Microsoft AI · MCP
The bridge between AI and your business systems.
Model Context Protocol is how AI agents reach enterprise systems. We've built MCP servers since the spec landed — for Business Central, Microsoft 365, Fusion 360, and for clients who need their own tools exposed to Claude or ChatGPT. Every server follows one rule: an agent that can write to your ERP gets guardrails the model can't talk its way past.
Three paths in
Run it yourself, have it installed, or have it built.
Free · open source
Business Central MCP
Install bc-mcp-proxy in an afternoon: pip install, an Azure App Registration, OAuth scopes, first call. Point Claude at your BC tenant and ask a customer's balance, their overdue invoices, their last purchase — in plain language, against live data. MIT-licensed, no account with us.
The setup guide
In your own Azure
BC MCP Pro
Business Central in Claude and ChatGPT — in the browser, on the desktop, on your phone — with nothing to install on anyone's device. We set it up in your own Azure subscription and Entra tenant in one session at your office. After that it runs on your side: we have no access to the service or to what your users ask.
Custom · fixed scope
MCP for your stack
MCP beyond the standard BC server — custom tools, write-actions with approval steps, other systems next to BC. Same plumbing as the Fusion 360 pipeline, pointed at your stack. A fixed-scope build, usually around two weeks. You get the code and the audit trail.
Scope a build
The plumbing
Three links in the chain.
01
Your AI client
Claude Desktop, VS Code, or Cursor — anything that speaks stdio MCP. No plugin to write, no SDK to learn.
02
bc-mcp-proxy, on your machine
The open-source proxy handles OAuth sign-in, silent token refresh, and reconnects. Your credentials never leave your device.
03
Microsoft's BC MCP server
The proxy connects over HTTPS to the Business Central MCP endpoint Microsoft hosts — with the scopes your admin granted and nothing more.
Security model in one line: the model gets read access to what your Entra admin scoped, every call can be traced once you switch on environment telemetry, and write-actions exist only where you've explicitly built them. Layered validation — schema bounds, business rules, post-build checks — because a system that can write to your ERP is unforgiving.
A word on data
For production, use the right Claude plan.
Against a live tenant, run this on Claude Team, Claude Enterprise, or the Anthropic API — Commercial Terms, no training on your inputs, and a DPA for GDPR. Not the consumer plans. We're an independent ISV, not affiliated with Anthropic; you choose the tier that matches your data.