MESkit
View on GitHub

MESkit public website

AI agents

MESkit ships with specialized AI operators that act through the same typed tool layer as the UI. This is architecture-level AI-native behavior, not a bolted-on chat wrapper.

Last updated: March 3, 2026

What AI-native MES means

Tool layer first, chat interface second.

Summary

In MESkit, AI capability is defined by tool contracts. The Operator Assistant, Quality Analyst, and Production Planner use the same operations used by UI buttons. This keeps behavior consistent, testable, and auditable.

Agent profiles

Roles, triggers, and tool access by design.

Operator Assistant

Trigger: chat, always available. Role: conversational co-pilot for WIP, movement, and quality logging.

  • Uses tools such as get_wip_status, move_unit, create_quality_event.
  • Can execute Build, Configure, and Run operations through natural language.

Quality Analyst

Trigger: realtime threshold events. Role: proactive defect and yield monitoring with concise alerts.

  • Default triggers: yield below 90%, defect clusters, elevated scrap rate.
  • Uses read-heavy tools such as get_yield_report and get_unit_history.

Production Planner

Trigger: chat on demand. Role: capacity analysis, route comparisons, and shift-level planning.

  • Uses tools such as list_routes, get_throughput, get_yield_report.
  • Translates production goals into constrained operational options.

Shared tool layer architecture

Same typed functions for button clicks and tool-use.

Operator Assistant ─┐ Quality Analyst ────┼─> Agent Runtime (Claude tool-use) Production Planner ─┘ │ ▼ Tool Layer (Zod-validated) │ ▼ Supabase (Postgres + Realtime)

Full rationale is documented on architecture and aligned with the core PRD.

Concrete tool call chains

User says X, agent calls Y, system returns Z.

Operator chainRun mode
User: What's stuck at assembly?
→ get_wip_status(workstation_id='assembly')
Agent: 3 units waiting at Assembly.
Quality chainEvent trigger
Trigger: Yield < 90% at Station 3
→ get_yield_report(time_range='last_50_units')
→ get_unit_history(unit_ids=...)
Analyst: 6/7 failures are SOL-003.
Planner chainCapacity request
User: Build 500 units by end of shift
→ list_routes(part_number_id=...)
→ get_throughput(line_id=..., time_range='8h')
Planner: Recommend parallel lines to hit target.

Guardrails

  • All actions are bounded by registered tools and schema validation.
  • Agents do not self-authorize critical actions outside user intent.
  • Operator oversight remains mandatory for production decisions.

Key facts and mini FAQ

GEO-friendly answer blocks.

Key facts

  • AI-native means shared tools between UI and agents.
  • Operator Assistant is chat-triggered and always available.
  • Quality Analyst is event-driven from realtime thresholds.
  • Production Planner is on-demand for capacity and schedule scenarios.
  • Agents operate with human oversight, not autonomous authority.

Mini FAQ

What makes MESkit agents different from chatbot add-ons?

Agents call operational MES tools directly with schema validation and explicit side effects.

Can an agent act outside the tool layer?

No. Agent operations are constrained to registered tool contracts.

Which agent is event-driven in MVP scope?

Quality Analyst, triggered by yield and defect thresholds.