Before your agents can answer questions about the business, something has to be true for them to read.

• "We tried pointing ChatGPT at the database and it made things up."
• Managers want to ask "how did we do last week against the same week last year?" without waiting on a report.
• An agent vendor asked for "clean data", and nobody knew what to send them.
• Someone connected a chatbot to production, and now someone else is asking what it can see.

An MCP server over the rebuilt schema. Read-only, always. It answers the questions the business actually asks, and it carries the caveats the data really has, written where the model will see them: which periods are partial, which figures are current through yesterday rather than this morning, which costs are missing because no system records them. A model that knows those things says so. One that doesn't gives you a confident wrong number.

Access is scoped per person. Everyone can see the aggregates, and guest names or pay rates stay with the few people who need them. Sign-in is the one you already use.

What changes: people get an answer in a minute instead of waiting on a report, and it comes from the same tables the finance numbers come from. When something looks wrong, they can ask which entries make it up, and the detail adds up to the total.

MCP, the Model Context Protocol. Claude, or any OpenAI-compatible model, including one you host yourself if the data shouldn't leave the building. PostgreSQL underneath. Microsoft Entra or Google sign-in for who is asking.

Anyone who wants agents writing to production. Everything here reads. Nothing I build lets a model change your data.