Since 2026 · mcp · claude · integration · logistics
An MCP Server That Turns Ten Systems Into One Question
Exposing a freight carrier's enterprise stack to Claude and ChatGPT as typed, auth-scoped tools under a single query surface.
The company had ten systems and no way to ask them a single question. This is the layer that lets you.
What it is. A production MCP server at TBM Carriers that exposes enterprise platforms (fleet maintenance, TMS, telematics, accounting, and CRM) to Claude and ChatGPT as typed, auth-scoped tools. Five are live; the goal is the whole stack.
Why it matters. Any question that crossed two systems was a human errand. Someone filed a request, someone else opened five separate UIs, and the answer arrived hours later, by which point the decision had usually been made without it.
Payoff. Cross-platform lookups that spanned five separate UIs went from about three hours to under fifteen seconds, and ops and finance staff now query live operational data themselves instead of queueing behind an analyst.
What this actually is
TBM Carriers is a logistics operator running the usual sprawl: a TMS for loads, telematics for the trucks, a maintenance system for the fleet, an accounting platform for the money, and a CRM for the customers. Each one is fine on its own. None of them talk.
The CEO's framing was a second brain: one place to ask the company a question and get an answer drawn from every system at once, including the questions nobody thought to ask.
The build is an MCP server. Each platform gets a set of typed tools with real auth scoping, so an assistant can read across the stack without anyone handing out standing credentials to five different systems.
The part that isn't the server
Standing up the connector was the tractable half. The harder half is integration into how people actually work (operations, finance, HR, billing, planning), because a query surface nobody reaches for is just a well-typed API.
That means meeting each function where its questions already live, and being specific about which ones are worth asking a model at all.
- Auth scoping per tool, so read access is granted deliberately rather than inherited
- Typed inputs and outputs, so the model fails loudly on a bad query instead of confidently inventing a plausible one
- One query surface across both Claude and ChatGPT, so the answer doesn't depend on which assistant someone happens to open
What I'd tell someone starting this
The temptation is to connect everything first and figure out the questions later. That produces a large, impressive, unused surface.
The order that worked was the reverse: find the lookup that a person does by hand every week, make that one instant, then let the next tool be pulled in by an actual question rather than pushed in by a roadmap.
Ongoing. The five platforms live today are the floor, not the target.