Ask Claude Desktop “which asset is losing us the most money this week” with Spall connected and the answer names the same press, the same dollar figure, and the same rate basis the Opportunities page shows. The assistant did not guess. It called one tool, read one public API endpoint, and quoted what came back. This guide is about what that connection is, why a plant would want it, and the lines it will not cross.
What MCP is, in one paragraph
The Model Context Protocol is an open standard that lets an AI assistant call outside tools while it answers a question. A tool is a named function with a documented input and output. The assistant decides when to call it, the tool runs on your machine, and the result goes back into the conversation. Claude Desktop, Cursor, and VS Code’s Copilot all speak it. Spall ships one small package, @spall/mcp, that gives any of them fourteen tools for reading your Spall tenant.
Why a plant would want this
The assistant is already open. A plant manager who drafts the morning email in Copilot, or a controls engineer who lives in Cursor, does not want a second chat window. With the package connected, “what was our top downtime reason yesterday” is answered where they already are, from the plant’s own data, with the reason named the way the floor names it.
Questions cross systems. The assistant can hold a maintenance vendor’s quote in one hand and Spall’s failure risk board in the other. “Is Press 2 worth the overhaul quote in this PDF, given how often it fails” is a question no single product answers, and it is a question an assistant with both documents can reason about. Spall’s part of that answer is the failure count and mean time between failures behind the risk number, so the reasoning starts from a figure with a basis.
IT asks the question early. “Can we use our own AI with it” comes up in the first technical call. The answer that satisfies an IT evaluator is not a vendor chatbot. It is a package they can read, a key they control, and a protocol their other tools already use. That is what this is.
What every answer carries
Each tool returns two parts. The first is a plain-text summary that names the basis: the rate source behind a dollar figure, the sample count behind a failure risk, the time window behind an OEE, or “not enough history” when Spall’s own model declined to estimate. The second is the raw data from the API, unchanged. The tool descriptions instruct the assistant to quote the basis alongside any number and to never invent a rate the API did not return.
That second part matters more than it looks. A large language model can still misread a table. When the raw data rides along with the summary, you can check the summary against it in the same conversation, and the assistant can too. A figure with its basis attached is a figure you can argue with. A figure without one is a guess with confidence.
The three refusals
No writes. Fourteen tools, fourteen reads. There is no tool that creates a KPI rule, edits a work order, acknowledges an alert, or writes a setpoint. Spall reads from your machines and never writes to one, and that promise holds one hop further out through your assistant.
No storage. The package keeps nothing. Every tool call is a fresh request through the public API, the same one your ERP integration or Excel connection uses. Your key travels with the request and stays in your own MCP client’s config. The package does not log it, cache it, or send it anywhere but Spall.
No open spend. Thirteen of the tools are plain reads under your key’s normal rate limit. The fourteenth, ask_factory, calls Spall’s own AI, so it needs a key with the ask scope turned on, and it draws against a daily question budget. A read-only key that leaks cannot run up a language-model bill, because it cannot ask.
Where it stands
The package and the seven public API endpoints behind it shipped in September 2026 and are proven against Spall’s demo plant. No customer floor has run it yet. That will change with the first pilot that connects an assistant, and this guide will say so when it does.
How to get it
Create a read-only key on the API keys page, add the ask scope if you want plain-language questions, and follow Connect your AI assistant to Spall with MCP for the config block your client needs. The install line is npx @spall/mcp.
Quick recap
- MCP lets Claude Desktop, Cursor, or Copilot call Spall’s public API as tools while answering a question.
- Fourteen tools, all reads: machines, OEE, ranked opportunities, downtime, quality, alerts, work orders, PM tasks, failure risk, the ledger, insights, golden runs, and ask.
- Every answer carries its basis and the raw data, so a quoted figure can be checked in the same conversation.
- Nothing is written, nothing is stored, and only the ask tool can spend, behind its own scope and budget.
- Proven on the demo plant as of September 2026, not yet on a customer floor.