What is MCP?
The Model Context Protocol (MCP) is an open standard for connecting AI systems to external data sources and tools. Think of it as a universal adapter between AI agents (like Claude or GPT) and the systems that hold your data.
Before MCP, every integration was custom. You'd build a REST API, write wrapper functions, create prompt templates, and hope the AI used them correctly. MCP standardises all of this into a single protocol that any AI client can speak.
The three primitives
MCP defines three types of capabilities that a server can expose:
Tools are functions the AI can call. A tool has a name, a description, input parameters, and returns structured output. For example:
Tool: schema_search
Description: Search for database tables matching a keyword
Input: { "query": "customer" }
Output: [{ "table": "ArCustomer", "rows": 1247, ... }]
Resources are data the AI can read. Think of them as documents or datasets that provide context: a data dictionary, a configuration reference, or a set of business rules.
Prompts are reusable templates that help the AI approach specific tasks. A "monthly KPI review" prompt might instruct the AI to pull five specific metrics and compare them to targets.
How Pharos implements MCP
Pharos is an MCP server built specifically for manufacturing ERP systems. It exposes 50+ tools across four categories:
Schema discovery. Let the AI navigate your data model. Search tables, explore columns, find relationships. The AI doesn't need to know your schema in advance; it discovers it.
Read operations. KPI dashboards, MRP analysis, customer risk, inventory health. Pre-built queries that return structured, analysable data.
Write operations. Labour posting, warehouse transfers, inventory adjustments. Every write goes through a confirmation gate: the AI proposes the transaction, the human approves it.
Analysis. Compare MRP runs, trace demand sources, explain why a suggestion was made. The kind of cross-system analysis that used to require a developer and a database client.
A real example
Here's what an actual MCP tool call looks like:
curl -s -X POST https://pharos.example.com/mcp/tools/kpi_dashboard \
-H "Authorization: Bearer $PHAROS_TOKEN" \
-d '{"period": "2026-03"}'
The response is structured JSON:
{
"period": "2026-03",
"metrics": {
"production_efficiency": { "value": 87.3, "target": 90.0 },
"on_time_delivery": { "value": 94.1, "target": 95.0 },
"scrap_rate": { "value": 2.1, "target": 3.0 },
"oee": { "value": 78.5, "target": 80.0 }
}
}
The AI receives this structured data and can reason about it: "Production efficiency is 2.7% below target. Scrap rate is within tolerance. OEE is the main drag. What's driving it?"
That follow-up question triggers another tool call, and the conversation continues.
Why this matters for manufacturing
Manufacturing ERP systems are notoriously hard to query. The schemas are complex, the table names are cryptic, and the relationships between entities aren't obvious. SYSPRO alone has thousands of tables.
MCP changes the equation. Instead of training users on SQL or building custom dashboards for every question, you give the AI the tools to explore and query the data itself. The human asks the question; the AI figures out how to answer it.
The road ahead
MCP is still early. The standard is evolving, and the ecosystem of MCP servers is growing. But the core idea, giving AI structured, discoverable access to real-world systems, is sound.
Pharos is our bet on this future. It's live in production today, and every week we add more tools, more data sources, and more ways for AI to work with manufacturing data.
Want to connect your ERP to AI? Learn more about Pharos or book a call to discuss your setup.