holds: your company’s rules instead of a database.

MCP in two paragraphs

MCP stands for Model Context Protocol. It’s an open standard for connecting AI assistants to things outside the chat. An MCP server offers a handful of tools, each with a name and a description. The AI reads those descriptions and decides on its own when to call a tool.

Claude, ChatGPT (Business, Enterprise and Edu), Microsoft Copilot through Copilot Studio, and coding tools like Cursor can all connect to remote MCP servers over HTTPS. You add the server once as a connector, sign in, and from then on the AI can use it in any conversation.

Why that matters for rules

Most companies feed their AI context by copying it in. A policy pasted into a chat, a handbook uploaded to a project, a few lines in custom instructions, a SKILL.md with the price list inside. Each of those is a copy. Copies age, and nobody knows how many there are.

An MCP server works the other way round. The AI asks at the moment it needs something, and the server answers with what’s true now. If the return window changed yesterday, today’s answer uses the new one. Every connected tool asks the same server, so ChatGPT in sales and Claude in support get the same rule.

What the AI should be able to ask

A server for business rules is only useful if the AI can ask the right questions. In our experience you need roughly these:

• Search for the rules that fit a question (“can we refund this order?”).
• Read one rule with its reason, owner and version, so the answer can cite it.
• Look up which rules a recurring task depends on.
• Propose a change when a user says a rule is wrong. The owner decides, not the AI.

That last one matters more than it looks. People correct the AI all the time. Without a way to send that correction to the person responsible, it’s lost after the chat ends.

What it should not do

The AI shouldn’t be able to change a rule. Reading and proposing, yes. Publishing, no. Otherwise one confident chat rewrites your refund policy.

It also shouldn’t see drafts. If someone is halfway through rewriting the discount rules, the AI should keep using the approved version until the new one is published. A server that serves whatever is in the folder will hand out half-finished work.

Hosting your own or using one

You can build an MCP server yourself. The protocol is open and there are good SDKs. For a team with developers and a clear use case, that’s a reasonable weekend project.

The hard part comes after the weekend. Who writes the rules down? Who approves them? What happens when two rules contradict each other? How do you see which rule an answer was based on? The server is the small part. Most of the work is keeping the content right.

Where to start

Ask yourself where your AI gets its company knowledge today. If the answer is “people paste it in” or “it’s in a few project files”, you have copies. Count them for one important rule, like your payment terms. That number tells you how much a single source would save.

CtxCore Brain is a hosted MCP server for exactly this: business rules with an owner and a version, approved before the AI sees them, readable from Claude, ChatGPT, Copilot and your own agents. If you’re still deciding what should go in, start with what a company brain is.