Every vendor is suddenly selling "AI agents for the shop floor." Strip away the marketing and most of them are built on the same quiet piece of plumbing: the Model Context Protocol (MCP). If you want to understand — or build — any of this, MCP is the concept to learn first. Here it is in plain English.
An LLM like Claude or GPT is brilliant at language and reasoning, but out of the box it knows nothing about your shop. It doesn't know machine 3 is down, or which jobs are behind, or what last week's scrap rate was. MCP is an open standard for handing an AI a set of tools (functions it can call) and resources (data it can read) so it can answer questions grounded in your reality instead of guessing.
The "frozen knowledge" problem
A raw LLM's knowledge is frozen at training time and it has zero access to your live systems. MCP is the bridge that thaws it — it lets the model reach into your ERP, your machines, and your spreadsheets at the moment you ask, instead of making something up.
The shop-floor version, concretely
Imagine an MCP server sitting in front of your FOCAS data and your ERP. Now an engineer can ask an AI, in plain language:
- "What's machine 3 doing right now, and how long has it been in cycle?"
- "Which jobs are behind schedule, and what's blocking them?"
- "Show me every alarm on the Fanuc cell in the last 24 hours."
- "What was our scrap rate on part 4471 last month?"
None of that is the AI being magic. It's the AI calling tools you exposed, which read data you already had — just trapped behind a dozen screens and logins nobody has time to open.
Why this fits manufacturing better than most industries
Manufacturing has two problems MCP is almost purpose-built for. First, data silos: the ERP knows one thing, the machine control another, a shared spreadsheet a third, and nothing talks. Second, tribal knowledge: the answer to "why does this part always chatter on op 20" lives in one senior guy's head. MCP gives an AI a way to read across the silos, and RAG (retrieval over your manuals and notes) gives it a way to surface the tribal knowledge on demand.
What it is NOT
- It is not letting an AI drive your machine. Start read-only. An agent that reads status is useful and safe; one that can press cycle-start is a different, much more dangerous conversation.
- It is not a replacement for your senior people. It's a way to make their knowledge searchable and their data reachable — a copilot, not an autopilot.
- It is not vendor lock-in. MCP is an open protocol; you can build your own server against your own systems.
The AI isn't the hard part. The hard part was always getting the shop's data into one place you could ask a question of. MCP is just the socket you plug the question into.
Convinced it's worth a prototype? The next post walks through building an MCP server for your CNC machines — including the guardrails that keep it read-only and safe.
