Real plant data for Claude, with no access to the process
An MCP server behind Faable Auth gives Claude read-only plant data for predictive maintenance. Every call is consented and logged, and no tool can touch a setpoint.
Illustrative scenario based on patterns we see across teams building on Faable.

The team
A food and beverage processing plant with tanks, pumps, heat exchangers and filling lines. Its maintenance team wanted to move from calendar-based to predictive maintenance by letting Claude read real trends: vibration, bearing temperatures, motor current, alarm history and past work orders.
The challenge
Sending internal plant data to a public model meant deciding exactly what leaves the plant, by which path and on whose authority, and being able to prove it afterwards. The AI also had to stay away from the process itself: no setpoints, no start or stop, no recipes. The feedback loop still needed the AI to write something back, so that one write had to be explicit, narrow and approved by a person.
How they built it
- 1
Plant to data API
Gateways next to the PLCs stream telemetry to the plant's data API, each with its own private key and a single write permission. The control network is never opened to the outside.
- 2
An MCP server with a closed list of tools
Claude reaches the data through a small MCP server: sensor history, alarm log and maintenance history, all read-only and aggregated per asset. The only write tool proposes a work order, which a technician approves in the maintenance system.
- 3
Faable Auth as the authorization server
Claude registers itself, the engineer signs in and sees a consent screen in plain words, and the token carries only what both the request and the engineer's role allow. Proposing work orders needs a separate, explicit approval.
- 4
Nothing critical to reach
No tool touches setpoints, recipes or start and stop. The token is bound to the MCP server by its audience, so the control systems reject it outright.
- 5
Every query attributed
Faable Auth logs every connection, consent and token. The MCP server logs every tool call with the person and the AI client behind it. People can disconnect the client at any time.
The result
A predictive maintenance loop on real plant data. Every query is attributed to a person and an AI client, every token is logged, and the AI holds no permission that touches the process.

