Skip to main content

Phase 3: Agentic AI & MCP Server

Status: Future phase — planned to build on Phases 1 & 2

Overview​

Phase 3 leverages the rich asset data, observations, and alerts from Phases 1 and 2 to enable intelligent, AI-powered insights and actions. Instead of traditional work order planning, Phase 3 uses agentic AI and MCP (Model Context Protocol) servers to allow natural language interaction with your maintenance data, intelligent recommendations, and context-aware automation.

What is MCP?​

The Model Context Protocol (MCP) is an open standard that enables AI models (LLMs) to securely connect to data sources and tools. In the context of Cyzag Blueprint, an MCP server exposes:

  • Asset data (readings, attributes, types, locations)
  • Historical trends and last known values
  • Operator rounds and observations
  • Phase 2 alerts and triage status

LLMs can then use these MCP servers to answer questions, analyze trends, and recommend actions — all with full context of your plant's operational data.

Core Capabilities​

Natural Language Queries​

Ask questions about your assets and operations in plain English:

  • "What's the current discharge pressure on Pump P-101?"
  • "Show me all pumps with bearing temperatures above 65°C in the last week"
  • "Which assets in Building 3 have the most observations flagged as critical?"
  • "What's the trend for vibration readings on Motor M-204 over the past month?"

The AI agent understands your domain model (assets, attributes, functional locations) and can query the data through the MCP server.

AI Agents with Full Context​

Unlike simple chatbots, agentic AI in Phase 3 has access to:

  • Asset structure: Types, attributes, relationships, and location hierarchy
  • Historical readings: Trends, patterns, anomalies over time
  • Observations: Operator notes, photos, and flagged issues from rounds
  • Alerts: Threshold violations and abnormal conditions from Phase 2
  • Maintenance history: Past observations, resolutions, and actions taken

This context allows agents to provide intelligent recommendations, identify root causes, and suggest priorities.

Intelligent Recommendations​

The AI can analyze patterns and recommend actions:

  • Prioritize maintenance: "Based on recent trends, these 5 assets should be inspected next"
  • Root cause analysis: "The pressure drop correlates with increased vibration readings on the upstream pump — likely a bearing issue"
  • Predictive insights: "Bearing temperature on Pump P-101 has been trending up for 3 weeks; recommend inspection before next scheduled round"
  • Comparative analysis: "Assets in Zone A are showing similar degradation patterns; consider reviewing maintenance procedures"

MCP Server Architecture​

The Cyzag Blueprint MCP server exposes tools (functions) that LLMs can invoke:

  • get_asset_by_id(asset_id) - Retrieve asset details and current attribute values
  • query_readings(asset_id, attribute_code, start_date, end_date) - Get historical readings for trending
  • list_alerts(severity, location, status) - Query Phase 2 alerts by criteria
  • get_observations(round_id) - Retrieve operator observations from a round
  • search_assets(query, filters) - Find assets matching criteria

The MCP server handles authentication, data access control, and formatting results for LLM consumption.

Example: AI-Assisted Maintenance Workflow​

  1. Operator completes round (Phase 1): Captures readings and notes unusual noise on Pump P-101
  2. Alert triggered (Phase 2): Vibration reading exceeds threshold, observation flagged as "medium" severity
  3. Supervisor queries AI agent (Phase 3): "What's going on with Pump P-101?"
  4. Agent responds with context:
    • "Pump P-101 has a medium-severity observation from today's round: unusual noise reported by Jane Smith"
    • "Vibration reading was 8.2 mm/s, which exceeds the threshold of 7.5 mm/s"
    • "Bearing temperature has been trending up over the past 2 weeks (from 55°C to 68°C)"
    • "Historical data shows similar patterns preceded a bearing failure in Pump P-103 last year"
    • "Recommendation: Inspect bearings and lubrication on Pump P-101 within 48 hours"

The supervisor can then decide whether to schedule immediate inspection or continue monitoring.

Example: AI-Assisted Troubleshooting at Blueprint Paper Mill​

To illustrate how Phase 3 provides intelligent insights, consider this scenario at Blueprint Paper Mill:

Scenario: Steam Boiler B-01 Efficiency Investigation​

Date: July 10, 2025 Engineer: Carlos Martinez (Reliability Engineer) Context: Steam production costs have increased over the past month

Carlos suspects an issue with Steam Boiler B-01 but isn't sure where to start investigating.

Phase 3 AI Agent Query​

Carlos opens the AI assistant and asks:

Carlos: "Why is Steam Boiler B-01 performance degrading?"

AI Agent Analysis (Using MCP Server)​

The agent queries the MCP server for context:

MCP Server Tool Invocations:
1. get_asset_by_id("steam-boiler-b-01")
   → Returns: Asset details, current attribute values, location

2. query_readings("steam-boiler-b-01", "steam_pressure", "2025-06-10", "2025-07-10")
   → Returns: 30 days of steam pressure readings

3. query_readings("steam-boiler-b-01", "flue_gas_temperature", "2025-06-10", "2025-07-10")
   → Returns: 30 days of flue gas temperature readings

4. query_readings("steam-boiler-b-01", "feedwater_conductivity", "2025-06-10", "2025-07-10")
   → Returns: 30 days of conductivity readings

5. list_alerts(asset_id="steam-boiler-b-01", start_date="2025-06-10")
   → Returns: 3 observations flagged in past 30 days

6. get_observations(asset_id="steam-boiler-b-01")
   → Returns: Operator notes and photos from recent rounds

AI Agent Response​

🤖 AI Analysis: Steam Boiler B-01 Performance Degradation

**Summary**:
Based on analysis of Steam Boiler B-01 data over the past 30 days,
I've identified two correlated issues that explain the performance
degradation:

**Issue 1: Increasing Flue Gas Temperature**
- Flue gas temperature has risen from 395°C (Jun 10) to 418°C (Jul 10)
- This is within the normal range (380–420°C) but trending toward
  the upper limit
- Higher flue gas temperature indicates heat is escaping up the stack
  rather than being transferred to water/steam
- **Impact**: Reduced thermal efficiency, higher fuel consumption

**Issue 2: Degrading Feedwater Conductivity**
- Feedwater conductivity has increased from 2.1 µS/cm (Jun 10) to
  4.8 µS/cm (Jul 10)
- Still below the 5 µS/cm limit but trending upward
- Higher conductivity indicates dissolved solids in feedwater
- **Impact**: Scale buildup on boiler tubes reduces heat transfer

**Correlation Analysis**:
The flue gas temperature increase correlates strongly with the
conductivity trend (R² = 0.87), suggesting **tube fouling** is the
root cause. As scale builds on tube surfaces, heat transfer
decreases, causing more heat to exit via flue gas.

**Recent Operator Observations**:
- Jul 8: "Steam pressure slightly lower than usual despite same firing rate"
- Jul 5: "Noticed white deposits on sight glass during inspection"

**Recommendation**:
1. Schedule boiler tube cleaning to remove scale deposits
2. Inspect feedwater treatment system for malfunction
3. Test boiler water chemistry after cleaning
4. Monitor flue gas temperature—expect 10–15°C drop after cleaning

**Historical Context**:
Similar pattern occurred in March 2024 (6 months ago). Tube cleaning
at that time restored efficiency and dropped flue gas temperature
from 415°C to 392°C within one day.

**Priority**: Medium
**Estimated Cost Impact**: ~$5,000/month in excess fuel costs if not addressed
**Recommended Action Timeline**: Schedule cleaning within next 2 weeks

[View Trend Charts] [Create Work Order] [Export Report]

Carlos's Response​

Carlos reviews the AI analysis and takes action:

  1. Validates findings: Checks trend charts and confirms correlation
  2. Schedules maintenance: Creates work order for tube cleaning during next planned outage (July 20)
  3. Investigates feedwater treatment: Assigns water treatment technician to inspect demineralizer system
  4. Monitors progress: Sets reminder to check flue gas temperature after cleaning to verify improvement

Outcome​

  • Root cause identified: Tube fouling due to feedwater quality degradation
  • Time saved: 30 minutes of AI analysis vs. hours of manual data review
  • Context provided: Historical comparison (March 2024) and cost impact estimate
  • Actionable recommendations: Clear next steps with priority and timeline

This example demonstrates how Phase 3 transforms raw data into intelligent insights that engineers can act on immediately.

Comparison: Without vs. With Phase 3 AI​

Without Phase 3 (Manual Investigation):

  1. Carlos exports 30 days of boiler data to spreadsheet (20 minutes)
  2. Creates trend charts in Excel (15 minutes)
  3. Reviews operator round notes manually (25 minutes)
  4. Searches maintenance history for similar issues (30 minutes)
  5. Consults with senior engineer (30 minutes)
  6. Writes analysis and recommendation (20 minutes)

Total time: ~2 hours 20 minutes

With Phase 3 AI:

  1. Carlos asks natural language question (30 seconds)
  2. AI agent queries MCP server and analyzes data (5 seconds)
  3. AI provides comprehensive analysis with recommendations (30 seconds)
  4. Carlos validates findings and takes action (5 minutes)

Total time: ~6 minutes

Time saved: ~2 hours 14 minutes (94% reduction)

Why Agentic AI?​

Agentic AI with MCP provides:

  • Flexibility: Natural language interaction without predefined workflows
  • Intelligence: Context-aware recommendations based on trends and patterns
  • Speed: Instant insights without navigating complex UIs
  • Accessibility: Anyone can query data and get answers, not just trained planners
  • Integration: MCP servers can connect to external tools for action execution if needed

Phase 3 augments operator rounds with intelligent, accessible insights that help teams make better decisions faster.

Technical Architecture​

MCP Server Components​

  • Data Access Layer: Secure queries to Phase 1 asset data and Phase 2 observations/alerts
  • Tool Registry: Defined functions (tools) available to LLMs
  • Authentication: Scoped access based on user roles and environments
  • Response Formatting: Convert database queries into LLM-friendly formats

AI Agent Integration​

  • Claude Code CLI: Agents can use the MCP server as a tool during conversations
  • Custom Agents: Build specialized agents for maintenance workflows using the Claude Agent SDK
  • Web Interface: Chat-based UI for supervisors and engineers to query data
  • Mobile Integration: Voice-enabled queries on mobile devices for field personnel

Security & Privacy​

  • MCP server enforces same authentication and scoping as Phase 1/2 APIs
  • No raw data exposed to external LLMs without user consent
  • Audit logs track all AI agent queries and actions
  • Sensitive data (photos, operator notes) can be excluded from MCP tools

Example Use Cases​

Use Case 1: Shift Handover​

Outgoing Supervisor: "Summarize critical issues from today's rounds in Building 2"

AI Agent:

  • "3 critical observations flagged today in Building 2:"
  • "1. Pump P-205: Discharge pressure low (35 bar, expected 40-50). Operator noted cavitation sounds."
  • "2. Motor M-312: Running hours counter stuck at 8,472 — not incrementing. Possible sensor failure."
  • "3. Heat Exchanger HX-401: Outlet temperature trending down over past 3 days (now 78°C, expected 85-90)."

Use Case 2: Trend Analysis​

Reliability Engineer: "Show me all centrifugal pumps with increasing vibration trends over the past month"

AI Agent:

  • "Found 7 pumps with upward vibration trends:"
  • "P-101: +15% (6.8 → 7.8 mm/s)"
  • "P-205: +22% (7.1 → 8.7 mm/s) — already flagged as medium severity"
  • "P-307: +8% (5.2 → 5.6 mm/s)"
  • "...would you like to see detailed charts or schedule inspections?"

Use Case 3: Root Cause Investigation​

Engineer: "Why did we get an alert on Heat Exchanger HX-201 yesterday?"

AI Agent:

  • "Alert triggered at 14:23 yesterday: Outlet temperature 72°C (below threshold of 80°C)"
  • "Inlet temperature was normal (90°C), suggesting heat transfer issue"
  • "Operator noted 'reduced flow' in observation"
  • "Differential pressure on upstream Pump P-150 has been low for 3 days (25 bar, expected 30-35)"
  • "Likely root cause: Pump P-150 performance degradation reducing flow through HX-201"
  • "Recommendation: Inspect Pump P-150 impeller and check for blockages"

What Phase 3 Builds On​

  • Phase 1 Data: Asset structure, attributes, readings, and offline sync provide the foundation
  • Phase 2 Intelligence: Observations and alerts give context for what needs attention
  • Phase 3 AI: Connects the dots, identifies patterns, and makes the data accessible through natural language

Each phase adds value independently, but together they create an intelligent maintenance system that learns from your operations.

See Also: