Session
MCP on a Unified Namespace to map out a Factory Shop Floor
Factories already publish data: PLC tags over MQTT, OPC UA servers, MTConnect agents, ERP exports. An AI agent pointed straight at that is lost: every plant names the same machine three ways, and "stopped" means something different in each source.
This talk shows how, in production, I give an agent a stable view of each plant through a Unified Namespace organized by the ISA-95 hierarchy, and how that view is served over MCP:
• Normalize before exposing. Source codes map to a canonical state × mode model (OPC UA Machinery, MTConnect), with versioned mappings. Unmapped inputs go to quarantine instead of creating assets.
• Design tools around questions, not tables. Coarse tools such as "losses for this line and period" versus generic query tools: accuracy, cost and latency trade-offs.
• Identity end to end. User → role → plant scope → tenant, enforced in Postgres row-level security.
• Other agents as first-class clients. An external partner agent subscribes to the same namespace over MQTT and calls the same tools under an integration role.
• Raw, canonical, metric. Parquet archive, Postgres + TimescaleDB canonical layer, versioned KPI snapshots, and why the agent only reads the last two.
Takeaways
1. A reference layout for exposing industrial data to agents through MCP.
2. Tool-granularity trade-offs, with examples of good and bad tools.
3. A pattern for letting third-party agents in without opening the database.
Yuri Winche Achermann
building AI agents for manufacturing
Aachen, Germany
Links
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top