Quick answer
An MCP server for Business Central exposes ERP data and actions as tools that Claude can
call — search customers, check availability by bin, list open orders, draft a quote. A
governed server runs every call under the signed-in user's Entra ID identity, separates
read from write tools, asks for approval before writes and logs everything. It does not
replace posting rules, permissions or your partner's judgement.
Everyone has seen a demo where an assistant "talks to the ERP". The Model Context Protocol (MCP) is what makes that repeatable instead of a one-off integration. This post explains what an MCP server for Business Central actually is, what Plur-e builds, and the boundaries we keep.
MCP is an open protocol in which a client (Claude Desktop, Claude Code, claude.ai, or your own agent) connects to a server that publishes tools with JSON schemas. The model decides which tool to call, the server executes it against the real system and returns structured results. The same server works with any MCP client.
Grouped by area, with a scope per tool:
| Area | Read | Write |
|---|
| Customers | search, card, ledger | — |
| Items | search, availability by location and bin, ledger | — |
| Sales | list documents | create quote, create order (approval) |
| Purchasing | open POs, vendor | — |
| Warehouse | activities, bin contents, count discrepancies | — |
| Documents | posted invoice/shipment/receipt | — |
A typical conversation: "Which open sales orders for Conchita Foods are short on stock at MAIN?" → list_sales_documents → get_availability per line → a table with the shortfalls and links to the records.
- Bypass permissions. Calls run as the user; if the user cannot see a customer, neither can Claude.
- Post without approval. Write tools are per role and prompt for approval; posting routines stay in Business Central.
- Invent numbers. Results carry the record URL so answers can be checked; the server returns data, not opinions.
- Train on your data. Anthropic does not train on API data by default.
A custom bot is one integration for one UI. An MCP server is one integration for every MCP client — and the tool schemas double as documentation. When Business Central adds a field, you update one tool.
A two-hour discovery workshop identifies the ten questions your team asks Business Central every day. A pilot of a few weeks ships the read tools for those questions; writes come after the audit trail is in place.